1
0
mirror of https://github.com/golang/go synced 2024-11-17 19:04:47 -07:00
go/test/live.go

725 lines
18 KiB
Go
Raw Normal View History

// errorcheckwithauto -0 -l -live -wb=0 -d=ssa/insert_resched_checks/off
// +build !ppc64,!ppc64le,!goexperiment.regabiargs
// ppc64 needs a better tighten pass to make f18 pass
// rescheduling checks need to be turned off because there are some live variables across the inserted check call
//
// For register ABI, liveness info changes slightly. See live_regabi.go.
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
// Copyright 2014 The Go Authors. All rights reserved.
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// liveness tests with inlining disabled.
// see also live2.go.
package main
func printnl()
//go:noescape
func printpointer(**int)
//go:noescape
func printintpointer(*int)
//go:noescape
func printstringpointer(*string)
//go:noescape
func printstring(string)
//go:noescape
func printbytepointer(*byte)
func printint(int)
func f1() {
var x *int // ERROR "stack object x \*int$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printpointer(&x) // ERROR "live at call to printpointer: x$"
printpointer(&x)
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
func f2(b bool) {
if b {
printint(0) // nothing live here
return
}
var x *int // ERROR "stack object x \*int$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printpointer(&x) // ERROR "live at call to printpointer: x$"
printpointer(&x)
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
func f3(b1, b2 bool) {
// Here x and y are ambiguously live. In previous go versions they
// were marked as live throughout the function to avoid being
// poisoned in GODEBUG=gcdead=1 mode; this is now no longer the
// case.
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printint(0)
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
if b1 == false {
printint(0)
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
return
}
if b2 {
var x *int // ERROR "stack object x \*int$"
printpointer(&x) // ERROR "live at call to printpointer: x$"
printpointer(&x)
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
} else {
var y *int // ERROR "stack object y \*int$"
printpointer(&y) // ERROR "live at call to printpointer: y$"
printpointer(&y)
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
printint(0) // nothing is live here
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
// The old algorithm treated x as live on all code that
// could flow to a return statement, so it included the
// function entry and code above the declaration of x
// but would not include an indirect use of x in an infinite loop.
// Check that these cases are handled correctly.
func f4(b1, b2 bool) { // x not live here
if b2 {
printint(0) // x not live here
return
}
var z **int
x := new(int) // ERROR "stack object x \*int$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
*x = 42
z = &x
printint(**z) // ERROR "live at call to printint: x$"
if b2 {
printint(1) // x not live here
return
}
for {
printint(**z) // ERROR "live at call to printint: x$"
}
}
func f5(b1 bool) {
var z **int
if b1 {
x := new(int) // ERROR "stack object x \*int$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
*x = 42
z = &x
} else {
y := new(int) // ERROR "stack object y \*int$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
*y = 54
z = &y
}
printint(**z) // nothing live here
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
// confusion about the _ result used to cause spurious "live at entry to f6: _".
func f6() (_, y string) {
y = "hello"
return
}
// confusion about addressed results used to cause "live at entry to f7: x".
func f7() (x string) { // ERROR "stack object x string"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
_ = &x
x = "hello"
return
}
// ignoring block returns used to cause "live at entry to f8: x, y".
func f8() (x, y string) {
return g8()
}
func g8() (string, string)
// ignoring block assignments used to cause "live at entry to f9: x"
// issue 7205
var i9 interface{}
func f9() bool {
g8()
x := i9
y := interface{}(g18()) // ERROR "live at call to convT: x.data$" "live at call to g18: x.data$" "stack object .autotmp_[0-9]+ \[2\]string$"
i9 = y // make y escape so the line above has to call convT
return x != y
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
// liveness formerly confused by UNDEF followed by RET,
// leading to "live at entry to f10: ~r1" (unnamed result).
func f10() string {
panic(1)
}
// liveness formerly confused by select, thinking runtime.selectgo
// can return to next instruction; it always jumps elsewhere.
// note that you have to use at least two cases in the select
// to get a true select; smaller selects compile to optimized helper functions.
var c chan *int
var b bool
// this used to have a spurious "live at entry to f11a: ~r0"
func f11a() *int {
select { // ERROR "stack object .autotmp_[0-9]+ \[2\]struct"
case <-c:
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
return nil
case <-c:
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
return nil
}
}
func f11b() *int {
p := new(int)
if b {
// At this point p is dead: the code here cannot
// get to the bottom of the function.
// This used to have a spurious "live at call to printint: p".
printint(1) // nothing live here!
select { // ERROR "stack object .autotmp_[0-9]+ \[2\]struct"
case <-c:
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
return nil
case <-c:
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
return nil
}
}
println(*p)
return nil
}
var sink *int
func f11c() *int {
p := new(int)
sink = p // prevent stack allocation, otherwise p is rematerializeable
if b {
// Unlike previous, the cases in this select fall through,
// so we can get to the println, so p is not dead.
printint(1) // ERROR "live at call to printint: p$"
select { // ERROR "live at call to selectgo: p$" "stack object .autotmp_[0-9]+ \[2\]struct"
case <-c:
case <-c:
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
}
println(*p)
return nil
}
// similarly, select{} does not fall through.
// this used to have a spurious "live at entry to f12: ~r0".
func f12() *int {
if b {
select {}
} else {
return nil
}
}
// incorrectly placed VARDEF annotations can cause missing liveness annotations.
// this used to be missing the fact that s is live during the call to g13 (because it is
// needed for the call to h13).
func f13() {
s := g14()
s = h13(s, g13(s)) // ERROR "live at call to g13: s.ptr$"
}
func g13(string) string
func h13(string, string) string
// more incorrectly placed VARDEF.
func f14() {
x := g14() // ERROR "stack object x string$"
printstringpointer(&x)
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
func g14() string
// Checking that various temporaries do not persist or cause
// ambiguously live values that must be zeroed.
// The exact temporary names are inconsequential but we are
// trying to check that there is only one at any given site,
// and also that none show up in "ambiguously live" messages.
var m map[string]int
runtime: add mapdelete_fast* Add benchmarks for map delete with int32/int64/string key Benchmark results on darwin/amd64 name old time/op new time/op delta MapDelete/Int32/1-8 151ns ± 8% 99ns ± 3% -34.39% (p=0.008 n=5+5) MapDelete/Int32/2-8 128ns ± 2% 111ns ±15% -13.40% (p=0.040 n=5+5) MapDelete/Int32/4-8 128ns ± 5% 114ns ± 2% -10.82% (p=0.008 n=5+5) MapDelete/Int64/1-8 144ns ± 0% 104ns ± 3% -27.53% (p=0.016 n=4+5) MapDelete/Int64/2-8 153ns ± 1% 126ns ± 3% -17.17% (p=0.008 n=5+5) MapDelete/Int64/4-8 178ns ± 3% 136ns ± 2% -23.60% (p=0.008 n=5+5) MapDelete/Str/1-8 187ns ± 3% 171ns ± 3% -8.54% (p=0.008 n=5+5) MapDelete/Str/2-8 221ns ± 3% 206ns ± 4% -7.18% (p=0.016 n=5+4) MapDelete/Str/4-8 256ns ± 5% 232ns ± 2% -9.36% (p=0.016 n=4+5) name old time/op new time/op delta BinaryTree17-8 2.78s ± 7% 2.70s ± 1% ~ (p=0.151 n=5+5) Fannkuch11-8 3.21s ± 2% 3.19s ± 1% ~ (p=0.310 n=5+5) FmtFprintfEmpty-8 49.1ns ± 3% 50.2ns ± 2% ~ (p=0.095 n=5+5) FmtFprintfString-8 78.6ns ± 4% 80.2ns ± 5% ~ (p=0.460 n=5+5) FmtFprintfInt-8 79.7ns ± 1% 81.0ns ± 3% ~ (p=0.103 n=5+5) FmtFprintfIntInt-8 117ns ± 2% 119ns ± 0% ~ (p=0.079 n=5+4) FmtFprintfPrefixedInt-8 153ns ± 1% 146ns ± 3% -4.19% (p=0.024 n=5+5) FmtFprintfFloat-8 239ns ± 1% 237ns ± 1% ~ (p=0.246 n=5+5) FmtManyArgs-8 506ns ± 2% 509ns ± 2% ~ (p=0.238 n=5+5) GobDecode-8 7.06ms ± 4% 6.86ms ± 1% ~ (p=0.222 n=5+5) GobEncode-8 6.01ms ± 5% 5.87ms ± 2% ~ (p=0.222 n=5+5) Gzip-8 246ms ± 4% 236ms ± 1% -4.12% (p=0.008 n=5+5) Gunzip-8 37.7ms ± 4% 37.3ms ± 1% ~ (p=0.841 n=5+5) HTTPClientServer-8 64.9µs ± 1% 64.4µs ± 0% -0.80% (p=0.032 n=5+4) JSONEncode-8 16.0ms ± 2% 16.2ms ±11% ~ (p=0.548 n=5+5) JSONDecode-8 53.2ms ± 2% 53.1ms ± 4% ~ (p=1.000 n=5+5) Mandelbrot200-8 4.33ms ± 2% 4.32ms ± 2% ~ (p=0.841 n=5+5) GoParse-8 3.24ms ± 2% 3.27ms ± 4% ~ (p=0.690 n=5+5) RegexpMatchEasy0_32-8 86.2ns ± 1% 85.2ns ± 3% ~ (p=0.286 n=5+5) RegexpMatchEasy0_1K-8 198ns ± 2% 199ns ± 1% ~ (p=0.310 n=5+5) RegexpMatchEasy1_32-8 82.6ns ± 2% 81.8ns ± 1% ~ (p=0.294 n=5+5) RegexpMatchEasy1_1K-8 359ns ± 2% 354ns ± 1% -1.39% (p=0.048 n=5+5) RegexpMatchMedium_32-8 123ns ± 2% 123ns ± 1% ~ (p=0.905 n=5+5) RegexpMatchMedium_1K-8 38.2µs ± 2% 38.6µs ± 8% ~ (p=0.690 n=5+5) RegexpMatchHard_32-8 1.92µs ± 2% 1.91µs ± 5% ~ (p=0.460 n=5+5) RegexpMatchHard_1K-8 57.6µs ± 1% 57.0µs ± 2% ~ (p=0.310 n=5+5) Revcomp-8 483ms ± 7% 441ms ± 1% -8.79% (p=0.016 n=5+4) Template-8 58.0ms ± 1% 58.2ms ± 7% ~ (p=0.310 n=5+5) TimeParse-8 324ns ± 6% 312ns ± 2% ~ (p=0.087 n=5+5) TimeFormat-8 330ns ± 1% 329ns ± 1% ~ (p=0.968 n=5+5) name old speed new speed delta GobDecode-8 109MB/s ± 4% 112MB/s ± 1% ~ (p=0.222 n=5+5) GobEncode-8 128MB/s ± 5% 131MB/s ± 2% ~ (p=0.222 n=5+5) Gzip-8 78.9MB/s ± 4% 82.3MB/s ± 1% +4.25% (p=0.008 n=5+5) Gunzip-8 514MB/s ± 4% 521MB/s ± 1% ~ (p=0.841 n=5+5) JSONEncode-8 121MB/s ± 2% 120MB/s ±10% ~ (p=0.548 n=5+5) JSONDecode-8 36.5MB/s ± 2% 36.6MB/s ± 4% ~ (p=1.000 n=5+5) GoParse-8 17.9MB/s ± 2% 17.7MB/s ± 4% ~ (p=0.730 n=5+5) RegexpMatchEasy0_32-8 371MB/s ± 1% 375MB/s ± 3% ~ (p=0.310 n=5+5) RegexpMatchEasy0_1K-8 5.15GB/s ± 1% 5.13GB/s ± 1% ~ (p=0.548 n=5+5) RegexpMatchEasy1_32-8 387MB/s ± 2% 391MB/s ± 1% ~ (p=0.310 n=5+5) RegexpMatchEasy1_1K-8 2.85GB/s ± 2% 2.89GB/s ± 1% ~ (p=0.056 n=5+5) RegexpMatchMedium_32-8 8.07MB/s ± 2% 8.06MB/s ± 1% ~ (p=0.730 n=5+5) RegexpMatchMedium_1K-8 26.8MB/s ± 2% 26.6MB/s ± 7% ~ (p=0.690 n=5+5) RegexpMatchHard_32-8 16.7MB/s ± 2% 16.7MB/s ± 5% ~ (p=0.421 n=5+5) RegexpMatchHard_1K-8 17.8MB/s ± 1% 18.0MB/s ± 2% ~ (p=0.310 n=5+5) Revcomp-8 527MB/s ± 6% 577MB/s ± 1% +9.44% (p=0.016 n=5+4) Template-8 33.5MB/s ± 1% 33.4MB/s ± 7% ~ (p=0.310 n=5+5) Updates #19495 Change-Id: Ib9ece1690813d9b4788455db43d30891e2138df5 Reviewed-on: https://go-review.googlesource.com/38172 Reviewed-by: Hugues Bruant <hugues.bruant@gmail.com> Reviewed-by: Keith Randall <khr@golang.org> Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org>
2017-03-14 12:11:28 -06:00
var mi map[interface{}]int
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
runtime: add mapdelete_fast* Add benchmarks for map delete with int32/int64/string key Benchmark results on darwin/amd64 name old time/op new time/op delta MapDelete/Int32/1-8 151ns ± 8% 99ns ± 3% -34.39% (p=0.008 n=5+5) MapDelete/Int32/2-8 128ns ± 2% 111ns ±15% -13.40% (p=0.040 n=5+5) MapDelete/Int32/4-8 128ns ± 5% 114ns ± 2% -10.82% (p=0.008 n=5+5) MapDelete/Int64/1-8 144ns ± 0% 104ns ± 3% -27.53% (p=0.016 n=4+5) MapDelete/Int64/2-8 153ns ± 1% 126ns ± 3% -17.17% (p=0.008 n=5+5) MapDelete/Int64/4-8 178ns ± 3% 136ns ± 2% -23.60% (p=0.008 n=5+5) MapDelete/Str/1-8 187ns ± 3% 171ns ± 3% -8.54% (p=0.008 n=5+5) MapDelete/Str/2-8 221ns ± 3% 206ns ± 4% -7.18% (p=0.016 n=5+4) MapDelete/Str/4-8 256ns ± 5% 232ns ± 2% -9.36% (p=0.016 n=4+5) name old time/op new time/op delta BinaryTree17-8 2.78s ± 7% 2.70s ± 1% ~ (p=0.151 n=5+5) Fannkuch11-8 3.21s ± 2% 3.19s ± 1% ~ (p=0.310 n=5+5) FmtFprintfEmpty-8 49.1ns ± 3% 50.2ns ± 2% ~ (p=0.095 n=5+5) FmtFprintfString-8 78.6ns ± 4% 80.2ns ± 5% ~ (p=0.460 n=5+5) FmtFprintfInt-8 79.7ns ± 1% 81.0ns ± 3% ~ (p=0.103 n=5+5) FmtFprintfIntInt-8 117ns ± 2% 119ns ± 0% ~ (p=0.079 n=5+4) FmtFprintfPrefixedInt-8 153ns ± 1% 146ns ± 3% -4.19% (p=0.024 n=5+5) FmtFprintfFloat-8 239ns ± 1% 237ns ± 1% ~ (p=0.246 n=5+5) FmtManyArgs-8 506ns ± 2% 509ns ± 2% ~ (p=0.238 n=5+5) GobDecode-8 7.06ms ± 4% 6.86ms ± 1% ~ (p=0.222 n=5+5) GobEncode-8 6.01ms ± 5% 5.87ms ± 2% ~ (p=0.222 n=5+5) Gzip-8 246ms ± 4% 236ms ± 1% -4.12% (p=0.008 n=5+5) Gunzip-8 37.7ms ± 4% 37.3ms ± 1% ~ (p=0.841 n=5+5) HTTPClientServer-8 64.9µs ± 1% 64.4µs ± 0% -0.80% (p=0.032 n=5+4) JSONEncode-8 16.0ms ± 2% 16.2ms ±11% ~ (p=0.548 n=5+5) JSONDecode-8 53.2ms ± 2% 53.1ms ± 4% ~ (p=1.000 n=5+5) Mandelbrot200-8 4.33ms ± 2% 4.32ms ± 2% ~ (p=0.841 n=5+5) GoParse-8 3.24ms ± 2% 3.27ms ± 4% ~ (p=0.690 n=5+5) RegexpMatchEasy0_32-8 86.2ns ± 1% 85.2ns ± 3% ~ (p=0.286 n=5+5) RegexpMatchEasy0_1K-8 198ns ± 2% 199ns ± 1% ~ (p=0.310 n=5+5) RegexpMatchEasy1_32-8 82.6ns ± 2% 81.8ns ± 1% ~ (p=0.294 n=5+5) RegexpMatchEasy1_1K-8 359ns ± 2% 354ns ± 1% -1.39% (p=0.048 n=5+5) RegexpMatchMedium_32-8 123ns ± 2% 123ns ± 1% ~ (p=0.905 n=5+5) RegexpMatchMedium_1K-8 38.2µs ± 2% 38.6µs ± 8% ~ (p=0.690 n=5+5) RegexpMatchHard_32-8 1.92µs ± 2% 1.91µs ± 5% ~ (p=0.460 n=5+5) RegexpMatchHard_1K-8 57.6µs ± 1% 57.0µs ± 2% ~ (p=0.310 n=5+5) Revcomp-8 483ms ± 7% 441ms ± 1% -8.79% (p=0.016 n=5+4) Template-8 58.0ms ± 1% 58.2ms ± 7% ~ (p=0.310 n=5+5) TimeParse-8 324ns ± 6% 312ns ± 2% ~ (p=0.087 n=5+5) TimeFormat-8 330ns ± 1% 329ns ± 1% ~ (p=0.968 n=5+5) name old speed new speed delta GobDecode-8 109MB/s ± 4% 112MB/s ± 1% ~ (p=0.222 n=5+5) GobEncode-8 128MB/s ± 5% 131MB/s ± 2% ~ (p=0.222 n=5+5) Gzip-8 78.9MB/s ± 4% 82.3MB/s ± 1% +4.25% (p=0.008 n=5+5) Gunzip-8 514MB/s ± 4% 521MB/s ± 1% ~ (p=0.841 n=5+5) JSONEncode-8 121MB/s ± 2% 120MB/s ±10% ~ (p=0.548 n=5+5) JSONDecode-8 36.5MB/s ± 2% 36.6MB/s ± 4% ~ (p=1.000 n=5+5) GoParse-8 17.9MB/s ± 2% 17.7MB/s ± 4% ~ (p=0.730 n=5+5) RegexpMatchEasy0_32-8 371MB/s ± 1% 375MB/s ± 3% ~ (p=0.310 n=5+5) RegexpMatchEasy0_1K-8 5.15GB/s ± 1% 5.13GB/s ± 1% ~ (p=0.548 n=5+5) RegexpMatchEasy1_32-8 387MB/s ± 2% 391MB/s ± 1% ~ (p=0.310 n=5+5) RegexpMatchEasy1_1K-8 2.85GB/s ± 2% 2.89GB/s ± 1% ~ (p=0.056 n=5+5) RegexpMatchMedium_32-8 8.07MB/s ± 2% 8.06MB/s ± 1% ~ (p=0.730 n=5+5) RegexpMatchMedium_1K-8 26.8MB/s ± 2% 26.6MB/s ± 7% ~ (p=0.690 n=5+5) RegexpMatchHard_32-8 16.7MB/s ± 2% 16.7MB/s ± 5% ~ (p=0.421 n=5+5) RegexpMatchHard_1K-8 17.8MB/s ± 1% 18.0MB/s ± 2% ~ (p=0.310 n=5+5) Revcomp-8 527MB/s ± 6% 577MB/s ± 1% +9.44% (p=0.016 n=5+4) Template-8 33.5MB/s ± 1% 33.4MB/s ± 7% ~ (p=0.310 n=5+5) Updates #19495 Change-Id: Ib9ece1690813d9b4788455db43d30891e2138df5 Reviewed-on: https://go-review.googlesource.com/38172 Reviewed-by: Hugues Bruant <hugues.bruant@gmail.com> Reviewed-by: Keith Randall <khr@golang.org> Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org>
2017-03-14 12:11:28 -06:00
// str and iface are used to ensure that a temp is required for runtime calls below.
cmd/compile: convert constants to interfaces without allocating The order pass is responsible for ensuring that values passed to runtime functions, including convT2E/convT2I, are addressable. Prior to this CL, this was always accomplished by creating a temp, which frequently escaped to the heap, causing allocations, perhaps most notably in code like: fmt.Println(1, 2, 3) // allocates three times None of the runtime routines modify the contents of the pointers they receive, so in the case of constants, instead of creating a temp value, we can create a static value. (Marking the static value as read-only provides protection against accidental attempts by the runtime to modify the constant data.) This improves code generation for code like: panic("abc") c <- 2 // c is a chan int which can now simply refer to "abc" and 2, rather than going by way of a temporary. It also allows us to optimize convT2E/convT2I, by recognizing static readonly values and directly constructing the interface. This CL adds ~0.5% to binary size, despite decreasing the size of many functions, because it also adds many static symbols. This binary size regression could be recovered in future (but currently unplanned) work. There is a lot of content-duplication in these symbols; this statement generates six new symbols, three containing an int 1 and three containing a pointer to the string "a": fmt.Println(1, 1, 1, "a", "a", "a") These symbols could be made content-addressable. Furthermore, these symbols are small, so the alignment and naming overhead is large. As with the go.strings section, these symbols could be hidden and have their alignment reduced. The changes to test/live.go make it impossible (at least with current optimization techniques) to place the values being passed to the runtime in static symbols, preserving autotmp creation. Fixes #18704 Benchmarks from fmt and go-kit's logging package: github.com/go-kit/kit/log name old time/op new time/op delta JSONLoggerSimple-8 1.91µs ± 2% 2.11µs ±22% ~ (p=1.000 n=9+10) JSONLoggerContextual-8 2.60µs ± 6% 2.43µs ± 2% -6.29% (p=0.000 n=9+10) Discard-8 101ns ± 2% 34ns ±14% -66.33% (p=0.000 n=10+9) OneWith-8 161ns ± 1% 102ns ±16% -36.78% (p=0.000 n=10+10) TwoWith-8 175ns ± 3% 106ns ± 7% -39.36% (p=0.000 n=10+9) TenWith-8 293ns ± 3% 227ns ±15% -22.44% (p=0.000 n=9+10) LogfmtLoggerSimple-8 704ns ± 2% 608ns ± 2% -13.65% (p=0.000 n=10+9) LogfmtLoggerContextual-8 962ns ± 1% 860ns ±17% -10.57% (p=0.003 n=9+10) NopLoggerSimple-8 188ns ± 1% 120ns ± 1% -36.39% (p=0.000 n=9+10) NopLoggerContextual-8 379ns ± 1% 243ns ± 0% -35.77% (p=0.000 n=9+10) ValueBindingTimestamp-8 577ns ± 1% 499ns ± 1% -13.51% (p=0.000 n=10+10) ValueBindingCaller-8 898ns ± 2% 844ns ± 2% -6.00% (p=0.000 n=10+10) name old alloc/op new alloc/op delta JSONLoggerSimple-8 904B ± 0% 872B ± 0% -3.54% (p=0.000 n=10+10) JSONLoggerContextual-8 1.20kB ± 0% 1.14kB ± 0% -5.33% (p=0.000 n=10+10) Discard-8 64.0B ± 0% 32.0B ± 0% -50.00% (p=0.000 n=10+10) OneWith-8 96.0B ± 0% 64.0B ± 0% -33.33% (p=0.000 n=10+10) TwoWith-8 160B ± 0% 128B ± 0% -20.00% (p=0.000 n=10+10) TenWith-8 672B ± 0% 640B ± 0% -4.76% (p=0.000 n=10+10) LogfmtLoggerSimple-8 128B ± 0% 96B ± 0% -25.00% (p=0.000 n=10+10) LogfmtLoggerContextual-8 304B ± 0% 240B ± 0% -21.05% (p=0.000 n=10+10) NopLoggerSimple-8 128B ± 0% 96B ± 0% -25.00% (p=0.000 n=10+10) NopLoggerContextual-8 304B ± 0% 240B ± 0% -21.05% (p=0.000 n=10+10) ValueBindingTimestamp-8 159B ± 0% 127B ± 0% -20.13% (p=0.000 n=10+10) ValueBindingCaller-8 112B ± 0% 80B ± 0% -28.57% (p=0.000 n=10+10) name old allocs/op new allocs/op delta JSONLoggerSimple-8 19.0 ± 0% 17.0 ± 0% -10.53% (p=0.000 n=10+10) JSONLoggerContextual-8 25.0 ± 0% 21.0 ± 0% -16.00% (p=0.000 n=10+10) Discard-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) OneWith-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) TwoWith-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) TenWith-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) LogfmtLoggerSimple-8 4.00 ± 0% 2.00 ± 0% -50.00% (p=0.000 n=10+10) LogfmtLoggerContextual-8 7.00 ± 0% 3.00 ± 0% -57.14% (p=0.000 n=10+10) NopLoggerSimple-8 4.00 ± 0% 2.00 ± 0% -50.00% (p=0.000 n=10+10) NopLoggerContextual-8 7.00 ± 0% 3.00 ± 0% -57.14% (p=0.000 n=10+10) ValueBindingTimestamp-8 5.00 ± 0% 3.00 ± 0% -40.00% (p=0.000 n=10+10) ValueBindingCaller-8 4.00 ± 0% 2.00 ± 0% -50.00% (p=0.000 n=10+10) fmt name old time/op new time/op delta SprintfPadding-8 88.9ns ± 3% 79.1ns ± 1% -11.09% (p=0.000 n=10+7) SprintfEmpty-8 12.6ns ± 3% 12.8ns ± 3% ~ (p=0.136 n=10+10) SprintfString-8 38.7ns ± 5% 26.9ns ± 6% -30.65% (p=0.000 n=10+10) SprintfTruncateString-8 56.7ns ± 2% 47.0ns ± 3% -17.05% (p=0.000 n=10+10) SprintfQuoteString-8 164ns ± 2% 153ns ± 2% -7.01% (p=0.000 n=10+10) SprintfInt-8 38.9ns ±15% 26.5ns ± 2% -31.93% (p=0.000 n=10+9) SprintfIntInt-8 60.3ns ± 9% 38.2ns ± 1% -36.67% (p=0.000 n=10+8) SprintfPrefixedInt-8 58.6ns ±13% 51.2ns ±11% -12.66% (p=0.001 n=10+10) SprintfFloat-8 71.4ns ± 3% 64.2ns ± 3% -10.08% (p=0.000 n=8+10) SprintfComplex-8 175ns ± 3% 159ns ± 2% -9.03% (p=0.000 n=10+10) SprintfBoolean-8 33.5ns ± 4% 25.7ns ± 5% -23.28% (p=0.000 n=10+10) SprintfHexString-8 65.3ns ± 3% 51.7ns ± 5% -20.86% (p=0.000 n=10+9) SprintfHexBytes-8 67.2ns ± 5% 67.9ns ± 4% ~ (p=0.383 n=10+10) SprintfBytes-8 129ns ± 7% 124ns ± 7% ~ (p=0.074 n=9+10) SprintfStringer-8 127ns ± 4% 126ns ± 8% ~ (p=0.506 n=9+10) SprintfStructure-8 357ns ± 3% 359ns ± 3% ~ (p=0.469 n=10+10) ManyArgs-8 203ns ± 6% 126ns ± 3% -37.94% (p=0.000 n=10+10) FprintInt-8 119ns ±10% 74ns ± 3% -37.54% (p=0.000 n=10+10) FprintfBytes-8 122ns ± 4% 120ns ± 3% ~ (p=0.124 n=10+10) FprintIntNoAlloc-8 78.2ns ± 5% 74.1ns ± 3% -5.28% (p=0.000 n=10+10) ScanInts-8 349µs ± 1% 349µs ± 0% ~ (p=0.606 n=9+8) ScanRecursiveInt-8 43.8ms ± 7% 40.1ms ± 2% -8.42% (p=0.000 n=10+10) ScanRecursiveIntReaderWrapper-8 43.5ms ± 4% 40.4ms ± 2% -7.16% (p=0.000 n=10+9) name old alloc/op new alloc/op delta SprintfPadding-8 24.0B ± 0% 16.0B ± 0% -33.33% (p=0.000 n=10+10) SprintfEmpty-8 0.00B 0.00B ~ (all equal) SprintfString-8 21.0B ± 0% 5.0B ± 0% -76.19% (p=0.000 n=10+10) SprintfTruncateString-8 32.0B ± 0% 16.0B ± 0% -50.00% (p=0.000 n=10+10) SprintfQuoteString-8 48.0B ± 0% 32.0B ± 0% -33.33% (p=0.000 n=10+10) SprintfInt-8 16.0B ± 0% 1.0B ± 0% -93.75% (p=0.000 n=10+10) SprintfIntInt-8 24.0B ± 0% 3.0B ± 0% -87.50% (p=0.000 n=10+10) SprintfPrefixedInt-8 72.0B ± 0% 64.0B ± 0% -11.11% (p=0.000 n=10+10) SprintfFloat-8 16.0B ± 0% 8.0B ± 0% -50.00% (p=0.000 n=10+10) SprintfComplex-8 48.0B ± 0% 32.0B ± 0% -33.33% (p=0.000 n=10+10) SprintfBoolean-8 8.00B ± 0% 4.00B ± 0% -50.00% (p=0.000 n=10+10) SprintfHexString-8 96.0B ± 0% 80.0B ± 0% -16.67% (p=0.000 n=10+10) SprintfHexBytes-8 112B ± 0% 112B ± 0% ~ (all equal) SprintfBytes-8 96.0B ± 0% 96.0B ± 0% ~ (all equal) SprintfStringer-8 32.0B ± 0% 32.0B ± 0% ~ (all equal) SprintfStructure-8 256B ± 0% 256B ± 0% ~ (all equal) ManyArgs-8 80.0B ± 0% 0.0B -100.00% (p=0.000 n=10+10) FprintInt-8 8.00B ± 0% 0.00B -100.00% (p=0.000 n=10+10) FprintfBytes-8 32.0B ± 0% 32.0B ± 0% ~ (all equal) FprintIntNoAlloc-8 0.00B 0.00B ~ (all equal) ScanInts-8 15.2kB ± 0% 15.2kB ± 0% ~ (p=0.248 n=9+10) ScanRecursiveInt-8 21.6kB ± 0% 21.6kB ± 0% ~ (all equal) ScanRecursiveIntReaderWrapper-8 21.7kB ± 0% 21.7kB ± 0% ~ (all equal) name old allocs/op new allocs/op delta SprintfPadding-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfEmpty-8 0.00 0.00 ~ (all equal) SprintfString-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfTruncateString-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfQuoteString-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfInt-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfIntInt-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) SprintfPrefixedInt-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfFloat-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfComplex-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfBoolean-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfHexString-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfHexBytes-8 2.00 ± 0% 2.00 ± 0% ~ (all equal) SprintfBytes-8 2.00 ± 0% 2.00 ± 0% ~ (all equal) SprintfStringer-8 4.00 ± 0% 4.00 ± 0% ~ (all equal) SprintfStructure-8 7.00 ± 0% 7.00 ± 0% ~ (all equal) ManyArgs-8 8.00 ± 0% 0.00 -100.00% (p=0.000 n=10+10) FprintInt-8 1.00 ± 0% 0.00 -100.00% (p=0.000 n=10+10) FprintfBytes-8 1.00 ± 0% 1.00 ± 0% ~ (all equal) FprintIntNoAlloc-8 0.00 0.00 ~ (all equal) ScanInts-8 1.60k ± 0% 1.60k ± 0% ~ (all equal) ScanRecursiveInt-8 1.71k ± 0% 1.71k ± 0% ~ (all equal) ScanRecursiveIntReaderWrapper-8 1.71k ± 0% 1.71k ± 0% ~ (all equal) Change-Id: I7ba72a25fea4140a0ba40a9f443103ed87cc69b5 Reviewed-on: https://go-review.googlesource.com/35554 Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Keith Randall <khr@golang.org>
2017-01-21 14:41:06 -07:00
func str() string
runtime: add mapdelete_fast* Add benchmarks for map delete with int32/int64/string key Benchmark results on darwin/amd64 name old time/op new time/op delta MapDelete/Int32/1-8 151ns ± 8% 99ns ± 3% -34.39% (p=0.008 n=5+5) MapDelete/Int32/2-8 128ns ± 2% 111ns ±15% -13.40% (p=0.040 n=5+5) MapDelete/Int32/4-8 128ns ± 5% 114ns ± 2% -10.82% (p=0.008 n=5+5) MapDelete/Int64/1-8 144ns ± 0% 104ns ± 3% -27.53% (p=0.016 n=4+5) MapDelete/Int64/2-8 153ns ± 1% 126ns ± 3% -17.17% (p=0.008 n=5+5) MapDelete/Int64/4-8 178ns ± 3% 136ns ± 2% -23.60% (p=0.008 n=5+5) MapDelete/Str/1-8 187ns ± 3% 171ns ± 3% -8.54% (p=0.008 n=5+5) MapDelete/Str/2-8 221ns ± 3% 206ns ± 4% -7.18% (p=0.016 n=5+4) MapDelete/Str/4-8 256ns ± 5% 232ns ± 2% -9.36% (p=0.016 n=4+5) name old time/op new time/op delta BinaryTree17-8 2.78s ± 7% 2.70s ± 1% ~ (p=0.151 n=5+5) Fannkuch11-8 3.21s ± 2% 3.19s ± 1% ~ (p=0.310 n=5+5) FmtFprintfEmpty-8 49.1ns ± 3% 50.2ns ± 2% ~ (p=0.095 n=5+5) FmtFprintfString-8 78.6ns ± 4% 80.2ns ± 5% ~ (p=0.460 n=5+5) FmtFprintfInt-8 79.7ns ± 1% 81.0ns ± 3% ~ (p=0.103 n=5+5) FmtFprintfIntInt-8 117ns ± 2% 119ns ± 0% ~ (p=0.079 n=5+4) FmtFprintfPrefixedInt-8 153ns ± 1% 146ns ± 3% -4.19% (p=0.024 n=5+5) FmtFprintfFloat-8 239ns ± 1% 237ns ± 1% ~ (p=0.246 n=5+5) FmtManyArgs-8 506ns ± 2% 509ns ± 2% ~ (p=0.238 n=5+5) GobDecode-8 7.06ms ± 4% 6.86ms ± 1% ~ (p=0.222 n=5+5) GobEncode-8 6.01ms ± 5% 5.87ms ± 2% ~ (p=0.222 n=5+5) Gzip-8 246ms ± 4% 236ms ± 1% -4.12% (p=0.008 n=5+5) Gunzip-8 37.7ms ± 4% 37.3ms ± 1% ~ (p=0.841 n=5+5) HTTPClientServer-8 64.9µs ± 1% 64.4µs ± 0% -0.80% (p=0.032 n=5+4) JSONEncode-8 16.0ms ± 2% 16.2ms ±11% ~ (p=0.548 n=5+5) JSONDecode-8 53.2ms ± 2% 53.1ms ± 4% ~ (p=1.000 n=5+5) Mandelbrot200-8 4.33ms ± 2% 4.32ms ± 2% ~ (p=0.841 n=5+5) GoParse-8 3.24ms ± 2% 3.27ms ± 4% ~ (p=0.690 n=5+5) RegexpMatchEasy0_32-8 86.2ns ± 1% 85.2ns ± 3% ~ (p=0.286 n=5+5) RegexpMatchEasy0_1K-8 198ns ± 2% 199ns ± 1% ~ (p=0.310 n=5+5) RegexpMatchEasy1_32-8 82.6ns ± 2% 81.8ns ± 1% ~ (p=0.294 n=5+5) RegexpMatchEasy1_1K-8 359ns ± 2% 354ns ± 1% -1.39% (p=0.048 n=5+5) RegexpMatchMedium_32-8 123ns ± 2% 123ns ± 1% ~ (p=0.905 n=5+5) RegexpMatchMedium_1K-8 38.2µs ± 2% 38.6µs ± 8% ~ (p=0.690 n=5+5) RegexpMatchHard_32-8 1.92µs ± 2% 1.91µs ± 5% ~ (p=0.460 n=5+5) RegexpMatchHard_1K-8 57.6µs ± 1% 57.0µs ± 2% ~ (p=0.310 n=5+5) Revcomp-8 483ms ± 7% 441ms ± 1% -8.79% (p=0.016 n=5+4) Template-8 58.0ms ± 1% 58.2ms ± 7% ~ (p=0.310 n=5+5) TimeParse-8 324ns ± 6% 312ns ± 2% ~ (p=0.087 n=5+5) TimeFormat-8 330ns ± 1% 329ns ± 1% ~ (p=0.968 n=5+5) name old speed new speed delta GobDecode-8 109MB/s ± 4% 112MB/s ± 1% ~ (p=0.222 n=5+5) GobEncode-8 128MB/s ± 5% 131MB/s ± 2% ~ (p=0.222 n=5+5) Gzip-8 78.9MB/s ± 4% 82.3MB/s ± 1% +4.25% (p=0.008 n=5+5) Gunzip-8 514MB/s ± 4% 521MB/s ± 1% ~ (p=0.841 n=5+5) JSONEncode-8 121MB/s ± 2% 120MB/s ±10% ~ (p=0.548 n=5+5) JSONDecode-8 36.5MB/s ± 2% 36.6MB/s ± 4% ~ (p=1.000 n=5+5) GoParse-8 17.9MB/s ± 2% 17.7MB/s ± 4% ~ (p=0.730 n=5+5) RegexpMatchEasy0_32-8 371MB/s ± 1% 375MB/s ± 3% ~ (p=0.310 n=5+5) RegexpMatchEasy0_1K-8 5.15GB/s ± 1% 5.13GB/s ± 1% ~ (p=0.548 n=5+5) RegexpMatchEasy1_32-8 387MB/s ± 2% 391MB/s ± 1% ~ (p=0.310 n=5+5) RegexpMatchEasy1_1K-8 2.85GB/s ± 2% 2.89GB/s ± 1% ~ (p=0.056 n=5+5) RegexpMatchMedium_32-8 8.07MB/s ± 2% 8.06MB/s ± 1% ~ (p=0.730 n=5+5) RegexpMatchMedium_1K-8 26.8MB/s ± 2% 26.6MB/s ± 7% ~ (p=0.690 n=5+5) RegexpMatchHard_32-8 16.7MB/s ± 2% 16.7MB/s ± 5% ~ (p=0.421 n=5+5) RegexpMatchHard_1K-8 17.8MB/s ± 1% 18.0MB/s ± 2% ~ (p=0.310 n=5+5) Revcomp-8 527MB/s ± 6% 577MB/s ± 1% +9.44% (p=0.016 n=5+4) Template-8 33.5MB/s ± 1% 33.4MB/s ± 7% ~ (p=0.310 n=5+5) Updates #19495 Change-Id: Ib9ece1690813d9b4788455db43d30891e2138df5 Reviewed-on: https://go-review.googlesource.com/38172 Reviewed-by: Hugues Bruant <hugues.bruant@gmail.com> Reviewed-by: Keith Randall <khr@golang.org> Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org>
2017-03-14 12:11:28 -06:00
func iface() interface{}
cmd/compile: convert constants to interfaces without allocating The order pass is responsible for ensuring that values passed to runtime functions, including convT2E/convT2I, are addressable. Prior to this CL, this was always accomplished by creating a temp, which frequently escaped to the heap, causing allocations, perhaps most notably in code like: fmt.Println(1, 2, 3) // allocates three times None of the runtime routines modify the contents of the pointers they receive, so in the case of constants, instead of creating a temp value, we can create a static value. (Marking the static value as read-only provides protection against accidental attempts by the runtime to modify the constant data.) This improves code generation for code like: panic("abc") c <- 2 // c is a chan int which can now simply refer to "abc" and 2, rather than going by way of a temporary. It also allows us to optimize convT2E/convT2I, by recognizing static readonly values and directly constructing the interface. This CL adds ~0.5% to binary size, despite decreasing the size of many functions, because it also adds many static symbols. This binary size regression could be recovered in future (but currently unplanned) work. There is a lot of content-duplication in these symbols; this statement generates six new symbols, three containing an int 1 and three containing a pointer to the string "a": fmt.Println(1, 1, 1, "a", "a", "a") These symbols could be made content-addressable. Furthermore, these symbols are small, so the alignment and naming overhead is large. As with the go.strings section, these symbols could be hidden and have their alignment reduced. The changes to test/live.go make it impossible (at least with current optimization techniques) to place the values being passed to the runtime in static symbols, preserving autotmp creation. Fixes #18704 Benchmarks from fmt and go-kit's logging package: github.com/go-kit/kit/log name old time/op new time/op delta JSONLoggerSimple-8 1.91µs ± 2% 2.11µs ±22% ~ (p=1.000 n=9+10) JSONLoggerContextual-8 2.60µs ± 6% 2.43µs ± 2% -6.29% (p=0.000 n=9+10) Discard-8 101ns ± 2% 34ns ±14% -66.33% (p=0.000 n=10+9) OneWith-8 161ns ± 1% 102ns ±16% -36.78% (p=0.000 n=10+10) TwoWith-8 175ns ± 3% 106ns ± 7% -39.36% (p=0.000 n=10+9) TenWith-8 293ns ± 3% 227ns ±15% -22.44% (p=0.000 n=9+10) LogfmtLoggerSimple-8 704ns ± 2% 608ns ± 2% -13.65% (p=0.000 n=10+9) LogfmtLoggerContextual-8 962ns ± 1% 860ns ±17% -10.57% (p=0.003 n=9+10) NopLoggerSimple-8 188ns ± 1% 120ns ± 1% -36.39% (p=0.000 n=9+10) NopLoggerContextual-8 379ns ± 1% 243ns ± 0% -35.77% (p=0.000 n=9+10) ValueBindingTimestamp-8 577ns ± 1% 499ns ± 1% -13.51% (p=0.000 n=10+10) ValueBindingCaller-8 898ns ± 2% 844ns ± 2% -6.00% (p=0.000 n=10+10) name old alloc/op new alloc/op delta JSONLoggerSimple-8 904B ± 0% 872B ± 0% -3.54% (p=0.000 n=10+10) JSONLoggerContextual-8 1.20kB ± 0% 1.14kB ± 0% -5.33% (p=0.000 n=10+10) Discard-8 64.0B ± 0% 32.0B ± 0% -50.00% (p=0.000 n=10+10) OneWith-8 96.0B ± 0% 64.0B ± 0% -33.33% (p=0.000 n=10+10) TwoWith-8 160B ± 0% 128B ± 0% -20.00% (p=0.000 n=10+10) TenWith-8 672B ± 0% 640B ± 0% -4.76% (p=0.000 n=10+10) LogfmtLoggerSimple-8 128B ± 0% 96B ± 0% -25.00% (p=0.000 n=10+10) LogfmtLoggerContextual-8 304B ± 0% 240B ± 0% -21.05% (p=0.000 n=10+10) NopLoggerSimple-8 128B ± 0% 96B ± 0% -25.00% (p=0.000 n=10+10) NopLoggerContextual-8 304B ± 0% 240B ± 0% -21.05% (p=0.000 n=10+10) ValueBindingTimestamp-8 159B ± 0% 127B ± 0% -20.13% (p=0.000 n=10+10) ValueBindingCaller-8 112B ± 0% 80B ± 0% -28.57% (p=0.000 n=10+10) name old allocs/op new allocs/op delta JSONLoggerSimple-8 19.0 ± 0% 17.0 ± 0% -10.53% (p=0.000 n=10+10) JSONLoggerContextual-8 25.0 ± 0% 21.0 ± 0% -16.00% (p=0.000 n=10+10) Discard-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) OneWith-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) TwoWith-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) TenWith-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) LogfmtLoggerSimple-8 4.00 ± 0% 2.00 ± 0% -50.00% (p=0.000 n=10+10) LogfmtLoggerContextual-8 7.00 ± 0% 3.00 ± 0% -57.14% (p=0.000 n=10+10) NopLoggerSimple-8 4.00 ± 0% 2.00 ± 0% -50.00% (p=0.000 n=10+10) NopLoggerContextual-8 7.00 ± 0% 3.00 ± 0% -57.14% (p=0.000 n=10+10) ValueBindingTimestamp-8 5.00 ± 0% 3.00 ± 0% -40.00% (p=0.000 n=10+10) ValueBindingCaller-8 4.00 ± 0% 2.00 ± 0% -50.00% (p=0.000 n=10+10) fmt name old time/op new time/op delta SprintfPadding-8 88.9ns ± 3% 79.1ns ± 1% -11.09% (p=0.000 n=10+7) SprintfEmpty-8 12.6ns ± 3% 12.8ns ± 3% ~ (p=0.136 n=10+10) SprintfString-8 38.7ns ± 5% 26.9ns ± 6% -30.65% (p=0.000 n=10+10) SprintfTruncateString-8 56.7ns ± 2% 47.0ns ± 3% -17.05% (p=0.000 n=10+10) SprintfQuoteString-8 164ns ± 2% 153ns ± 2% -7.01% (p=0.000 n=10+10) SprintfInt-8 38.9ns ±15% 26.5ns ± 2% -31.93% (p=0.000 n=10+9) SprintfIntInt-8 60.3ns ± 9% 38.2ns ± 1% -36.67% (p=0.000 n=10+8) SprintfPrefixedInt-8 58.6ns ±13% 51.2ns ±11% -12.66% (p=0.001 n=10+10) SprintfFloat-8 71.4ns ± 3% 64.2ns ± 3% -10.08% (p=0.000 n=8+10) SprintfComplex-8 175ns ± 3% 159ns ± 2% -9.03% (p=0.000 n=10+10) SprintfBoolean-8 33.5ns ± 4% 25.7ns ± 5% -23.28% (p=0.000 n=10+10) SprintfHexString-8 65.3ns ± 3% 51.7ns ± 5% -20.86% (p=0.000 n=10+9) SprintfHexBytes-8 67.2ns ± 5% 67.9ns ± 4% ~ (p=0.383 n=10+10) SprintfBytes-8 129ns ± 7% 124ns ± 7% ~ (p=0.074 n=9+10) SprintfStringer-8 127ns ± 4% 126ns ± 8% ~ (p=0.506 n=9+10) SprintfStructure-8 357ns ± 3% 359ns ± 3% ~ (p=0.469 n=10+10) ManyArgs-8 203ns ± 6% 126ns ± 3% -37.94% (p=0.000 n=10+10) FprintInt-8 119ns ±10% 74ns ± 3% -37.54% (p=0.000 n=10+10) FprintfBytes-8 122ns ± 4% 120ns ± 3% ~ (p=0.124 n=10+10) FprintIntNoAlloc-8 78.2ns ± 5% 74.1ns ± 3% -5.28% (p=0.000 n=10+10) ScanInts-8 349µs ± 1% 349µs ± 0% ~ (p=0.606 n=9+8) ScanRecursiveInt-8 43.8ms ± 7% 40.1ms ± 2% -8.42% (p=0.000 n=10+10) ScanRecursiveIntReaderWrapper-8 43.5ms ± 4% 40.4ms ± 2% -7.16% (p=0.000 n=10+9) name old alloc/op new alloc/op delta SprintfPadding-8 24.0B ± 0% 16.0B ± 0% -33.33% (p=0.000 n=10+10) SprintfEmpty-8 0.00B 0.00B ~ (all equal) SprintfString-8 21.0B ± 0% 5.0B ± 0% -76.19% (p=0.000 n=10+10) SprintfTruncateString-8 32.0B ± 0% 16.0B ± 0% -50.00% (p=0.000 n=10+10) SprintfQuoteString-8 48.0B ± 0% 32.0B ± 0% -33.33% (p=0.000 n=10+10) SprintfInt-8 16.0B ± 0% 1.0B ± 0% -93.75% (p=0.000 n=10+10) SprintfIntInt-8 24.0B ± 0% 3.0B ± 0% -87.50% (p=0.000 n=10+10) SprintfPrefixedInt-8 72.0B ± 0% 64.0B ± 0% -11.11% (p=0.000 n=10+10) SprintfFloat-8 16.0B ± 0% 8.0B ± 0% -50.00% (p=0.000 n=10+10) SprintfComplex-8 48.0B ± 0% 32.0B ± 0% -33.33% (p=0.000 n=10+10) SprintfBoolean-8 8.00B ± 0% 4.00B ± 0% -50.00% (p=0.000 n=10+10) SprintfHexString-8 96.0B ± 0% 80.0B ± 0% -16.67% (p=0.000 n=10+10) SprintfHexBytes-8 112B ± 0% 112B ± 0% ~ (all equal) SprintfBytes-8 96.0B ± 0% 96.0B ± 0% ~ (all equal) SprintfStringer-8 32.0B ± 0% 32.0B ± 0% ~ (all equal) SprintfStructure-8 256B ± 0% 256B ± 0% ~ (all equal) ManyArgs-8 80.0B ± 0% 0.0B -100.00% (p=0.000 n=10+10) FprintInt-8 8.00B ± 0% 0.00B -100.00% (p=0.000 n=10+10) FprintfBytes-8 32.0B ± 0% 32.0B ± 0% ~ (all equal) FprintIntNoAlloc-8 0.00B 0.00B ~ (all equal) ScanInts-8 15.2kB ± 0% 15.2kB ± 0% ~ (p=0.248 n=9+10) ScanRecursiveInt-8 21.6kB ± 0% 21.6kB ± 0% ~ (all equal) ScanRecursiveIntReaderWrapper-8 21.7kB ± 0% 21.7kB ± 0% ~ (all equal) name old allocs/op new allocs/op delta SprintfPadding-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfEmpty-8 0.00 0.00 ~ (all equal) SprintfString-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfTruncateString-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfQuoteString-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfInt-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfIntInt-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) SprintfPrefixedInt-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfFloat-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfComplex-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfBoolean-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfHexString-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfHexBytes-8 2.00 ± 0% 2.00 ± 0% ~ (all equal) SprintfBytes-8 2.00 ± 0% 2.00 ± 0% ~ (all equal) SprintfStringer-8 4.00 ± 0% 4.00 ± 0% ~ (all equal) SprintfStructure-8 7.00 ± 0% 7.00 ± 0% ~ (all equal) ManyArgs-8 8.00 ± 0% 0.00 -100.00% (p=0.000 n=10+10) FprintInt-8 1.00 ± 0% 0.00 -100.00% (p=0.000 n=10+10) FprintfBytes-8 1.00 ± 0% 1.00 ± 0% ~ (all equal) FprintIntNoAlloc-8 0.00 0.00 ~ (all equal) ScanInts-8 1.60k ± 0% 1.60k ± 0% ~ (all equal) ScanRecursiveInt-8 1.71k ± 0% 1.71k ± 0% ~ (all equal) ScanRecursiveIntReaderWrapper-8 1.71k ± 0% 1.71k ± 0% ~ (all equal) Change-Id: I7ba72a25fea4140a0ba40a9f443103ed87cc69b5 Reviewed-on: https://go-review.googlesource.com/35554 Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Keith Randall <khr@golang.org>
2017-01-21 14:41:06 -07:00
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
func f16() {
if b {
delete(mi, iface()) // ERROR "stack object .autotmp_[0-9]+ interface \{\}$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
delete(mi, iface())
delete(mi, iface())
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
var m2s map[string]*byte
var m2 map[[2]string]*byte
var x2 [2]string
var bp *byte
func f17a(p *byte) { // ERROR "live at entry to f17a: p$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
if b {
m2[x2] = p // ERROR "live at call to mapassign: p$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
m2[x2] = p // ERROR "live at call to mapassign: p$"
m2[x2] = p // ERROR "live at call to mapassign: p$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
func f17b(p *byte) { // ERROR "live at entry to f17b: p$"
// key temporary
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
if b {
runtime: add mapassign_fast* Add benchmarks for map assignment with int32/int64/string key Benchmark results on darwin/amd64 name old time/op new time/op delta MapAssignInt32_255-8 24.7ns ± 3% 17.4ns ± 2% -29.75% (p=0.000 n=10+10) MapAssignInt32_64k-8 45.5ns ± 4% 37.6ns ± 4% -17.18% (p=0.000 n=10+10) MapAssignInt64_255-8 26.0ns ± 3% 17.9ns ± 4% -31.03% (p=0.000 n=10+10) MapAssignInt64_64k-8 46.9ns ± 5% 38.7ns ± 2% -17.53% (p=0.000 n=9+10) MapAssignStr_255-8 47.8ns ± 3% 24.8ns ± 4% -48.01% (p=0.000 n=10+10) MapAssignStr_64k-8 83.0ns ± 3% 51.9ns ± 3% -37.45% (p=0.000 n=10+9) name old time/op new time/op delta BinaryTree17-8 3.11s ±19% 2.78s ± 3% ~ (p=0.095 n=5+5) Fannkuch11-8 3.26s ± 1% 3.21s ± 2% ~ (p=0.056 n=5+5) FmtFprintfEmpty-8 50.3ns ± 1% 50.8ns ± 2% ~ (p=0.246 n=5+5) FmtFprintfString-8 82.7ns ± 4% 80.1ns ± 5% ~ (p=0.238 n=5+5) FmtFprintfInt-8 82.6ns ± 2% 81.9ns ± 3% ~ (p=0.508 n=5+5) FmtFprintfIntInt-8 124ns ± 4% 121ns ± 3% ~ (p=0.111 n=5+5) FmtFprintfPrefixedInt-8 158ns ± 6% 160ns ± 2% ~ (p=0.341 n=5+5) FmtFprintfFloat-8 249ns ± 2% 245ns ± 2% ~ (p=0.095 n=5+5) FmtManyArgs-8 513ns ± 2% 519ns ± 3% ~ (p=0.151 n=5+5) GobDecode-8 7.48ms ±12% 7.11ms ± 2% ~ (p=0.222 n=5+5) GobEncode-8 6.25ms ± 1% 6.03ms ± 2% -3.56% (p=0.008 n=5+5) Gzip-8 252ms ± 4% 252ms ± 4% ~ (p=1.000 n=5+5) Gunzip-8 38.4ms ± 3% 38.6ms ± 2% ~ (p=0.690 n=5+5) HTTPClientServer-8 76.9µs ±41% 66.4µs ± 6% ~ (p=0.310 n=5+5) JSONEncode-8 16.5ms ± 3% 16.7ms ± 3% ~ (p=0.421 n=5+5) JSONDecode-8 54.6ms ± 1% 54.3ms ± 2% ~ (p=0.548 n=5+5) Mandelbrot200-8 4.45ms ± 3% 4.47ms ± 1% ~ (p=0.841 n=5+5) GoParse-8 3.43ms ± 1% 3.32ms ± 2% -3.28% (p=0.008 n=5+5) RegexpMatchEasy0_32-8 88.2ns ± 3% 89.4ns ± 2% ~ (p=0.333 n=5+5) RegexpMatchEasy0_1K-8 205ns ± 1% 206ns ± 1% ~ (p=0.905 n=5+5) RegexpMatchEasy1_32-8 85.1ns ± 1% 85.5ns ± 5% ~ (p=0.690 n=5+5) RegexpMatchEasy1_1K-8 365ns ± 1% 371ns ± 9% ~ (p=1.000 n=5+5) RegexpMatchMedium_32-8 129ns ± 2% 128ns ± 3% ~ (p=0.730 n=5+5) RegexpMatchMedium_1K-8 39.8µs ± 0% 39.7µs ± 4% ~ (p=0.730 n=4+5) RegexpMatchHard_32-8 1.99µs ± 3% 2.05µs ±16% ~ (p=0.794 n=5+5) RegexpMatchHard_1K-8 59.3µs ± 1% 60.3µs ± 7% ~ (p=1.000 n=5+5) Revcomp-8 1.36s ±63% 0.52s ± 5% ~ (p=0.095 n=5+5) Template-8 62.6ms ±14% 60.5ms ± 5% ~ (p=0.690 n=5+5) TimeParse-8 330ns ± 2% 324ns ± 2% ~ (p=0.087 n=5+5) TimeFormat-8 350ns ± 3% 340ns ± 1% -2.86% (p=0.008 n=5+5) name old speed new speed delta GobDecode-8 103MB/s ±11% 108MB/s ± 2% ~ (p=0.222 n=5+5) GobEncode-8 123MB/s ± 1% 127MB/s ± 2% +3.71% (p=0.008 n=5+5) Gzip-8 77.1MB/s ± 4% 76.9MB/s ± 3% ~ (p=1.000 n=5+5) Gunzip-8 505MB/s ± 3% 503MB/s ± 2% ~ (p=0.690 n=5+5) JSONEncode-8 118MB/s ± 3% 116MB/s ± 3% ~ (p=0.421 n=5+5) JSONDecode-8 35.5MB/s ± 1% 35.8MB/s ± 2% ~ (p=0.397 n=5+5) GoParse-8 16.9MB/s ± 1% 17.4MB/s ± 2% +3.45% (p=0.008 n=5+5) RegexpMatchEasy0_32-8 363MB/s ± 3% 358MB/s ± 2% ~ (p=0.421 n=5+5) RegexpMatchEasy0_1K-8 4.98GB/s ± 1% 4.97GB/s ± 1% ~ (p=0.548 n=5+5) RegexpMatchEasy1_32-8 376MB/s ± 1% 375MB/s ± 5% ~ (p=0.690 n=5+5) RegexpMatchEasy1_1K-8 2.80GB/s ± 1% 2.76GB/s ± 9% ~ (p=0.841 n=5+5) RegexpMatchMedium_32-8 7.73MB/s ± 1% 7.76MB/s ± 3% ~ (p=0.730 n=5+5) RegexpMatchMedium_1K-8 25.8MB/s ± 0% 25.8MB/s ± 4% ~ (p=0.651 n=4+5) RegexpMatchHard_32-8 16.1MB/s ± 3% 15.7MB/s ±14% ~ (p=0.794 n=5+5) RegexpMatchHard_1K-8 17.3MB/s ± 1% 17.0MB/s ± 7% ~ (p=0.984 n=5+5) Revcomp-8 273MB/s ±83% 488MB/s ± 5% ~ (p=0.095 n=5+5) Template-8 31.1MB/s ±13% 32.1MB/s ± 5% ~ (p=0.690 n=5+5) Updates #19495 Change-Id: I116e9a2a4594769318b22d736464de8a98499909 Reviewed-on: https://go-review.googlesource.com/38091 Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com> Reviewed-by: Keith Randall <khr@golang.org> Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com> TryBot-Result: Gobot Gobot <gobot@golang.org>
2017-03-12 15:47:59 -06:00
m2s[str()] = p // ERROR "live at call to mapassign_faststr: p$" "live at call to str: p$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
runtime: add mapassign_fast* Add benchmarks for map assignment with int32/int64/string key Benchmark results on darwin/amd64 name old time/op new time/op delta MapAssignInt32_255-8 24.7ns ± 3% 17.4ns ± 2% -29.75% (p=0.000 n=10+10) MapAssignInt32_64k-8 45.5ns ± 4% 37.6ns ± 4% -17.18% (p=0.000 n=10+10) MapAssignInt64_255-8 26.0ns ± 3% 17.9ns ± 4% -31.03% (p=0.000 n=10+10) MapAssignInt64_64k-8 46.9ns ± 5% 38.7ns ± 2% -17.53% (p=0.000 n=9+10) MapAssignStr_255-8 47.8ns ± 3% 24.8ns ± 4% -48.01% (p=0.000 n=10+10) MapAssignStr_64k-8 83.0ns ± 3% 51.9ns ± 3% -37.45% (p=0.000 n=10+9) name old time/op new time/op delta BinaryTree17-8 3.11s ±19% 2.78s ± 3% ~ (p=0.095 n=5+5) Fannkuch11-8 3.26s ± 1% 3.21s ± 2% ~ (p=0.056 n=5+5) FmtFprintfEmpty-8 50.3ns ± 1% 50.8ns ± 2% ~ (p=0.246 n=5+5) FmtFprintfString-8 82.7ns ± 4% 80.1ns ± 5% ~ (p=0.238 n=5+5) FmtFprintfInt-8 82.6ns ± 2% 81.9ns ± 3% ~ (p=0.508 n=5+5) FmtFprintfIntInt-8 124ns ± 4% 121ns ± 3% ~ (p=0.111 n=5+5) FmtFprintfPrefixedInt-8 158ns ± 6% 160ns ± 2% ~ (p=0.341 n=5+5) FmtFprintfFloat-8 249ns ± 2% 245ns ± 2% ~ (p=0.095 n=5+5) FmtManyArgs-8 513ns ± 2% 519ns ± 3% ~ (p=0.151 n=5+5) GobDecode-8 7.48ms ±12% 7.11ms ± 2% ~ (p=0.222 n=5+5) GobEncode-8 6.25ms ± 1% 6.03ms ± 2% -3.56% (p=0.008 n=5+5) Gzip-8 252ms ± 4% 252ms ± 4% ~ (p=1.000 n=5+5) Gunzip-8 38.4ms ± 3% 38.6ms ± 2% ~ (p=0.690 n=5+5) HTTPClientServer-8 76.9µs ±41% 66.4µs ± 6% ~ (p=0.310 n=5+5) JSONEncode-8 16.5ms ± 3% 16.7ms ± 3% ~ (p=0.421 n=5+5) JSONDecode-8 54.6ms ± 1% 54.3ms ± 2% ~ (p=0.548 n=5+5) Mandelbrot200-8 4.45ms ± 3% 4.47ms ± 1% ~ (p=0.841 n=5+5) GoParse-8 3.43ms ± 1% 3.32ms ± 2% -3.28% (p=0.008 n=5+5) RegexpMatchEasy0_32-8 88.2ns ± 3% 89.4ns ± 2% ~ (p=0.333 n=5+5) RegexpMatchEasy0_1K-8 205ns ± 1% 206ns ± 1% ~ (p=0.905 n=5+5) RegexpMatchEasy1_32-8 85.1ns ± 1% 85.5ns ± 5% ~ (p=0.690 n=5+5) RegexpMatchEasy1_1K-8 365ns ± 1% 371ns ± 9% ~ (p=1.000 n=5+5) RegexpMatchMedium_32-8 129ns ± 2% 128ns ± 3% ~ (p=0.730 n=5+5) RegexpMatchMedium_1K-8 39.8µs ± 0% 39.7µs ± 4% ~ (p=0.730 n=4+5) RegexpMatchHard_32-8 1.99µs ± 3% 2.05µs ±16% ~ (p=0.794 n=5+5) RegexpMatchHard_1K-8 59.3µs ± 1% 60.3µs ± 7% ~ (p=1.000 n=5+5) Revcomp-8 1.36s ±63% 0.52s ± 5% ~ (p=0.095 n=5+5) Template-8 62.6ms ±14% 60.5ms ± 5% ~ (p=0.690 n=5+5) TimeParse-8 330ns ± 2% 324ns ± 2% ~ (p=0.087 n=5+5) TimeFormat-8 350ns ± 3% 340ns ± 1% -2.86% (p=0.008 n=5+5) name old speed new speed delta GobDecode-8 103MB/s ±11% 108MB/s ± 2% ~ (p=0.222 n=5+5) GobEncode-8 123MB/s ± 1% 127MB/s ± 2% +3.71% (p=0.008 n=5+5) Gzip-8 77.1MB/s ± 4% 76.9MB/s ± 3% ~ (p=1.000 n=5+5) Gunzip-8 505MB/s ± 3% 503MB/s ± 2% ~ (p=0.690 n=5+5) JSONEncode-8 118MB/s ± 3% 116MB/s ± 3% ~ (p=0.421 n=5+5) JSONDecode-8 35.5MB/s ± 1% 35.8MB/s ± 2% ~ (p=0.397 n=5+5) GoParse-8 16.9MB/s ± 1% 17.4MB/s ± 2% +3.45% (p=0.008 n=5+5) RegexpMatchEasy0_32-8 363MB/s ± 3% 358MB/s ± 2% ~ (p=0.421 n=5+5) RegexpMatchEasy0_1K-8 4.98GB/s ± 1% 4.97GB/s ± 1% ~ (p=0.548 n=5+5) RegexpMatchEasy1_32-8 376MB/s ± 1% 375MB/s ± 5% ~ (p=0.690 n=5+5) RegexpMatchEasy1_1K-8 2.80GB/s ± 1% 2.76GB/s ± 9% ~ (p=0.841 n=5+5) RegexpMatchMedium_32-8 7.73MB/s ± 1% 7.76MB/s ± 3% ~ (p=0.730 n=5+5) RegexpMatchMedium_1K-8 25.8MB/s ± 0% 25.8MB/s ± 4% ~ (p=0.651 n=4+5) RegexpMatchHard_32-8 16.1MB/s ± 3% 15.7MB/s ±14% ~ (p=0.794 n=5+5) RegexpMatchHard_1K-8 17.3MB/s ± 1% 17.0MB/s ± 7% ~ (p=0.984 n=5+5) Revcomp-8 273MB/s ±83% 488MB/s ± 5% ~ (p=0.095 n=5+5) Template-8 31.1MB/s ±13% 32.1MB/s ± 5% ~ (p=0.690 n=5+5) Updates #19495 Change-Id: I116e9a2a4594769318b22d736464de8a98499909 Reviewed-on: https://go-review.googlesource.com/38091 Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com> Reviewed-by: Keith Randall <khr@golang.org> Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com> TryBot-Result: Gobot Gobot <gobot@golang.org>
2017-03-12 15:47:59 -06:00
m2s[str()] = p // ERROR "live at call to mapassign_faststr: p$" "live at call to str: p$"
m2s[str()] = p // ERROR "live at call to mapassign_faststr: p$" "live at call to str: p$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
func f17c() {
// key and value temporaries
if b {
runtime: add mapassign_fast* Add benchmarks for map assignment with int32/int64/string key Benchmark results on darwin/amd64 name old time/op new time/op delta MapAssignInt32_255-8 24.7ns ± 3% 17.4ns ± 2% -29.75% (p=0.000 n=10+10) MapAssignInt32_64k-8 45.5ns ± 4% 37.6ns ± 4% -17.18% (p=0.000 n=10+10) MapAssignInt64_255-8 26.0ns ± 3% 17.9ns ± 4% -31.03% (p=0.000 n=10+10) MapAssignInt64_64k-8 46.9ns ± 5% 38.7ns ± 2% -17.53% (p=0.000 n=9+10) MapAssignStr_255-8 47.8ns ± 3% 24.8ns ± 4% -48.01% (p=0.000 n=10+10) MapAssignStr_64k-8 83.0ns ± 3% 51.9ns ± 3% -37.45% (p=0.000 n=10+9) name old time/op new time/op delta BinaryTree17-8 3.11s ±19% 2.78s ± 3% ~ (p=0.095 n=5+5) Fannkuch11-8 3.26s ± 1% 3.21s ± 2% ~ (p=0.056 n=5+5) FmtFprintfEmpty-8 50.3ns ± 1% 50.8ns ± 2% ~ (p=0.246 n=5+5) FmtFprintfString-8 82.7ns ± 4% 80.1ns ± 5% ~ (p=0.238 n=5+5) FmtFprintfInt-8 82.6ns ± 2% 81.9ns ± 3% ~ (p=0.508 n=5+5) FmtFprintfIntInt-8 124ns ± 4% 121ns ± 3% ~ (p=0.111 n=5+5) FmtFprintfPrefixedInt-8 158ns ± 6% 160ns ± 2% ~ (p=0.341 n=5+5) FmtFprintfFloat-8 249ns ± 2% 245ns ± 2% ~ (p=0.095 n=5+5) FmtManyArgs-8 513ns ± 2% 519ns ± 3% ~ (p=0.151 n=5+5) GobDecode-8 7.48ms ±12% 7.11ms ± 2% ~ (p=0.222 n=5+5) GobEncode-8 6.25ms ± 1% 6.03ms ± 2% -3.56% (p=0.008 n=5+5) Gzip-8 252ms ± 4% 252ms ± 4% ~ (p=1.000 n=5+5) Gunzip-8 38.4ms ± 3% 38.6ms ± 2% ~ (p=0.690 n=5+5) HTTPClientServer-8 76.9µs ±41% 66.4µs ± 6% ~ (p=0.310 n=5+5) JSONEncode-8 16.5ms ± 3% 16.7ms ± 3% ~ (p=0.421 n=5+5) JSONDecode-8 54.6ms ± 1% 54.3ms ± 2% ~ (p=0.548 n=5+5) Mandelbrot200-8 4.45ms ± 3% 4.47ms ± 1% ~ (p=0.841 n=5+5) GoParse-8 3.43ms ± 1% 3.32ms ± 2% -3.28% (p=0.008 n=5+5) RegexpMatchEasy0_32-8 88.2ns ± 3% 89.4ns ± 2% ~ (p=0.333 n=5+5) RegexpMatchEasy0_1K-8 205ns ± 1% 206ns ± 1% ~ (p=0.905 n=5+5) RegexpMatchEasy1_32-8 85.1ns ± 1% 85.5ns ± 5% ~ (p=0.690 n=5+5) RegexpMatchEasy1_1K-8 365ns ± 1% 371ns ± 9% ~ (p=1.000 n=5+5) RegexpMatchMedium_32-8 129ns ± 2% 128ns ± 3% ~ (p=0.730 n=5+5) RegexpMatchMedium_1K-8 39.8µs ± 0% 39.7µs ± 4% ~ (p=0.730 n=4+5) RegexpMatchHard_32-8 1.99µs ± 3% 2.05µs ±16% ~ (p=0.794 n=5+5) RegexpMatchHard_1K-8 59.3µs ± 1% 60.3µs ± 7% ~ (p=1.000 n=5+5) Revcomp-8 1.36s ±63% 0.52s ± 5% ~ (p=0.095 n=5+5) Template-8 62.6ms ±14% 60.5ms ± 5% ~ (p=0.690 n=5+5) TimeParse-8 330ns ± 2% 324ns ± 2% ~ (p=0.087 n=5+5) TimeFormat-8 350ns ± 3% 340ns ± 1% -2.86% (p=0.008 n=5+5) name old speed new speed delta GobDecode-8 103MB/s ±11% 108MB/s ± 2% ~ (p=0.222 n=5+5) GobEncode-8 123MB/s ± 1% 127MB/s ± 2% +3.71% (p=0.008 n=5+5) Gzip-8 77.1MB/s ± 4% 76.9MB/s ± 3% ~ (p=1.000 n=5+5) Gunzip-8 505MB/s ± 3% 503MB/s ± 2% ~ (p=0.690 n=5+5) JSONEncode-8 118MB/s ± 3% 116MB/s ± 3% ~ (p=0.421 n=5+5) JSONDecode-8 35.5MB/s ± 1% 35.8MB/s ± 2% ~ (p=0.397 n=5+5) GoParse-8 16.9MB/s ± 1% 17.4MB/s ± 2% +3.45% (p=0.008 n=5+5) RegexpMatchEasy0_32-8 363MB/s ± 3% 358MB/s ± 2% ~ (p=0.421 n=5+5) RegexpMatchEasy0_1K-8 4.98GB/s ± 1% 4.97GB/s ± 1% ~ (p=0.548 n=5+5) RegexpMatchEasy1_32-8 376MB/s ± 1% 375MB/s ± 5% ~ (p=0.690 n=5+5) RegexpMatchEasy1_1K-8 2.80GB/s ± 1% 2.76GB/s ± 9% ~ (p=0.841 n=5+5) RegexpMatchMedium_32-8 7.73MB/s ± 1% 7.76MB/s ± 3% ~ (p=0.730 n=5+5) RegexpMatchMedium_1K-8 25.8MB/s ± 0% 25.8MB/s ± 4% ~ (p=0.651 n=4+5) RegexpMatchHard_32-8 16.1MB/s ± 3% 15.7MB/s ±14% ~ (p=0.794 n=5+5) RegexpMatchHard_1K-8 17.3MB/s ± 1% 17.0MB/s ± 7% ~ (p=0.984 n=5+5) Revcomp-8 273MB/s ±83% 488MB/s ± 5% ~ (p=0.095 n=5+5) Template-8 31.1MB/s ±13% 32.1MB/s ± 5% ~ (p=0.690 n=5+5) Updates #19495 Change-Id: I116e9a2a4594769318b22d736464de8a98499909 Reviewed-on: https://go-review.googlesource.com/38091 Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com> Reviewed-by: Keith Randall <khr@golang.org> Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com> TryBot-Result: Gobot Gobot <gobot@golang.org>
2017-03-12 15:47:59 -06:00
m2s[str()] = f17d() // ERROR "live at call to f17d: .autotmp_[0-9]+$" "live at call to mapassign_faststr: .autotmp_[0-9]+$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
runtime: add mapassign_fast* Add benchmarks for map assignment with int32/int64/string key Benchmark results on darwin/amd64 name old time/op new time/op delta MapAssignInt32_255-8 24.7ns ± 3% 17.4ns ± 2% -29.75% (p=0.000 n=10+10) MapAssignInt32_64k-8 45.5ns ± 4% 37.6ns ± 4% -17.18% (p=0.000 n=10+10) MapAssignInt64_255-8 26.0ns ± 3% 17.9ns ± 4% -31.03% (p=0.000 n=10+10) MapAssignInt64_64k-8 46.9ns ± 5% 38.7ns ± 2% -17.53% (p=0.000 n=9+10) MapAssignStr_255-8 47.8ns ± 3% 24.8ns ± 4% -48.01% (p=0.000 n=10+10) MapAssignStr_64k-8 83.0ns ± 3% 51.9ns ± 3% -37.45% (p=0.000 n=10+9) name old time/op new time/op delta BinaryTree17-8 3.11s ±19% 2.78s ± 3% ~ (p=0.095 n=5+5) Fannkuch11-8 3.26s ± 1% 3.21s ± 2% ~ (p=0.056 n=5+5) FmtFprintfEmpty-8 50.3ns ± 1% 50.8ns ± 2% ~ (p=0.246 n=5+5) FmtFprintfString-8 82.7ns ± 4% 80.1ns ± 5% ~ (p=0.238 n=5+5) FmtFprintfInt-8 82.6ns ± 2% 81.9ns ± 3% ~ (p=0.508 n=5+5) FmtFprintfIntInt-8 124ns ± 4% 121ns ± 3% ~ (p=0.111 n=5+5) FmtFprintfPrefixedInt-8 158ns ± 6% 160ns ± 2% ~ (p=0.341 n=5+5) FmtFprintfFloat-8 249ns ± 2% 245ns ± 2% ~ (p=0.095 n=5+5) FmtManyArgs-8 513ns ± 2% 519ns ± 3% ~ (p=0.151 n=5+5) GobDecode-8 7.48ms ±12% 7.11ms ± 2% ~ (p=0.222 n=5+5) GobEncode-8 6.25ms ± 1% 6.03ms ± 2% -3.56% (p=0.008 n=5+5) Gzip-8 252ms ± 4% 252ms ± 4% ~ (p=1.000 n=5+5) Gunzip-8 38.4ms ± 3% 38.6ms ± 2% ~ (p=0.690 n=5+5) HTTPClientServer-8 76.9µs ±41% 66.4µs ± 6% ~ (p=0.310 n=5+5) JSONEncode-8 16.5ms ± 3% 16.7ms ± 3% ~ (p=0.421 n=5+5) JSONDecode-8 54.6ms ± 1% 54.3ms ± 2% ~ (p=0.548 n=5+5) Mandelbrot200-8 4.45ms ± 3% 4.47ms ± 1% ~ (p=0.841 n=5+5) GoParse-8 3.43ms ± 1% 3.32ms ± 2% -3.28% (p=0.008 n=5+5) RegexpMatchEasy0_32-8 88.2ns ± 3% 89.4ns ± 2% ~ (p=0.333 n=5+5) RegexpMatchEasy0_1K-8 205ns ± 1% 206ns ± 1% ~ (p=0.905 n=5+5) RegexpMatchEasy1_32-8 85.1ns ± 1% 85.5ns ± 5% ~ (p=0.690 n=5+5) RegexpMatchEasy1_1K-8 365ns ± 1% 371ns ± 9% ~ (p=1.000 n=5+5) RegexpMatchMedium_32-8 129ns ± 2% 128ns ± 3% ~ (p=0.730 n=5+5) RegexpMatchMedium_1K-8 39.8µs ± 0% 39.7µs ± 4% ~ (p=0.730 n=4+5) RegexpMatchHard_32-8 1.99µs ± 3% 2.05µs ±16% ~ (p=0.794 n=5+5) RegexpMatchHard_1K-8 59.3µs ± 1% 60.3µs ± 7% ~ (p=1.000 n=5+5) Revcomp-8 1.36s ±63% 0.52s ± 5% ~ (p=0.095 n=5+5) Template-8 62.6ms ±14% 60.5ms ± 5% ~ (p=0.690 n=5+5) TimeParse-8 330ns ± 2% 324ns ± 2% ~ (p=0.087 n=5+5) TimeFormat-8 350ns ± 3% 340ns ± 1% -2.86% (p=0.008 n=5+5) name old speed new speed delta GobDecode-8 103MB/s ±11% 108MB/s ± 2% ~ (p=0.222 n=5+5) GobEncode-8 123MB/s ± 1% 127MB/s ± 2% +3.71% (p=0.008 n=5+5) Gzip-8 77.1MB/s ± 4% 76.9MB/s ± 3% ~ (p=1.000 n=5+5) Gunzip-8 505MB/s ± 3% 503MB/s ± 2% ~ (p=0.690 n=5+5) JSONEncode-8 118MB/s ± 3% 116MB/s ± 3% ~ (p=0.421 n=5+5) JSONDecode-8 35.5MB/s ± 1% 35.8MB/s ± 2% ~ (p=0.397 n=5+5) GoParse-8 16.9MB/s ± 1% 17.4MB/s ± 2% +3.45% (p=0.008 n=5+5) RegexpMatchEasy0_32-8 363MB/s ± 3% 358MB/s ± 2% ~ (p=0.421 n=5+5) RegexpMatchEasy0_1K-8 4.98GB/s ± 1% 4.97GB/s ± 1% ~ (p=0.548 n=5+5) RegexpMatchEasy1_32-8 376MB/s ± 1% 375MB/s ± 5% ~ (p=0.690 n=5+5) RegexpMatchEasy1_1K-8 2.80GB/s ± 1% 2.76GB/s ± 9% ~ (p=0.841 n=5+5) RegexpMatchMedium_32-8 7.73MB/s ± 1% 7.76MB/s ± 3% ~ (p=0.730 n=5+5) RegexpMatchMedium_1K-8 25.8MB/s ± 0% 25.8MB/s ± 4% ~ (p=0.651 n=4+5) RegexpMatchHard_32-8 16.1MB/s ± 3% 15.7MB/s ±14% ~ (p=0.794 n=5+5) RegexpMatchHard_1K-8 17.3MB/s ± 1% 17.0MB/s ± 7% ~ (p=0.984 n=5+5) Revcomp-8 273MB/s ±83% 488MB/s ± 5% ~ (p=0.095 n=5+5) Template-8 31.1MB/s ±13% 32.1MB/s ± 5% ~ (p=0.690 n=5+5) Updates #19495 Change-Id: I116e9a2a4594769318b22d736464de8a98499909 Reviewed-on: https://go-review.googlesource.com/38091 Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com> Reviewed-by: Keith Randall <khr@golang.org> Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com> TryBot-Result: Gobot Gobot <gobot@golang.org>
2017-03-12 15:47:59 -06:00
m2s[str()] = f17d() // ERROR "live at call to f17d: .autotmp_[0-9]+$" "live at call to mapassign_faststr: .autotmp_[0-9]+$"
m2s[str()] = f17d() // ERROR "live at call to f17d: .autotmp_[0-9]+$" "live at call to mapassign_faststr: .autotmp_[0-9]+$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
func f17d() *byte
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
func g18() [2]string
func f18() {
// key temporary for mapaccess.
// temporary introduced by orderexpr.
var z *byte
if b {
z = m2[g18()] // ERROR "stack object .autotmp_[0-9]+ \[2\]string$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
z = m2[g18()]
z = m2[g18()]
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printbytepointer(z)
}
var ch chan *byte
cmd/compile: convert constants to interfaces without allocating The order pass is responsible for ensuring that values passed to runtime functions, including convT2E/convT2I, are addressable. Prior to this CL, this was always accomplished by creating a temp, which frequently escaped to the heap, causing allocations, perhaps most notably in code like: fmt.Println(1, 2, 3) // allocates three times None of the runtime routines modify the contents of the pointers they receive, so in the case of constants, instead of creating a temp value, we can create a static value. (Marking the static value as read-only provides protection against accidental attempts by the runtime to modify the constant data.) This improves code generation for code like: panic("abc") c <- 2 // c is a chan int which can now simply refer to "abc" and 2, rather than going by way of a temporary. It also allows us to optimize convT2E/convT2I, by recognizing static readonly values and directly constructing the interface. This CL adds ~0.5% to binary size, despite decreasing the size of many functions, because it also adds many static symbols. This binary size regression could be recovered in future (but currently unplanned) work. There is a lot of content-duplication in these symbols; this statement generates six new symbols, three containing an int 1 and three containing a pointer to the string "a": fmt.Println(1, 1, 1, "a", "a", "a") These symbols could be made content-addressable. Furthermore, these symbols are small, so the alignment and naming overhead is large. As with the go.strings section, these symbols could be hidden and have their alignment reduced. The changes to test/live.go make it impossible (at least with current optimization techniques) to place the values being passed to the runtime in static symbols, preserving autotmp creation. Fixes #18704 Benchmarks from fmt and go-kit's logging package: github.com/go-kit/kit/log name old time/op new time/op delta JSONLoggerSimple-8 1.91µs ± 2% 2.11µs ±22% ~ (p=1.000 n=9+10) JSONLoggerContextual-8 2.60µs ± 6% 2.43µs ± 2% -6.29% (p=0.000 n=9+10) Discard-8 101ns ± 2% 34ns ±14% -66.33% (p=0.000 n=10+9) OneWith-8 161ns ± 1% 102ns ±16% -36.78% (p=0.000 n=10+10) TwoWith-8 175ns ± 3% 106ns ± 7% -39.36% (p=0.000 n=10+9) TenWith-8 293ns ± 3% 227ns ±15% -22.44% (p=0.000 n=9+10) LogfmtLoggerSimple-8 704ns ± 2% 608ns ± 2% -13.65% (p=0.000 n=10+9) LogfmtLoggerContextual-8 962ns ± 1% 860ns ±17% -10.57% (p=0.003 n=9+10) NopLoggerSimple-8 188ns ± 1% 120ns ± 1% -36.39% (p=0.000 n=9+10) NopLoggerContextual-8 379ns ± 1% 243ns ± 0% -35.77% (p=0.000 n=9+10) ValueBindingTimestamp-8 577ns ± 1% 499ns ± 1% -13.51% (p=0.000 n=10+10) ValueBindingCaller-8 898ns ± 2% 844ns ± 2% -6.00% (p=0.000 n=10+10) name old alloc/op new alloc/op delta JSONLoggerSimple-8 904B ± 0% 872B ± 0% -3.54% (p=0.000 n=10+10) JSONLoggerContextual-8 1.20kB ± 0% 1.14kB ± 0% -5.33% (p=0.000 n=10+10) Discard-8 64.0B ± 0% 32.0B ± 0% -50.00% (p=0.000 n=10+10) OneWith-8 96.0B ± 0% 64.0B ± 0% -33.33% (p=0.000 n=10+10) TwoWith-8 160B ± 0% 128B ± 0% -20.00% (p=0.000 n=10+10) TenWith-8 672B ± 0% 640B ± 0% -4.76% (p=0.000 n=10+10) LogfmtLoggerSimple-8 128B ± 0% 96B ± 0% -25.00% (p=0.000 n=10+10) LogfmtLoggerContextual-8 304B ± 0% 240B ± 0% -21.05% (p=0.000 n=10+10) NopLoggerSimple-8 128B ± 0% 96B ± 0% -25.00% (p=0.000 n=10+10) NopLoggerContextual-8 304B ± 0% 240B ± 0% -21.05% (p=0.000 n=10+10) ValueBindingTimestamp-8 159B ± 0% 127B ± 0% -20.13% (p=0.000 n=10+10) ValueBindingCaller-8 112B ± 0% 80B ± 0% -28.57% (p=0.000 n=10+10) name old allocs/op new allocs/op delta JSONLoggerSimple-8 19.0 ± 0% 17.0 ± 0% -10.53% (p=0.000 n=10+10) JSONLoggerContextual-8 25.0 ± 0% 21.0 ± 0% -16.00% (p=0.000 n=10+10) Discard-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) OneWith-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) TwoWith-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) TenWith-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) LogfmtLoggerSimple-8 4.00 ± 0% 2.00 ± 0% -50.00% (p=0.000 n=10+10) LogfmtLoggerContextual-8 7.00 ± 0% 3.00 ± 0% -57.14% (p=0.000 n=10+10) NopLoggerSimple-8 4.00 ± 0% 2.00 ± 0% -50.00% (p=0.000 n=10+10) NopLoggerContextual-8 7.00 ± 0% 3.00 ± 0% -57.14% (p=0.000 n=10+10) ValueBindingTimestamp-8 5.00 ± 0% 3.00 ± 0% -40.00% (p=0.000 n=10+10) ValueBindingCaller-8 4.00 ± 0% 2.00 ± 0% -50.00% (p=0.000 n=10+10) fmt name old time/op new time/op delta SprintfPadding-8 88.9ns ± 3% 79.1ns ± 1% -11.09% (p=0.000 n=10+7) SprintfEmpty-8 12.6ns ± 3% 12.8ns ± 3% ~ (p=0.136 n=10+10) SprintfString-8 38.7ns ± 5% 26.9ns ± 6% -30.65% (p=0.000 n=10+10) SprintfTruncateString-8 56.7ns ± 2% 47.0ns ± 3% -17.05% (p=0.000 n=10+10) SprintfQuoteString-8 164ns ± 2% 153ns ± 2% -7.01% (p=0.000 n=10+10) SprintfInt-8 38.9ns ±15% 26.5ns ± 2% -31.93% (p=0.000 n=10+9) SprintfIntInt-8 60.3ns ± 9% 38.2ns ± 1% -36.67% (p=0.000 n=10+8) SprintfPrefixedInt-8 58.6ns ±13% 51.2ns ±11% -12.66% (p=0.001 n=10+10) SprintfFloat-8 71.4ns ± 3% 64.2ns ± 3% -10.08% (p=0.000 n=8+10) SprintfComplex-8 175ns ± 3% 159ns ± 2% -9.03% (p=0.000 n=10+10) SprintfBoolean-8 33.5ns ± 4% 25.7ns ± 5% -23.28% (p=0.000 n=10+10) SprintfHexString-8 65.3ns ± 3% 51.7ns ± 5% -20.86% (p=0.000 n=10+9) SprintfHexBytes-8 67.2ns ± 5% 67.9ns ± 4% ~ (p=0.383 n=10+10) SprintfBytes-8 129ns ± 7% 124ns ± 7% ~ (p=0.074 n=9+10) SprintfStringer-8 127ns ± 4% 126ns ± 8% ~ (p=0.506 n=9+10) SprintfStructure-8 357ns ± 3% 359ns ± 3% ~ (p=0.469 n=10+10) ManyArgs-8 203ns ± 6% 126ns ± 3% -37.94% (p=0.000 n=10+10) FprintInt-8 119ns ±10% 74ns ± 3% -37.54% (p=0.000 n=10+10) FprintfBytes-8 122ns ± 4% 120ns ± 3% ~ (p=0.124 n=10+10) FprintIntNoAlloc-8 78.2ns ± 5% 74.1ns ± 3% -5.28% (p=0.000 n=10+10) ScanInts-8 349µs ± 1% 349µs ± 0% ~ (p=0.606 n=9+8) ScanRecursiveInt-8 43.8ms ± 7% 40.1ms ± 2% -8.42% (p=0.000 n=10+10) ScanRecursiveIntReaderWrapper-8 43.5ms ± 4% 40.4ms ± 2% -7.16% (p=0.000 n=10+9) name old alloc/op new alloc/op delta SprintfPadding-8 24.0B ± 0% 16.0B ± 0% -33.33% (p=0.000 n=10+10) SprintfEmpty-8 0.00B 0.00B ~ (all equal) SprintfString-8 21.0B ± 0% 5.0B ± 0% -76.19% (p=0.000 n=10+10) SprintfTruncateString-8 32.0B ± 0% 16.0B ± 0% -50.00% (p=0.000 n=10+10) SprintfQuoteString-8 48.0B ± 0% 32.0B ± 0% -33.33% (p=0.000 n=10+10) SprintfInt-8 16.0B ± 0% 1.0B ± 0% -93.75% (p=0.000 n=10+10) SprintfIntInt-8 24.0B ± 0% 3.0B ± 0% -87.50% (p=0.000 n=10+10) SprintfPrefixedInt-8 72.0B ± 0% 64.0B ± 0% -11.11% (p=0.000 n=10+10) SprintfFloat-8 16.0B ± 0% 8.0B ± 0% -50.00% (p=0.000 n=10+10) SprintfComplex-8 48.0B ± 0% 32.0B ± 0% -33.33% (p=0.000 n=10+10) SprintfBoolean-8 8.00B ± 0% 4.00B ± 0% -50.00% (p=0.000 n=10+10) SprintfHexString-8 96.0B ± 0% 80.0B ± 0% -16.67% (p=0.000 n=10+10) SprintfHexBytes-8 112B ± 0% 112B ± 0% ~ (all equal) SprintfBytes-8 96.0B ± 0% 96.0B ± 0% ~ (all equal) SprintfStringer-8 32.0B ± 0% 32.0B ± 0% ~ (all equal) SprintfStructure-8 256B ± 0% 256B ± 0% ~ (all equal) ManyArgs-8 80.0B ± 0% 0.0B -100.00% (p=0.000 n=10+10) FprintInt-8 8.00B ± 0% 0.00B -100.00% (p=0.000 n=10+10) FprintfBytes-8 32.0B ± 0% 32.0B ± 0% ~ (all equal) FprintIntNoAlloc-8 0.00B 0.00B ~ (all equal) ScanInts-8 15.2kB ± 0% 15.2kB ± 0% ~ (p=0.248 n=9+10) ScanRecursiveInt-8 21.6kB ± 0% 21.6kB ± 0% ~ (all equal) ScanRecursiveIntReaderWrapper-8 21.7kB ± 0% 21.7kB ± 0% ~ (all equal) name old allocs/op new allocs/op delta SprintfPadding-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfEmpty-8 0.00 0.00 ~ (all equal) SprintfString-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfTruncateString-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfQuoteString-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfInt-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfIntInt-8 3.00 ± 0% 1.00 ± 0% -66.67% (p=0.000 n=10+10) SprintfPrefixedInt-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfFloat-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfComplex-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfBoolean-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfHexString-8 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10) SprintfHexBytes-8 2.00 ± 0% 2.00 ± 0% ~ (all equal) SprintfBytes-8 2.00 ± 0% 2.00 ± 0% ~ (all equal) SprintfStringer-8 4.00 ± 0% 4.00 ± 0% ~ (all equal) SprintfStructure-8 7.00 ± 0% 7.00 ± 0% ~ (all equal) ManyArgs-8 8.00 ± 0% 0.00 -100.00% (p=0.000 n=10+10) FprintInt-8 1.00 ± 0% 0.00 -100.00% (p=0.000 n=10+10) FprintfBytes-8 1.00 ± 0% 1.00 ± 0% ~ (all equal) FprintIntNoAlloc-8 0.00 0.00 ~ (all equal) ScanInts-8 1.60k ± 0% 1.60k ± 0% ~ (all equal) ScanRecursiveInt-8 1.71k ± 0% 1.71k ± 0% ~ (all equal) ScanRecursiveIntReaderWrapper-8 1.71k ± 0% 1.71k ± 0% ~ (all equal) Change-Id: I7ba72a25fea4140a0ba40a9f443103ed87cc69b5 Reviewed-on: https://go-review.googlesource.com/35554 Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Keith Randall <khr@golang.org>
2017-01-21 14:41:06 -07:00
// byteptr is used to ensure that a temp is required for runtime calls below.
func byteptr() *byte
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
func f19() {
// dest temporary for channel receive.
var z *byte
if b {
z = <-ch // ERROR "stack object .autotmp_[0-9]+ \*byte$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
z = <-ch
z = <-ch // ERROR "live at call to chanrecv1: .autotmp_[0-9]+$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printbytepointer(z)
}
func f20() {
// src temporary for channel send
if b {
ch <- byteptr() // ERROR "stack object .autotmp_[0-9]+ \*byte$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
ch <- byteptr()
ch <- byteptr()
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
func f21() {
// key temporary for mapaccess using array literal key.
var z *byte
if b {
z = m2[[2]string{"x", "y"}] // ERROR "stack object .autotmp_[0-9]+ \[2\]string$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
z = m2[[2]string{"x", "y"}]
z = m2[[2]string{"x", "y"}]
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printbytepointer(z)
}
func f23() {
// key temporary for two-result map access using array literal key.
var z *byte
var ok bool
if b {
z, ok = m2[[2]string{"x", "y"}] // ERROR "stack object .autotmp_[0-9]+ \[2\]string$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
z, ok = m2[[2]string{"x", "y"}]
z, ok = m2[[2]string{"x", "y"}]
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printbytepointer(z)
print(ok)
}
func f24() {
// key temporary for map access using array literal key.
// value temporary too.
if b {
m2[[2]string{"x", "y"}] = nil // ERROR "stack object .autotmp_[0-9]+ \[2\]string$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
m2[[2]string{"x", "y"}] = nil
m2[[2]string{"x", "y"}] = nil
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
cmd/compile, cmd/link, runtime: make defers low-cost through inline code and extra funcdata Generate inline code at defer time to save the args of defer calls to unique (autotmp) stack slots, and generate inline code at exit time to check which defer calls were made and make the associated function/method/interface calls. We remember that a particular defer statement was reached by storing in the deferBits variable (always stored on the stack). At exit time, we check the bits of the deferBits variable to determine which defer function calls to make (in reverse order). These low-cost defers are only used for functions where no defers appear in loops. In addition, we don't do these low-cost defers if there are too many defer statements or too many exits in a function (to limit code increase). When a function uses open-coded defers, we produce extra FUNCDATA_OpenCodedDeferInfo information that specifies the number of defers, and for each defer, the stack slots where the closure and associated args have been stored. The funcdata also includes the location of the deferBits variable. Therefore, for panics, we can use this funcdata to determine exactly which defers are active, and call the appropriate functions/methods/closures with the correct arguments for each active defer. In order to unwind the stack correctly after a recover(), we need to add an extra code segment to functions with open-coded defers that simply calls deferreturn() and returns. This segment is not reachable by the normal function, but is returned to by the runtime during recovery. We set the liveness information of this deferreturn() to be the same as the liveness at the first function call during the last defer exit code (so all return values and all stack slots needed by the defer calls will be live). I needed to increase the stackguard constant from 880 to 896, because of a small amount of new code in deferreturn(). The -N flag disables open-coded defers. '-d defer' prints out the kind of defer being used at each defer statement (heap-allocated, stack-allocated, or open-coded). Cost of defer statement [ go test -run NONE -bench BenchmarkDefer$ runtime ] With normal (stack-allocated) defers only: 35.4 ns/op With open-coded defers: 5.6 ns/op Cost of function call alone (remove defer keyword): 4.4 ns/op Text size increase (including funcdata) for go binary without/with open-coded defers: 0.09% The average size increase (including funcdata) for only the functions that use open-coded defers is 1.1%. The cost of a panic followed by a recover got noticeably slower, since panic processing now requires a scan of the stack for open-coded defer frames. This scan is required, even if no frames are using open-coded defers: Cost of panic and recover [ go test -run NONE -bench BenchmarkPanicRecover runtime ] Without open-coded defers: 62.0 ns/op With open-coded defers: 255 ns/op A CGO Go-to-C-to-Go benchmark got noticeably faster because of open-coded defers: CGO Go-to-C-to-Go benchmark [cd misc/cgo/test; go test -run NONE -bench BenchmarkCGoCallback ] Without open-coded defers: 443 ns/op With open-coded defers: 347 ns/op Updates #14939 (defer performance) Updates #34481 (design doc) Change-Id: I63b1a60d1ebf28126f55ee9fd7ecffe9cb23d1ff Reviewed-on: https://go-review.googlesource.com/c/go/+/202340 Reviewed-by: Austin Clements <austin@google.com>
2019-06-24 13:59:22 -06:00
// Non-open-coded defers should not cause autotmps. (Open-coded defers do create extra autotmps).
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
func f25(b bool) {
cmd/compile, cmd/link, runtime: make defers low-cost through inline code and extra funcdata Generate inline code at defer time to save the args of defer calls to unique (autotmp) stack slots, and generate inline code at exit time to check which defer calls were made and make the associated function/method/interface calls. We remember that a particular defer statement was reached by storing in the deferBits variable (always stored on the stack). At exit time, we check the bits of the deferBits variable to determine which defer function calls to make (in reverse order). These low-cost defers are only used for functions where no defers appear in loops. In addition, we don't do these low-cost defers if there are too many defer statements or too many exits in a function (to limit code increase). When a function uses open-coded defers, we produce extra FUNCDATA_OpenCodedDeferInfo information that specifies the number of defers, and for each defer, the stack slots where the closure and associated args have been stored. The funcdata also includes the location of the deferBits variable. Therefore, for panics, we can use this funcdata to determine exactly which defers are active, and call the appropriate functions/methods/closures with the correct arguments for each active defer. In order to unwind the stack correctly after a recover(), we need to add an extra code segment to functions with open-coded defers that simply calls deferreturn() and returns. This segment is not reachable by the normal function, but is returned to by the runtime during recovery. We set the liveness information of this deferreturn() to be the same as the liveness at the first function call during the last defer exit code (so all return values and all stack slots needed by the defer calls will be live). I needed to increase the stackguard constant from 880 to 896, because of a small amount of new code in deferreturn(). The -N flag disables open-coded defers. '-d defer' prints out the kind of defer being used at each defer statement (heap-allocated, stack-allocated, or open-coded). Cost of defer statement [ go test -run NONE -bench BenchmarkDefer$ runtime ] With normal (stack-allocated) defers only: 35.4 ns/op With open-coded defers: 5.6 ns/op Cost of function call alone (remove defer keyword): 4.4 ns/op Text size increase (including funcdata) for go binary without/with open-coded defers: 0.09% The average size increase (including funcdata) for only the functions that use open-coded defers is 1.1%. The cost of a panic followed by a recover got noticeably slower, since panic processing now requires a scan of the stack for open-coded defer frames. This scan is required, even if no frames are using open-coded defers: Cost of panic and recover [ go test -run NONE -bench BenchmarkPanicRecover runtime ] Without open-coded defers: 62.0 ns/op With open-coded defers: 255 ns/op A CGO Go-to-C-to-Go benchmark got noticeably faster because of open-coded defers: CGO Go-to-C-to-Go benchmark [cd misc/cgo/test; go test -run NONE -bench BenchmarkCGoCallback ] Without open-coded defers: 443 ns/op With open-coded defers: 347 ns/op Updates #14939 (defer performance) Updates #34481 (design doc) Change-Id: I63b1a60d1ebf28126f55ee9fd7ecffe9cb23d1ff Reviewed-on: https://go-review.googlesource.com/c/go/+/202340 Reviewed-by: Austin Clements <austin@google.com>
2019-06-24 13:59:22 -06:00
for i := 0; i < 2; i++ {
// Put in loop to make sure defer is not open-coded
defer g25()
}
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
if b {
return
}
var x string
x = g14()
printstring(x)
cmd/compile, cmd/link, runtime: make defers low-cost through inline code and extra funcdata Generate inline code at defer time to save the args of defer calls to unique (autotmp) stack slots, and generate inline code at exit time to check which defer calls were made and make the associated function/method/interface calls. We remember that a particular defer statement was reached by storing in the deferBits variable (always stored on the stack). At exit time, we check the bits of the deferBits variable to determine which defer function calls to make (in reverse order). These low-cost defers are only used for functions where no defers appear in loops. In addition, we don't do these low-cost defers if there are too many defer statements or too many exits in a function (to limit code increase). When a function uses open-coded defers, we produce extra FUNCDATA_OpenCodedDeferInfo information that specifies the number of defers, and for each defer, the stack slots where the closure and associated args have been stored. The funcdata also includes the location of the deferBits variable. Therefore, for panics, we can use this funcdata to determine exactly which defers are active, and call the appropriate functions/methods/closures with the correct arguments for each active defer. In order to unwind the stack correctly after a recover(), we need to add an extra code segment to functions with open-coded defers that simply calls deferreturn() and returns. This segment is not reachable by the normal function, but is returned to by the runtime during recovery. We set the liveness information of this deferreturn() to be the same as the liveness at the first function call during the last defer exit code (so all return values and all stack slots needed by the defer calls will be live). I needed to increase the stackguard constant from 880 to 896, because of a small amount of new code in deferreturn(). The -N flag disables open-coded defers. '-d defer' prints out the kind of defer being used at each defer statement (heap-allocated, stack-allocated, or open-coded). Cost of defer statement [ go test -run NONE -bench BenchmarkDefer$ runtime ] With normal (stack-allocated) defers only: 35.4 ns/op With open-coded defers: 5.6 ns/op Cost of function call alone (remove defer keyword): 4.4 ns/op Text size increase (including funcdata) for go binary without/with open-coded defers: 0.09% The average size increase (including funcdata) for only the functions that use open-coded defers is 1.1%. The cost of a panic followed by a recover got noticeably slower, since panic processing now requires a scan of the stack for open-coded defer frames. This scan is required, even if no frames are using open-coded defers: Cost of panic and recover [ go test -run NONE -bench BenchmarkPanicRecover runtime ] Without open-coded defers: 62.0 ns/op With open-coded defers: 255 ns/op A CGO Go-to-C-to-Go benchmark got noticeably faster because of open-coded defers: CGO Go-to-C-to-Go benchmark [cd misc/cgo/test; go test -run NONE -bench BenchmarkCGoCallback ] Without open-coded defers: 443 ns/op With open-coded defers: 347 ns/op Updates #14939 (defer performance) Updates #34481 (design doc) Change-Id: I63b1a60d1ebf28126f55ee9fd7ecffe9cb23d1ff Reviewed-on: https://go-review.googlesource.com/c/go/+/202340 Reviewed-by: Austin Clements <austin@google.com>
2019-06-24 13:59:22 -06:00
return
}
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
func g25()
// non-escaping ... slices passed to function call should die on return,
// so that the temporaries do not stack and do not cause ambiguously
// live variables.
func f26(b bool) {
if b {
print26((*int)(nil), (*int)(nil), (*int)(nil)) // ERROR "stack object .autotmp_[0-9]+ \[3\]interface \{\}$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
print26((*int)(nil), (*int)(nil), (*int)(nil))
print26((*int)(nil), (*int)(nil), (*int)(nil))
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printnl()
}
//go:noescape
func print26(...interface{})
// non-escaping closures passed to function call should die on return
func f27(b bool) {
x := 0
if b {
call27(func() { x++ }) // ERROR "stack object .autotmp_[0-9]+ struct \{"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
call27(func() { x++ })
call27(func() { x++ })
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printnl()
}
// but defer does escape to later execution in the function
func f27defer(b bool) {
x := 0
if b {
defer call27(func() { x++ }) // ERROR "stack object .autotmp_[0-9]+ struct \{"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
defer call27(func() { x++ }) // ERROR "stack object .autotmp_[0-9]+ struct \{"
cmd/compile, cmd/link, runtime: make defers low-cost through inline code and extra funcdata Generate inline code at defer time to save the args of defer calls to unique (autotmp) stack slots, and generate inline code at exit time to check which defer calls were made and make the associated function/method/interface calls. We remember that a particular defer statement was reached by storing in the deferBits variable (always stored on the stack). At exit time, we check the bits of the deferBits variable to determine which defer function calls to make (in reverse order). These low-cost defers are only used for functions where no defers appear in loops. In addition, we don't do these low-cost defers if there are too many defer statements or too many exits in a function (to limit code increase). When a function uses open-coded defers, we produce extra FUNCDATA_OpenCodedDeferInfo information that specifies the number of defers, and for each defer, the stack slots where the closure and associated args have been stored. The funcdata also includes the location of the deferBits variable. Therefore, for panics, we can use this funcdata to determine exactly which defers are active, and call the appropriate functions/methods/closures with the correct arguments for each active defer. In order to unwind the stack correctly after a recover(), we need to add an extra code segment to functions with open-coded defers that simply calls deferreturn() and returns. This segment is not reachable by the normal function, but is returned to by the runtime during recovery. We set the liveness information of this deferreturn() to be the same as the liveness at the first function call during the last defer exit code (so all return values and all stack slots needed by the defer calls will be live). I needed to increase the stackguard constant from 880 to 896, because of a small amount of new code in deferreturn(). The -N flag disables open-coded defers. '-d defer' prints out the kind of defer being used at each defer statement (heap-allocated, stack-allocated, or open-coded). Cost of defer statement [ go test -run NONE -bench BenchmarkDefer$ runtime ] With normal (stack-allocated) defers only: 35.4 ns/op With open-coded defers: 5.6 ns/op Cost of function call alone (remove defer keyword): 4.4 ns/op Text size increase (including funcdata) for go binary without/with open-coded defers: 0.09% The average size increase (including funcdata) for only the functions that use open-coded defers is 1.1%. The cost of a panic followed by a recover got noticeably slower, since panic processing now requires a scan of the stack for open-coded defer frames. This scan is required, even if no frames are using open-coded defers: Cost of panic and recover [ go test -run NONE -bench BenchmarkPanicRecover runtime ] Without open-coded defers: 62.0 ns/op With open-coded defers: 255 ns/op A CGO Go-to-C-to-Go benchmark got noticeably faster because of open-coded defers: CGO Go-to-C-to-Go benchmark [cd misc/cgo/test; go test -run NONE -bench BenchmarkCGoCallback ] Without open-coded defers: 443 ns/op With open-coded defers: 347 ns/op Updates #14939 (defer performance) Updates #34481 (design doc) Change-Id: I63b1a60d1ebf28126f55ee9fd7ecffe9cb23d1ff Reviewed-on: https://go-review.googlesource.com/c/go/+/202340 Reviewed-by: Austin Clements <austin@google.com>
2019-06-24 13:59:22 -06:00
printnl() // ERROR "live at call to printnl: .autotmp_[0-9]+ .autotmp_[0-9]+"
return // ERROR "live at indirect call: .autotmp_[0-9]+"
}
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
// and newproc (go) escapes to the heap
func f27go(b bool) {
x := 0
if b {
go call27(func() { x++ }) // ERROR "live at call to newobject: &x$" "live at call to newobject: &x .autotmp_[0-9]+$" "live at call to newproc: &x$" // allocate two closures, the func literal, and the wrapper for go
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
go call27(func() { x++ }) // ERROR "live at call to newobject: &x$" "live at call to newobject: .autotmp_[0-9]+$" // allocate two closures, the func literal, and the wrapper for go
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printnl()
}
//go:noescape
func call27(func())
// concatstring slice should die on return
var s1, s2, s3, s4, s5, s6, s7, s8, s9, s10 string
func f28(b bool) {
if b {
printstring(s1 + s2 + s3 + s4 + s5 + s6 + s7 + s8 + s9 + s10) // ERROR "stack object .autotmp_[0-9]+ \[10\]string$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
printstring(s1 + s2 + s3 + s4 + s5 + s6 + s7 + s8 + s9 + s10)
printstring(s1 + s2 + s3 + s4 + s5 + s6 + s7 + s8 + s9 + s10)
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
// map iterator should die on end of range loop
func f29(b bool) {
if b {
for k := range m { // ERROR "live at call to mapiterinit: .autotmp_[0-9]+$" "live at call to mapiternext: .autotmp_[0-9]+$" "stack object .autotmp_[0-9]+ map.iter\[string\]int$"
printstring(k) // ERROR "live at call to printstring: .autotmp_[0-9]+$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
}
for k := range m { // ERROR "live at call to mapiterinit: .autotmp_[0-9]+$" "live at call to mapiternext: .autotmp_[0-9]+$"
printstring(k) // ERROR "live at call to printstring: .autotmp_[0-9]+$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
for k := range m { // ERROR "live at call to mapiterinit: .autotmp_[0-9]+$" "live at call to mapiternext: .autotmp_[0-9]+$"
printstring(k) // ERROR "live at call to printstring: .autotmp_[0-9]+$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
}
// copy of array of pointers should die at end of range loop
cmd/compile: simplify slice/array range loops for some element sizes In range loops over slices and arrays besides a variable to track the index an extra variable containing the address of the current element is used. To compute a pointer to the next element the elements size is added to the address. On 386 and amd64 an element of size 1, 2, 4 or 8 bytes can by copied from an array using a MOV instruction with suitable addressing mode that uses the start address of the array, the index of the element and element size as scaling factor. Thereby, for arrays and slices with suitable element size we can avoid keeping and incrementing an extra variable to compute the next elements address. Shrinks cmd/go by 4 kilobytes. AMD64: name old time/op new time/op delta BinaryTree17 2.66s ± 7% 2.54s ± 0% -4.53% (p=0.000 n=10+8) Fannkuch11 3.02s ± 1% 3.02s ± 1% ~ (p=0.579 n=10+10) FmtFprintfEmpty 45.6ns ± 1% 42.2ns ± 1% -7.46% (p=0.000 n=10+10) FmtFprintfString 69.8ns ± 1% 70.4ns ± 1% +0.84% (p=0.041 n=10+10) FmtFprintfInt 80.1ns ± 1% 79.0ns ± 1% -1.35% (p=0.000 n=10+10) FmtFprintfIntInt 127ns ± 1% 125ns ± 1% -1.00% (p=0.007 n=10+9) FmtFprintfPrefixedInt 158ns ± 2% 152ns ± 1% -4.11% (p=0.000 n=10+10) FmtFprintfFloat 218ns ± 1% 214ns ± 1% -1.61% (p=0.000 n=10+10) FmtManyArgs 508ns ± 1% 504ns ± 1% -0.93% (p=0.001 n=9+10) GobDecode 6.76ms ± 1% 6.78ms ± 1% ~ (p=0.353 n=10+10) GobEncode 5.84ms ± 1% 5.77ms ± 1% -1.31% (p=0.000 n=10+9) Gzip 223ms ± 1% 218ms ± 1% -2.39% (p=0.000 n=10+10) Gunzip 40.3ms ± 1% 40.4ms ± 3% ~ (p=0.796 n=10+10) HTTPClientServer 73.5µs ± 0% 73.3µs ± 0% -0.28% (p=0.000 n=10+9) JSONEncode 12.7ms ± 1% 12.6ms ± 8% ~ (p=0.173 n=8+10) JSONDecode 57.5ms ± 1% 56.1ms ± 2% -2.40% (p=0.000 n=10+10) Mandelbrot200 3.80ms ± 1% 3.86ms ± 6% ~ (p=0.579 n=10+10) GoParse 3.25ms ± 1% 3.23ms ± 1% ~ (p=0.052 n=10+10) RegexpMatchEasy0_32 74.4ns ± 1% 76.9ns ± 1% +3.39% (p=0.000 n=10+10) RegexpMatchEasy0_1K 243ns ± 2% 248ns ± 1% +1.86% (p=0.000 n=10+8) RegexpMatchEasy1_32 71.0ns ± 2% 72.8ns ± 1% +2.55% (p=0.000 n=10+10) RegexpMatchEasy1_1K 370ns ± 1% 383ns ± 0% +3.39% (p=0.000 n=10+9) RegexpMatchMedium_32 107ns ± 0% 113ns ± 1% +5.33% (p=0.000 n=6+10) RegexpMatchMedium_1K 35.0µs ± 1% 36.0µs ± 1% +3.13% (p=0.000 n=10+10) RegexpMatchHard_32 1.65µs ± 1% 1.69µs ± 1% +2.23% (p=0.000 n=10+9) RegexpMatchHard_1K 49.8µs ± 1% 50.6µs ± 1% +1.59% (p=0.000 n=10+10) Revcomp 398ms ± 1% 396ms ± 1% -0.51% (p=0.043 n=10+10) Template 63.4ms ± 1% 60.8ms ± 0% -4.11% (p=0.000 n=10+9) TimeParse 318ns ± 1% 322ns ± 1% +1.10% (p=0.005 n=10+10) TimeFormat 323ns ± 1% 336ns ± 1% +4.15% (p=0.000 n=10+10) Updates: #15809. Change-Id: I55915aaf6d26768e12247f8a8edf14e7630726d1 Reviewed-on: https://go-review.googlesource.com/38061 Run-TryBot: Martin Möhrmann <moehrmann@google.com> Reviewed-by: Keith Randall <khr@golang.org>
2016-12-18 12:13:58 -07:00
var pstructarr [10]pstruct
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
// Struct size chosen to make pointer to element in pstructarr
cmd/compile: simplify slice/array range loops for some element sizes In range loops over slices and arrays besides a variable to track the index an extra variable containing the address of the current element is used. To compute a pointer to the next element the elements size is added to the address. On 386 and amd64 an element of size 1, 2, 4 or 8 bytes can by copied from an array using a MOV instruction with suitable addressing mode that uses the start address of the array, the index of the element and element size as scaling factor. Thereby, for arrays and slices with suitable element size we can avoid keeping and incrementing an extra variable to compute the next elements address. Shrinks cmd/go by 4 kilobytes. AMD64: name old time/op new time/op delta BinaryTree17 2.66s ± 7% 2.54s ± 0% -4.53% (p=0.000 n=10+8) Fannkuch11 3.02s ± 1% 3.02s ± 1% ~ (p=0.579 n=10+10) FmtFprintfEmpty 45.6ns ± 1% 42.2ns ± 1% -7.46% (p=0.000 n=10+10) FmtFprintfString 69.8ns ± 1% 70.4ns ± 1% +0.84% (p=0.041 n=10+10) FmtFprintfInt 80.1ns ± 1% 79.0ns ± 1% -1.35% (p=0.000 n=10+10) FmtFprintfIntInt 127ns ± 1% 125ns ± 1% -1.00% (p=0.007 n=10+9) FmtFprintfPrefixedInt 158ns ± 2% 152ns ± 1% -4.11% (p=0.000 n=10+10) FmtFprintfFloat 218ns ± 1% 214ns ± 1% -1.61% (p=0.000 n=10+10) FmtManyArgs 508ns ± 1% 504ns ± 1% -0.93% (p=0.001 n=9+10) GobDecode 6.76ms ± 1% 6.78ms ± 1% ~ (p=0.353 n=10+10) GobEncode 5.84ms ± 1% 5.77ms ± 1% -1.31% (p=0.000 n=10+9) Gzip 223ms ± 1% 218ms ± 1% -2.39% (p=0.000 n=10+10) Gunzip 40.3ms ± 1% 40.4ms ± 3% ~ (p=0.796 n=10+10) HTTPClientServer 73.5µs ± 0% 73.3µs ± 0% -0.28% (p=0.000 n=10+9) JSONEncode 12.7ms ± 1% 12.6ms ± 8% ~ (p=0.173 n=8+10) JSONDecode 57.5ms ± 1% 56.1ms ± 2% -2.40% (p=0.000 n=10+10) Mandelbrot200 3.80ms ± 1% 3.86ms ± 6% ~ (p=0.579 n=10+10) GoParse 3.25ms ± 1% 3.23ms ± 1% ~ (p=0.052 n=10+10) RegexpMatchEasy0_32 74.4ns ± 1% 76.9ns ± 1% +3.39% (p=0.000 n=10+10) RegexpMatchEasy0_1K 243ns ± 2% 248ns ± 1% +1.86% (p=0.000 n=10+8) RegexpMatchEasy1_32 71.0ns ± 2% 72.8ns ± 1% +2.55% (p=0.000 n=10+10) RegexpMatchEasy1_1K 370ns ± 1% 383ns ± 0% +3.39% (p=0.000 n=10+9) RegexpMatchMedium_32 107ns ± 0% 113ns ± 1% +5.33% (p=0.000 n=6+10) RegexpMatchMedium_1K 35.0µs ± 1% 36.0µs ± 1% +3.13% (p=0.000 n=10+10) RegexpMatchHard_32 1.65µs ± 1% 1.69µs ± 1% +2.23% (p=0.000 n=10+9) RegexpMatchHard_1K 49.8µs ± 1% 50.6µs ± 1% +1.59% (p=0.000 n=10+10) Revcomp 398ms ± 1% 396ms ± 1% -0.51% (p=0.043 n=10+10) Template 63.4ms ± 1% 60.8ms ± 0% -4.11% (p=0.000 n=10+9) TimeParse 318ns ± 1% 322ns ± 1% +1.10% (p=0.005 n=10+10) TimeFormat 323ns ± 1% 336ns ± 1% +4.15% (p=0.000 n=10+10) Updates: #15809. Change-Id: I55915aaf6d26768e12247f8a8edf14e7630726d1 Reviewed-on: https://go-review.googlesource.com/38061 Run-TryBot: Martin Möhrmann <moehrmann@google.com> Reviewed-by: Keith Randall <khr@golang.org>
2016-12-18 12:13:58 -07:00
// not computable by strength reduction.
type pstruct struct {
intp *int
_ [8]byte
}
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
func f30(b bool) {
// live temp during printintpointer(p):
cmd/compile: simplify slice/array range loops for some element sizes In range loops over slices and arrays besides a variable to track the index an extra variable containing the address of the current element is used. To compute a pointer to the next element the elements size is added to the address. On 386 and amd64 an element of size 1, 2, 4 or 8 bytes can by copied from an array using a MOV instruction with suitable addressing mode that uses the start address of the array, the index of the element and element size as scaling factor. Thereby, for arrays and slices with suitable element size we can avoid keeping and incrementing an extra variable to compute the next elements address. Shrinks cmd/go by 4 kilobytes. AMD64: name old time/op new time/op delta BinaryTree17 2.66s ± 7% 2.54s ± 0% -4.53% (p=0.000 n=10+8) Fannkuch11 3.02s ± 1% 3.02s ± 1% ~ (p=0.579 n=10+10) FmtFprintfEmpty 45.6ns ± 1% 42.2ns ± 1% -7.46% (p=0.000 n=10+10) FmtFprintfString 69.8ns ± 1% 70.4ns ± 1% +0.84% (p=0.041 n=10+10) FmtFprintfInt 80.1ns ± 1% 79.0ns ± 1% -1.35% (p=0.000 n=10+10) FmtFprintfIntInt 127ns ± 1% 125ns ± 1% -1.00% (p=0.007 n=10+9) FmtFprintfPrefixedInt 158ns ± 2% 152ns ± 1% -4.11% (p=0.000 n=10+10) FmtFprintfFloat 218ns ± 1% 214ns ± 1% -1.61% (p=0.000 n=10+10) FmtManyArgs 508ns ± 1% 504ns ± 1% -0.93% (p=0.001 n=9+10) GobDecode 6.76ms ± 1% 6.78ms ± 1% ~ (p=0.353 n=10+10) GobEncode 5.84ms ± 1% 5.77ms ± 1% -1.31% (p=0.000 n=10+9) Gzip 223ms ± 1% 218ms ± 1% -2.39% (p=0.000 n=10+10) Gunzip 40.3ms ± 1% 40.4ms ± 3% ~ (p=0.796 n=10+10) HTTPClientServer 73.5µs ± 0% 73.3µs ± 0% -0.28% (p=0.000 n=10+9) JSONEncode 12.7ms ± 1% 12.6ms ± 8% ~ (p=0.173 n=8+10) JSONDecode 57.5ms ± 1% 56.1ms ± 2% -2.40% (p=0.000 n=10+10) Mandelbrot200 3.80ms ± 1% 3.86ms ± 6% ~ (p=0.579 n=10+10) GoParse 3.25ms ± 1% 3.23ms ± 1% ~ (p=0.052 n=10+10) RegexpMatchEasy0_32 74.4ns ± 1% 76.9ns ± 1% +3.39% (p=0.000 n=10+10) RegexpMatchEasy0_1K 243ns ± 2% 248ns ± 1% +1.86% (p=0.000 n=10+8) RegexpMatchEasy1_32 71.0ns ± 2% 72.8ns ± 1% +2.55% (p=0.000 n=10+10) RegexpMatchEasy1_1K 370ns ± 1% 383ns ± 0% +3.39% (p=0.000 n=10+9) RegexpMatchMedium_32 107ns ± 0% 113ns ± 1% +5.33% (p=0.000 n=6+10) RegexpMatchMedium_1K 35.0µs ± 1% 36.0µs ± 1% +3.13% (p=0.000 n=10+10) RegexpMatchHard_32 1.65µs ± 1% 1.69µs ± 1% +2.23% (p=0.000 n=10+9) RegexpMatchHard_1K 49.8µs ± 1% 50.6µs ± 1% +1.59% (p=0.000 n=10+10) Revcomp 398ms ± 1% 396ms ± 1% -0.51% (p=0.043 n=10+10) Template 63.4ms ± 1% 60.8ms ± 0% -4.11% (p=0.000 n=10+9) TimeParse 318ns ± 1% 322ns ± 1% +1.10% (p=0.005 n=10+10) TimeFormat 323ns ± 1% 336ns ± 1% +4.15% (p=0.000 n=10+10) Updates: #15809. Change-Id: I55915aaf6d26768e12247f8a8edf14e7630726d1 Reviewed-on: https://go-review.googlesource.com/38061 Run-TryBot: Martin Möhrmann <moehrmann@google.com> Reviewed-by: Keith Randall <khr@golang.org>
2016-12-18 12:13:58 -07:00
// the internal iterator pointer if a pointer to pstruct in pstructarr
// can not be easily computed by strength reduction.
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
if b {
for _, p := range pstructarr { // ERROR "stack object .autotmp_[0-9]+ \[10\]pstruct$"
printintpointer(p.intp) // ERROR "live at call to printintpointer: .autotmp_[0-9]+$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
}
for _, p := range pstructarr {
printintpointer(p.intp) // ERROR "live at call to printintpointer: .autotmp_[0-9]+$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
for _, p := range pstructarr {
printintpointer(p.intp) // ERROR "live at call to printintpointer: .autotmp_[0-9]+$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
}
// conversion to interface should not leave temporary behind
func f31(b1, b2, b3 bool) {
if b1 {
g31(g18()) // ERROR "stack object .autotmp_[0-9]+ \[2\]string$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
if b2 {
h31(g18()) // ERROR "live at call to convT: .autotmp_[0-9]+$" "live at call to newobject: .autotmp_[0-9]+$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
if b3 {
panic(g18())
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
print(b3)
}
func g31(interface{})
func h31(...interface{})
// non-escaping partial functions passed to function call should die on return
type T32 int
func (t *T32) Inc() { // ERROR "live at entry to \(\*T32\).Inc: t$"
*t++
}
var t32 T32
func f32(b bool) {
if b {
call32(t32.Inc) // ERROR "stack object .autotmp_[0-9]+ struct \{"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
call32(t32.Inc)
call32(t32.Inc)
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
//go:noescape
func call32(func())
// temporaries introduced during if conditions and && || expressions
// should die once the condition has been acted upon.
var m33 map[interface{}]int
func f33() {
if m33[byteptr()] == 0 { // ERROR "stack object .autotmp_[0-9]+ interface \{\}$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printnl()
return
} else {
printnl()
}
printnl()
}
func f34() {
if m33[byteptr()] == 0 { // ERROR "stack object .autotmp_[0-9]+ interface \{\}$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printnl()
return
}
printnl()
}
func f35() {
if m33[byteptr()] == 0 && // ERROR "stack object .autotmp_[0-9]+ interface \{\}"
m33[byteptr()] == 0 { // ERROR "stack object .autotmp_[0-9]+ interface \{\}"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printnl()
return
}
printnl()
}
func f36() {
if m33[byteptr()] == 0 || // ERROR "stack object .autotmp_[0-9]+ interface \{\}"
m33[byteptr()] == 0 { // ERROR "stack object .autotmp_[0-9]+ interface \{\}"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printnl()
return
}
printnl()
}
func f37() {
if (m33[byteptr()] == 0 || // ERROR "stack object .autotmp_[0-9]+ interface \{\}"
m33[byteptr()] == 0) && // ERROR "stack object .autotmp_[0-9]+ interface \{\}"
m33[byteptr()] == 0 {
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printnl()
return
}
printnl()
}
// select temps should disappear in the case bodies
var c38 chan string
func fc38() chan string
func fi38(int) *string
func fb38() *bool
func f38(b bool) {
// we don't care what temps are printed on the lines with output.
// we care that the println lines have no live variables
// and therefore no output.
if b {
select { // ERROR "live at call to selectgo:( .autotmp_[0-9]+)+$" "stack object .autotmp_[0-9]+ \[4\]struct \{"
case <-fc38():
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printnl()
case fc38() <- *fi38(1): // ERROR "live at call to fc38:( .autotmp_[0-9]+)+$" "live at call to fi38:( .autotmp_[0-9]+)+$" "stack object .autotmp_[0-9]+ string$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printnl()
case *fi38(2) = <-fc38(): // ERROR "live at call to fc38:( .autotmp_[0-9]+)+$" "live at call to fi38:( .autotmp_[0-9]+)+$" "stack object .autotmp_[0-9]+ string$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printnl()
case *fi38(3), *fb38() = <-fc38(): // ERROR "stack object .autotmp_[0-9]+ string$" "live at call to fc38:( .autotmp_[0-9]+)+$" "live at call to fi38:( .autotmp_[0-9]+)+$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
printnl()
}
printnl()
}
printnl()
}
// issue 8097: mishandling of x = x during return.
func f39() (x []int) {
x = []int{1}
printnl() // ERROR "live at call to printnl: .autotmp_[0-9]+$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
return x
}
func f39a() (x []int) {
x = []int{1}
printnl() // ERROR "live at call to printnl: .autotmp_[0-9]+$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
return
}
func f39b() (x [10]*int) {
x = [10]*int{}
x[0] = new(int) // ERROR "live at call to newobject: x$"
printnl() // ERROR "live at call to printnl: x$"
return x
}
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
func f39c() (x [10]*int) {
x = [10]*int{}
x[0] = new(int) // ERROR "live at call to newobject: x$"
printnl() // ERROR "live at call to printnl: x$"
return
}
// issue 8142: lost 'addrtaken' bit on inlined variables.
// no inlining in this test, so just checking that non-inlined works.
type T40 struct {
m map[int]int
}
//go:noescape
func useT40(*T40)
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
func newT40() *T40 {
ret := T40{}
ret.m = make(map[int]int, 42) // ERROR "live at call to makemap: &ret$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
return &ret
}
func bad40() {
t := newT40()
_ = t
printnl()
}
func good40() {
ret := T40{} // ERROR "stack object ret T40$"
ret.m = make(map[int]int) // ERROR "live at call to fastrand: .autotmp_[0-9]+$" "stack object .autotmp_[0-9]+ map.hdr\[int\]int$"
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
t := &ret
printnl() // ERROR "live at call to printnl: ret$"
// Note: ret is live at the printnl because the compiler moves &ret
// from before the printnl to after.
useT40(t)
cmd/compile: better job of naming compound types Compound AUTO types weren't named previously. That was because live variable analysis (plive.go) doesn't handle spilling to compound types. It can't handle them because there is no valid place to put VARDEFs when regalloc is spilling compound types. compound types = multiword builtin types: complex, string, slice, and interface. Instead, we split named AUTOs into individual one-word variables. For example, a string s gets split into a byte ptr s.ptr and an integer s.len. Those two variables can be spilled to / restored from independently. As a result, live variable analysis can handle them because they are one-word objects. This CL will change how AUTOs are described in DWARF information. Consider the code: func f(s string, i int) int { x := s[i:i+5] g() return lookup(x) } The old compiler would spill x to two consecutive slots on the stack, both named x (at offsets 0 and 8). The new compiler spills the pointer of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a constant (5) and can be rematerialized for the call to lookup. So compound objects may not be spilled in their entirety, and even if they are they won't necessarily be contiguous. Such is the price of optimization. Re-enable live variable analysis tests. One test remains disabled, it fails because of #14904. Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801 Reviewed-on: https://go-review.googlesource.com/21233 Run-TryBot: Keith Randall <khr@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: David Chase <drchase@google.com>
2016-03-28 12:25:17 -06:00
}
func ddd1(x, y *int) { // ERROR "live at entry to ddd1: x y$"
ddd2(x, y) // ERROR "stack object .autotmp_[0-9]+ \[2\]\*int$"
printnl()
// Note: no .?autotmp live at printnl. See issue 16996.
}
func ddd2(a ...*int) { // ERROR "live at entry to ddd2: a$"
sink = a[0]
}
// issue 16016: autogenerated wrapper should have arguments live
type T struct{}
func (*T) Foo(ptr *int) {}
type R struct{ *T }
// issue 18860: output arguments must be live all the time if there is a defer.
// In particular, at printint r must be live.
func f41(p, q *int) (r *int) { // ERROR "live at entry to f41: p q$"
r = p
cmd/compile, cmd/link, runtime: make defers low-cost through inline code and extra funcdata Generate inline code at defer time to save the args of defer calls to unique (autotmp) stack slots, and generate inline code at exit time to check which defer calls were made and make the associated function/method/interface calls. We remember that a particular defer statement was reached by storing in the deferBits variable (always stored on the stack). At exit time, we check the bits of the deferBits variable to determine which defer function calls to make (in reverse order). These low-cost defers are only used for functions where no defers appear in loops. In addition, we don't do these low-cost defers if there are too many defer statements or too many exits in a function (to limit code increase). When a function uses open-coded defers, we produce extra FUNCDATA_OpenCodedDeferInfo information that specifies the number of defers, and for each defer, the stack slots where the closure and associated args have been stored. The funcdata also includes the location of the deferBits variable. Therefore, for panics, we can use this funcdata to determine exactly which defers are active, and call the appropriate functions/methods/closures with the correct arguments for each active defer. In order to unwind the stack correctly after a recover(), we need to add an extra code segment to functions with open-coded defers that simply calls deferreturn() and returns. This segment is not reachable by the normal function, but is returned to by the runtime during recovery. We set the liveness information of this deferreturn() to be the same as the liveness at the first function call during the last defer exit code (so all return values and all stack slots needed by the defer calls will be live). I needed to increase the stackguard constant from 880 to 896, because of a small amount of new code in deferreturn(). The -N flag disables open-coded defers. '-d defer' prints out the kind of defer being used at each defer statement (heap-allocated, stack-allocated, or open-coded). Cost of defer statement [ go test -run NONE -bench BenchmarkDefer$ runtime ] With normal (stack-allocated) defers only: 35.4 ns/op With open-coded defers: 5.6 ns/op Cost of function call alone (remove defer keyword): 4.4 ns/op Text size increase (including funcdata) for go binary without/with open-coded defers: 0.09% The average size increase (including funcdata) for only the functions that use open-coded defers is 1.1%. The cost of a panic followed by a recover got noticeably slower, since panic processing now requires a scan of the stack for open-coded defer frames. This scan is required, even if no frames are using open-coded defers: Cost of panic and recover [ go test -run NONE -bench BenchmarkPanicRecover runtime ] Without open-coded defers: 62.0 ns/op With open-coded defers: 255 ns/op A CGO Go-to-C-to-Go benchmark got noticeably faster because of open-coded defers: CGO Go-to-C-to-Go benchmark [cd misc/cgo/test; go test -run NONE -bench BenchmarkCGoCallback ] Without open-coded defers: 443 ns/op With open-coded defers: 347 ns/op Updates #14939 (defer performance) Updates #34481 (design doc) Change-Id: I63b1a60d1ebf28126f55ee9fd7ecffe9cb23d1ff Reviewed-on: https://go-review.googlesource.com/c/go/+/202340 Reviewed-by: Austin Clements <austin@google.com>
2019-06-24 13:59:22 -06:00
defer func() {
recover()
}()
cmd/compile, cmd/link, runtime: make defers low-cost through inline code and extra funcdata Generate inline code at defer time to save the args of defer calls to unique (autotmp) stack slots, and generate inline code at exit time to check which defer calls were made and make the associated function/method/interface calls. We remember that a particular defer statement was reached by storing in the deferBits variable (always stored on the stack). At exit time, we check the bits of the deferBits variable to determine which defer function calls to make (in reverse order). These low-cost defers are only used for functions where no defers appear in loops. In addition, we don't do these low-cost defers if there are too many defer statements or too many exits in a function (to limit code increase). When a function uses open-coded defers, we produce extra FUNCDATA_OpenCodedDeferInfo information that specifies the number of defers, and for each defer, the stack slots where the closure and associated args have been stored. The funcdata also includes the location of the deferBits variable. Therefore, for panics, we can use this funcdata to determine exactly which defers are active, and call the appropriate functions/methods/closures with the correct arguments for each active defer. In order to unwind the stack correctly after a recover(), we need to add an extra code segment to functions with open-coded defers that simply calls deferreturn() and returns. This segment is not reachable by the normal function, but is returned to by the runtime during recovery. We set the liveness information of this deferreturn() to be the same as the liveness at the first function call during the last defer exit code (so all return values and all stack slots needed by the defer calls will be live). I needed to increase the stackguard constant from 880 to 896, because of a small amount of new code in deferreturn(). The -N flag disables open-coded defers. '-d defer' prints out the kind of defer being used at each defer statement (heap-allocated, stack-allocated, or open-coded). Cost of defer statement [ go test -run NONE -bench BenchmarkDefer$ runtime ] With normal (stack-allocated) defers only: 35.4 ns/op With open-coded defers: 5.6 ns/op Cost of function call alone (remove defer keyword): 4.4 ns/op Text size increase (including funcdata) for go binary without/with open-coded defers: 0.09% The average size increase (including funcdata) for only the functions that use open-coded defers is 1.1%. The cost of a panic followed by a recover got noticeably slower, since panic processing now requires a scan of the stack for open-coded defer frames. This scan is required, even if no frames are using open-coded defers: Cost of panic and recover [ go test -run NONE -bench BenchmarkPanicRecover runtime ] Without open-coded defers: 62.0 ns/op With open-coded defers: 255 ns/op A CGO Go-to-C-to-Go benchmark got noticeably faster because of open-coded defers: CGO Go-to-C-to-Go benchmark [cd misc/cgo/test; go test -run NONE -bench BenchmarkCGoCallback ] Without open-coded defers: 443 ns/op With open-coded defers: 347 ns/op Updates #14939 (defer performance) Updates #34481 (design doc) Change-Id: I63b1a60d1ebf28126f55ee9fd7ecffe9cb23d1ff Reviewed-on: https://go-review.googlesource.com/c/go/+/202340 Reviewed-by: Austin Clements <austin@google.com>
2019-06-24 13:59:22 -06:00
printint(0) // ERROR "live at call to printint: q r .autotmp_[0-9]+$"
r = q
cmd/compile, cmd/link, runtime: make defers low-cost through inline code and extra funcdata Generate inline code at defer time to save the args of defer calls to unique (autotmp) stack slots, and generate inline code at exit time to check which defer calls were made and make the associated function/method/interface calls. We remember that a particular defer statement was reached by storing in the deferBits variable (always stored on the stack). At exit time, we check the bits of the deferBits variable to determine which defer function calls to make (in reverse order). These low-cost defers are only used for functions where no defers appear in loops. In addition, we don't do these low-cost defers if there are too many defer statements or too many exits in a function (to limit code increase). When a function uses open-coded defers, we produce extra FUNCDATA_OpenCodedDeferInfo information that specifies the number of defers, and for each defer, the stack slots where the closure and associated args have been stored. The funcdata also includes the location of the deferBits variable. Therefore, for panics, we can use this funcdata to determine exactly which defers are active, and call the appropriate functions/methods/closures with the correct arguments for each active defer. In order to unwind the stack correctly after a recover(), we need to add an extra code segment to functions with open-coded defers that simply calls deferreturn() and returns. This segment is not reachable by the normal function, but is returned to by the runtime during recovery. We set the liveness information of this deferreturn() to be the same as the liveness at the first function call during the last defer exit code (so all return values and all stack slots needed by the defer calls will be live). I needed to increase the stackguard constant from 880 to 896, because of a small amount of new code in deferreturn(). The -N flag disables open-coded defers. '-d defer' prints out the kind of defer being used at each defer statement (heap-allocated, stack-allocated, or open-coded). Cost of defer statement [ go test -run NONE -bench BenchmarkDefer$ runtime ] With normal (stack-allocated) defers only: 35.4 ns/op With open-coded defers: 5.6 ns/op Cost of function call alone (remove defer keyword): 4.4 ns/op Text size increase (including funcdata) for go binary without/with open-coded defers: 0.09% The average size increase (including funcdata) for only the functions that use open-coded defers is 1.1%. The cost of a panic followed by a recover got noticeably slower, since panic processing now requires a scan of the stack for open-coded defer frames. This scan is required, even if no frames are using open-coded defers: Cost of panic and recover [ go test -run NONE -bench BenchmarkPanicRecover runtime ] Without open-coded defers: 62.0 ns/op With open-coded defers: 255 ns/op A CGO Go-to-C-to-Go benchmark got noticeably faster because of open-coded defers: CGO Go-to-C-to-Go benchmark [cd misc/cgo/test; go test -run NONE -bench BenchmarkCGoCallback ] Without open-coded defers: 443 ns/op With open-coded defers: 347 ns/op Updates #14939 (defer performance) Updates #34481 (design doc) Change-Id: I63b1a60d1ebf28126f55ee9fd7ecffe9cb23d1ff Reviewed-on: https://go-review.googlesource.com/c/go/+/202340 Reviewed-by: Austin Clements <austin@google.com>
2019-06-24 13:59:22 -06:00
return // ERROR "live at call to f41.func1: r .autotmp_[0-9]+$"
}
func f42() {
var p, q, r int
f43([]*int{&p, &q, &r}) // ERROR "stack object .autotmp_[0-9]+ \[3\]\*int$"
f43([]*int{&p, &r, &q})
f43([]*int{&q, &p, &r})
}
//go:noescape
func f43(a []*int)
// Assigning to a sub-element that makes up an entire local variable
// should clobber that variable.
func f44(f func() [2]*int) interface{} { // ERROR "live at entry to f44: f"
type T struct {
s [1][2]*int
}
ret := T{}
ret.s[0] = f()
return ret // ERROR "stack object .autotmp_[0-9]+ T"
}