mirror of
https://github.com/golang/go
synced 2024-11-18 21:54:49 -07:00
cmd/compile: avoid a spill in append fast path
Instead of spilling newlen, recalculate it. This removes a spill from the fast path, at the cost of a cheap recalculation on the (rare) growth path. This uses 8 bytes less of stack space. It generates two more bytes of code, but that is due to suboptimal register allocation; see far below. Runtime append microbenchmarks are all over the map, presumably due to incidental code movement. Sample code: func s(b []byte) []byte { b = append(b, 1, 2, 3) return b } Before: "".s t=1 size=160 args=0x30 locals=0x48 0x0000 00000 (append.go:8) TEXT "".s(SB), $72-48 0x0000 00000 (append.go:8) MOVQ (TLS), CX 0x0009 00009 (append.go:8) CMPQ SP, 16(CX) 0x000d 00013 (append.go:8) JLS 149 0x0013 00019 (append.go:8) SUBQ $72, SP 0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB) 0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB) 0x0017 00023 (append.go:9) MOVQ "".b+88(FP), CX 0x001c 00028 (append.go:9) LEAQ 3(CX), DX 0x0020 00032 (append.go:9) MOVQ DX, "".autotmp_0+64(SP) 0x0025 00037 (append.go:9) MOVQ "".b+96(FP), BX 0x002a 00042 (append.go:9) CMPQ DX, BX 0x002d 00045 (append.go:9) JGT $0, 86 0x002f 00047 (append.go:8) MOVQ "".b+80(FP), AX 0x0034 00052 (append.go:9) MOVB $1, (AX)(CX*1) 0x0038 00056 (append.go:9) MOVB $2, 1(AX)(CX*1) 0x003d 00061 (append.go:9) MOVB $3, 2(AX)(CX*1) 0x0042 00066 (append.go:10) MOVQ AX, "".~r1+104(FP) 0x0047 00071 (append.go:10) MOVQ DX, "".~r1+112(FP) 0x004c 00076 (append.go:10) MOVQ BX, "".~r1+120(FP) 0x0051 00081 (append.go:10) ADDQ $72, SP 0x0055 00085 (append.go:10) RET 0x0056 00086 (append.go:9) LEAQ type.[]uint8(SB), AX 0x005d 00093 (append.go:9) MOVQ AX, (SP) 0x0061 00097 (append.go:9) MOVQ "".b+80(FP), BP 0x0066 00102 (append.go:9) MOVQ BP, 8(SP) 0x006b 00107 (append.go:9) MOVQ CX, 16(SP) 0x0070 00112 (append.go:9) MOVQ BX, 24(SP) 0x0075 00117 (append.go:9) MOVQ DX, 32(SP) 0x007a 00122 (append.go:9) PCDATA $0, $0 0x007a 00122 (append.go:9) CALL runtime.growslice(SB) 0x007f 00127 (append.go:9) MOVQ 40(SP), AX 0x0084 00132 (append.go:9) MOVQ 56(SP), BX 0x0089 00137 (append.go:8) MOVQ "".b+88(FP), CX 0x008e 00142 (append.go:9) MOVQ "".autotmp_0+64(SP), DX 0x0093 00147 (append.go:9) JMP 52 0x0095 00149 (append.go:9) NOP 0x0095 00149 (append.go:8) CALL runtime.morestack_noctxt(SB) 0x009a 00154 (append.go:8) JMP 0 After: "".s t=1 size=176 args=0x30 locals=0x40 0x0000 00000 (append.go:8) TEXT "".s(SB), $64-48 0x0000 00000 (append.go:8) MOVQ (TLS), CX 0x0009 00009 (append.go:8) CMPQ SP, 16(CX) 0x000d 00013 (append.go:8) JLS 151 0x0013 00019 (append.go:8) SUBQ $64, SP 0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB) 0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB) 0x0017 00023 (append.go:9) MOVQ "".b+80(FP), CX 0x001c 00028 (append.go:9) LEAQ 3(CX), DX 0x0020 00032 (append.go:9) MOVQ "".b+88(FP), BX 0x0025 00037 (append.go:9) CMPQ DX, BX 0x0028 00040 (append.go:9) JGT $0, 81 0x002a 00042 (append.go:8) MOVQ "".b+72(FP), AX 0x002f 00047 (append.go:9) MOVB $1, (AX)(CX*1) 0x0033 00051 (append.go:9) MOVB $2, 1(AX)(CX*1) 0x0038 00056 (append.go:9) MOVB $3, 2(AX)(CX*1) 0x003d 00061 (append.go:10) MOVQ AX, "".~r1+96(FP) 0x0042 00066 (append.go:10) MOVQ DX, "".~r1+104(FP) 0x0047 00071 (append.go:10) MOVQ BX, "".~r1+112(FP) 0x004c 00076 (append.go:10) ADDQ $64, SP 0x0050 00080 (append.go:10) RET 0x0051 00081 (append.go:9) LEAQ type.[]uint8(SB), AX 0x0058 00088 (append.go:9) MOVQ AX, (SP) 0x005c 00092 (append.go:9) MOVQ "".b+72(FP), BP 0x0061 00097 (append.go:9) MOVQ BP, 8(SP) 0x0066 00102 (append.go:9) MOVQ CX, 16(SP) 0x006b 00107 (append.go:9) MOVQ BX, 24(SP) 0x0070 00112 (append.go:9) MOVQ DX, 32(SP) 0x0075 00117 (append.go:9) PCDATA $0, $0 0x0075 00117 (append.go:9) CALL runtime.growslice(SB) 0x007a 00122 (append.go:9) MOVQ 40(SP), AX 0x007f 00127 (append.go:9) MOVQ 48(SP), CX 0x0084 00132 (append.go:9) MOVQ 56(SP), BX 0x0089 00137 (append.go:9) ADDQ $3, CX 0x008d 00141 (append.go:9) MOVQ CX, DX 0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX 0x0095 00149 (append.go:9) JMP 47 0x0097 00151 (append.go:9) NOP 0x0097 00151 (append.go:8) CALL runtime.morestack_noctxt(SB) 0x009c 00156 (append.go:8) JMP 0 Observe that in the following sequence, we should use DX directly instead of using CX as a temporary register, which would make the new code a strict improvement on the old: 0x007f 00127 (append.go:9) MOVQ 48(SP), CX 0x0084 00132 (append.go:9) MOVQ 56(SP), BX 0x0089 00137 (append.go:9) ADDQ $3, CX 0x008d 00141 (append.go:9) MOVQ CX, DX 0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX Change-Id: I4ee50b18fa53865901d2d7f86c2cbb54c6fa6924 Reviewed-on: https://go-review.googlesource.com/21812 Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Keith Randall <khr@golang.org>
This commit is contained in:
parent
974c201f74
commit
6b33b0e98e
@ -337,12 +337,13 @@ var (
|
||||
memVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "mem"}}
|
||||
|
||||
// dummy nodes for temporary variables
|
||||
ptrVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "ptr"}}
|
||||
capVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "cap"}}
|
||||
typVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "typ"}}
|
||||
idataVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "idata"}}
|
||||
okVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "ok"}}
|
||||
deltaVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "delta"}}
|
||||
ptrVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "ptr"}}
|
||||
newlenVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "newlen"}}
|
||||
capVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "cap"}}
|
||||
typVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "typ"}}
|
||||
idataVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "idata"}}
|
||||
okVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "ok"}}
|
||||
deltaVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "delta"}}
|
||||
)
|
||||
|
||||
// startBlock sets the current block we're generating code in to b.
|
||||
@ -2089,15 +2090,16 @@ func (s *state) expr(n *Node) *ssa.Value {
|
||||
// exprAppend converts an OAPPEND node n to an ssa.Value, adds it to s, and returns the Value.
|
||||
func (s *state) exprAppend(n *Node) *ssa.Value {
|
||||
// append(s, e1, e2, e3). Compile like:
|
||||
// ptr,len,cap := s
|
||||
// ptr, len, cap := s
|
||||
// newlen := len + 3
|
||||
// if newlen > s.cap {
|
||||
// ptr,_,cap = growslice(s, newlen)
|
||||
// ptr, len, cap = growslice(s, newlen)
|
||||
// newlen = len + 3 // recalculate to avoid a spill
|
||||
// }
|
||||
// *(ptr+len) = e1
|
||||
// *(ptr+len+1) = e2
|
||||
// *(ptr+len+2) = e3
|
||||
// makeslice(ptr,newlen,cap)
|
||||
// makeslice(ptr, newlen, cap)
|
||||
|
||||
et := n.Type.Elem()
|
||||
pt := Ptrto(et)
|
||||
@ -2117,6 +2119,7 @@ func (s *state) exprAppend(n *Node) *ssa.Value {
|
||||
nl := s.newValue2(s.ssaOp(OADD, Types[TINT]), Types[TINT], l, s.constInt(Types[TINT], nargs))
|
||||
cmp := s.newValue2(s.ssaOp(OGT, Types[TINT]), Types[TBOOL], nl, c)
|
||||
s.vars[&ptrVar] = p
|
||||
s.vars[&newlenVar] = nl
|
||||
s.vars[&capVar] = c
|
||||
b := s.endBlock()
|
||||
b.Kind = ssa.BlockIf
|
||||
@ -2132,8 +2135,7 @@ func (s *state) exprAppend(n *Node) *ssa.Value {
|
||||
r := s.rtcall(growslice, true, []*Type{pt, Types[TINT], Types[TINT]}, taddr, p, l, c, nl)
|
||||
|
||||
s.vars[&ptrVar] = r[0]
|
||||
// Note: we don't need to read r[1], the result's length. It will be nl.
|
||||
// (or maybe we should, we just have to spill/restore nl otherwise?)
|
||||
s.vars[&newlenVar] = s.newValue2(s.ssaOp(OADD, Types[TINT]), Types[TINT], r[1], s.constInt(Types[TINT], nargs))
|
||||
s.vars[&capVar] = r[2]
|
||||
b = s.endBlock()
|
||||
b.AddEdgeTo(assign)
|
||||
@ -2154,8 +2156,9 @@ func (s *state) exprAppend(n *Node) *ssa.Value {
|
||||
}
|
||||
}
|
||||
|
||||
p = s.variable(&ptrVar, pt) // generates phi for ptr
|
||||
c = s.variable(&capVar, Types[TINT]) // generates phi for cap
|
||||
p = s.variable(&ptrVar, pt) // generates phi for ptr
|
||||
nl = s.variable(&newlenVar, Types[TINT]) // generates phi for nl
|
||||
c = s.variable(&capVar, Types[TINT]) // generates phi for cap
|
||||
p2 := s.newValue2(ssa.OpPtrIndex, pt, p, l)
|
||||
// TODO: just one write barrier call for all of these writes?
|
||||
// TODO: maybe just one writeBarrier.enabled check?
|
||||
@ -2178,6 +2181,7 @@ func (s *state) exprAppend(n *Node) *ssa.Value {
|
||||
|
||||
// make result
|
||||
delete(s.vars, &ptrVar)
|
||||
delete(s.vars, &newlenVar)
|
||||
delete(s.vars, &capVar)
|
||||
return s.newValue3(ssa.OpSliceMake, n.Type, p, nl, c)
|
||||
}
|
||||
|
@ -37,6 +37,12 @@ func makeslice(t *slicetype, len64, cap64 int64) slice {
|
||||
// It is passed the slice type, the old slice, and the desired new minimum capacity,
|
||||
// and it returns a new slice with at least that capacity, with the old data
|
||||
// copied into it.
|
||||
// The new slice's length is set to the old slice's length,
|
||||
// NOT to the new requested capacity.
|
||||
// This is for codegen convenience. The old slice's length is used immediately
|
||||
// to calculate where to write new values during an append.
|
||||
// TODO: When the old backend is gone, reconsider this decision.
|
||||
// The SSA backend might prefer the new length or to return only ptr/cap and save stack space.
|
||||
func growslice(t *slicetype, old slice, cap int) slice {
|
||||
if raceenabled {
|
||||
callerpc := getcallerpc(unsafe.Pointer(&t))
|
||||
|
Loading…
Reference in New Issue
Block a user