2014-08-12 17:48:49 -06:00
|
|
|
// Copyright 2014 The Go Authors. All rights reserved.
|
|
|
|
|
// Use of this source code is governed by a BSD-style
|
|
|
|
|
// license that can be found in the LICENSE file.
|
|
|
|
|
|
2014-12-05 17:13:20 -07:00
|
|
|
// +build ppc64 ppc64le
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2014-11-18 13:19:26 -07:00
|
|
|
#include "go_asm.h"
|
|
|
|
|
#include "go_tls.h"
|
2014-08-12 17:48:49 -06:00
|
|
|
#include "funcdata.h"
|
2014-10-27 15:27:03 -06:00
|
|
|
#include "textflag.h"
|
2015-10-08 03:34:29 -06:00
|
|
|
#include "asm_ppc64x.h"
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2014-10-27 15:27:03 -06:00
|
|
|
TEXT runtime·rt0_go(SB),NOSPLIT,$0
|
2014-12-16 15:14:00 -07:00
|
|
|
// R1 = stack; R3 = argc; R4 = argv; R13 = C TLS base pointer
|
|
|
|
|
|
2014-08-12 17:48:49 -06:00
|
|
|
// initialize essential registers
|
|
|
|
|
BL runtime·reginit(SB)
|
|
|
|
|
|
2015-10-08 03:34:29 -06:00
|
|
|
SUB $(FIXED_FRAME+16), R1
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
MOVD R2, 24(R1) // stash the TOC pointer away again now we've created a new frame
|
|
|
|
|
MOVW R3, FIXED_FRAME+0(R1) // argc
|
|
|
|
|
MOVD R4, FIXED_FRAME+8(R1) // argv
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
// create istack out of the given (operating system) stack.
|
|
|
|
|
// _cgo_init may update stackguard.
|
|
|
|
|
MOVD $runtime·g0(SB), g
|
|
|
|
|
MOVD $(-64*1024), R31
|
|
|
|
|
ADD R31, R1, R3
|
2015-01-05 09:29:21 -07:00
|
|
|
MOVD R3, g_stackguard0(g)
|
|
|
|
|
MOVD R3, g_stackguard1(g)
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVD R3, (g_stack+stack_lo)(g)
|
|
|
|
|
MOVD R1, (g_stack+stack_hi)(g)
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2014-12-16 16:34:55 -07:00
|
|
|
// if there is a _cgo_init, call it using the gcc ABI.
|
|
|
|
|
MOVD _cgo_init(SB), R12
|
|
|
|
|
CMP R0, R12
|
|
|
|
|
BEQ nocgo
|
|
|
|
|
MOVD R12, CTR // r12 = "global function entry point"
|
|
|
|
|
MOVD R13, R5 // arg 2: TLS base pointer
|
|
|
|
|
MOVD $setg_gcc<>(SB), R4 // arg 1: setg
|
|
|
|
|
MOVD g, R3 // arg 0: G
|
|
|
|
|
// C functions expect 32 bytes of space on caller stack frame
|
|
|
|
|
// and a 16-byte aligned R1
|
|
|
|
|
MOVD R1, R14 // save current stack
|
|
|
|
|
SUB $32, R1 // reserve 32 bytes
|
|
|
|
|
RLDCR $0, R1, $~15, R1 // 16-byte align
|
|
|
|
|
BL (CTR) // may clobber R0, R3-R12
|
|
|
|
|
MOVD R14, R1 // restore stack
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
MOVD 24(R1), R2
|
2014-12-16 16:34:55 -07:00
|
|
|
XOR R0, R0 // fix R0
|
|
|
|
|
|
|
|
|
|
nocgo:
|
|
|
|
|
// update stackguard after _cgo_init
|
|
|
|
|
MOVD (g_stack+stack_lo)(g), R3
|
|
|
|
|
ADD $const__StackGuard, R3
|
|
|
|
|
MOVD R3, g_stackguard0(g)
|
|
|
|
|
MOVD R3, g_stackguard1(g)
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
// set the per-goroutine and per-mach "registers"
|
|
|
|
|
MOVD $runtime·m0(SB), R3
|
|
|
|
|
|
|
|
|
|
// save m->g0 = g0
|
|
|
|
|
MOVD g, m_g0(R3)
|
|
|
|
|
// save m0 to g0->m
|
|
|
|
|
MOVD R3, g_m(g)
|
|
|
|
|
|
|
|
|
|
BL runtime·check(SB)
|
|
|
|
|
|
|
|
|
|
// args are already prepared
|
|
|
|
|
BL runtime·args(SB)
|
|
|
|
|
BL runtime·osinit(SB)
|
|
|
|
|
BL runtime·schedinit(SB)
|
|
|
|
|
|
|
|
|
|
// create a new goroutine to start program
|
2015-03-29 17:38:20 -06:00
|
|
|
MOVD $runtime·mainPC(SB), R3 // entry
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVDU R3, -8(R1)
|
|
|
|
|
MOVDU R0, -8(R1)
|
|
|
|
|
MOVDU R0, -8(R1)
|
2015-10-29 17:36:08 -06:00
|
|
|
MOVDU R0, -8(R1)
|
|
|
|
|
MOVDU R0, -8(R1)
|
|
|
|
|
MOVDU R0, -8(R1)
|
2014-08-12 17:48:49 -06:00
|
|
|
BL runtime·newproc(SB)
|
2015-10-29 17:36:08 -06:00
|
|
|
ADD $(16+FIXED_FRAME), R1
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
// start this M
|
|
|
|
|
BL runtime·mstart(SB)
|
|
|
|
|
|
|
|
|
|
MOVD R0, 1(R0)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2015-03-29 17:38:20 -06:00
|
|
|
DATA runtime·mainPC+0(SB)/8,$runtime·main(SB)
|
|
|
|
|
GLOBL runtime·mainPC(SB),RODATA,$8
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·breakpoint(SB),NOSPLIT|NOFRAME,$0-0
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD R0, 2(R0) // TODO: TD
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·asminit(SB),NOSPLIT|NOFRAME,$0-0
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT _cgo_reginit(SB),NOSPLIT|NOFRAME,$0-0
|
2014-12-16 16:34:55 -07:00
|
|
|
// crosscall_ppc64 and crosscall2 need to reginit, but can't
|
|
|
|
|
// get at the 'runtime.reginit' symbol.
|
|
|
|
|
BR runtime·reginit(SB)
|
|
|
|
|
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·reginit(SB),NOSPLIT|NOFRAME,$0-0
|
2014-08-12 17:48:49 -06:00
|
|
|
// set R0 to zero, it's expected by the toolchain
|
|
|
|
|
XOR R0, R0
|
|
|
|
|
// initialize essential FP registers
|
|
|
|
|
FMOVD $4503601774854144.0, F27
|
|
|
|
|
FMOVD $0.5, F29
|
|
|
|
|
FSUB F29, F29, F28
|
|
|
|
|
FADD F29, F29, F30
|
|
|
|
|
FADD F30, F30, F31
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* go-routine
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
// void gosave(Gobuf*)
|
|
|
|
|
// save state in Gobuf; setjmp
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·gosave(SB), NOSPLIT|NOFRAME, $0-8
|
2014-10-30 13:58:30 -06:00
|
|
|
MOVD buf+0(FP), R3
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD R1, gobuf_sp(R3)
|
|
|
|
|
MOVD LR, R31
|
|
|
|
|
MOVD R31, gobuf_pc(R3)
|
|
|
|
|
MOVD g, gobuf_g(R3)
|
|
|
|
|
MOVD R0, gobuf_lr(R3)
|
|
|
|
|
MOVD R0, gobuf_ret(R3)
|
|
|
|
|
MOVD R0, gobuf_ctxt(R3)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
// void gogo(Gobuf*)
|
|
|
|
|
// restore state from Gobuf; longjmp
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·gogo(SB), NOSPLIT|NOFRAME, $0-8
|
2014-10-30 13:58:30 -06:00
|
|
|
MOVD buf+0(FP), R5
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD gobuf_g(R5), g // make sure g is not nil
|
2014-12-16 15:14:00 -07:00
|
|
|
BL runtime·save_g(SB)
|
|
|
|
|
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD 0(g), R4
|
|
|
|
|
MOVD gobuf_sp(R5), R1
|
|
|
|
|
MOVD gobuf_lr(R5), R31
|
|
|
|
|
MOVD R31, LR
|
|
|
|
|
MOVD gobuf_ret(R5), R3
|
|
|
|
|
MOVD gobuf_ctxt(R5), R11
|
|
|
|
|
MOVD R0, gobuf_sp(R5)
|
|
|
|
|
MOVD R0, gobuf_ret(R5)
|
|
|
|
|
MOVD R0, gobuf_lr(R5)
|
|
|
|
|
MOVD R0, gobuf_ctxt(R5)
|
|
|
|
|
CMP R0, R0 // set condition codes for == test, needed by stack split
|
2015-10-08 03:49:39 -06:00
|
|
|
MOVD gobuf_pc(R5), R12
|
|
|
|
|
MOVD R12, CTR
|
2014-08-12 17:48:49 -06:00
|
|
|
BR (CTR)
|
|
|
|
|
|
2014-10-27 15:27:03 -06:00
|
|
|
// void mcall(fn func(*g))
|
2014-08-12 17:48:49 -06:00
|
|
|
// Switch to m->g0's stack, call fn(g).
|
|
|
|
|
// Fn must never return. It should gogo(&g->sched)
|
|
|
|
|
// to keep running g.
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·mcall(SB), NOSPLIT|NOFRAME, $0-8
|
2014-08-12 17:48:49 -06:00
|
|
|
// Save caller state in g->sched
|
|
|
|
|
MOVD R1, (g_sched+gobuf_sp)(g)
|
|
|
|
|
MOVD LR, R31
|
|
|
|
|
MOVD R31, (g_sched+gobuf_pc)(g)
|
|
|
|
|
MOVD R0, (g_sched+gobuf_lr)(g)
|
|
|
|
|
MOVD g, (g_sched+gobuf_g)(g)
|
|
|
|
|
|
|
|
|
|
// Switch to m->g0 & its stack, call fn.
|
|
|
|
|
MOVD g, R3
|
|
|
|
|
MOVD g_m(g), R8
|
|
|
|
|
MOVD m_g0(R8), g
|
2014-12-16 15:14:00 -07:00
|
|
|
BL runtime·save_g(SB)
|
2014-08-12 17:48:49 -06:00
|
|
|
CMP g, R3
|
|
|
|
|
BNE 2(PC)
|
|
|
|
|
BR runtime·badmcall(SB)
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVD fn+0(FP), R11 // context
|
2015-10-08 03:49:39 -06:00
|
|
|
MOVD 0(R11), R12 // code pointer
|
|
|
|
|
MOVD R12, CTR
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVD (g_sched+gobuf_sp)(g), R1 // sp = m->g0->sched.sp
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVDU R3, -8(R1)
|
|
|
|
|
MOVDU R0, -8(R1)
|
2015-10-29 17:36:08 -06:00
|
|
|
MOVDU R0, -8(R1)
|
|
|
|
|
MOVDU R0, -8(R1)
|
|
|
|
|
MOVDU R0, -8(R1)
|
2014-08-12 17:48:49 -06:00
|
|
|
BL (CTR)
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
MOVD 24(R1), R2
|
2014-08-12 17:48:49 -06:00
|
|
|
BR runtime·badmcall2(SB)
|
|
|
|
|
|
2014-11-18 13:50:36 -07:00
|
|
|
// systemstack_switch is a dummy routine that systemstack leaves at the bottom
|
2014-08-12 19:55:43 -06:00
|
|
|
// of the G stack. We need to distinguish the routine that
|
|
|
|
|
// lives at the bottom of the G stack from the one that lives
|
2014-11-18 13:50:36 -07:00
|
|
|
// at the top of the system stack because the one at the top of
|
|
|
|
|
// the system stack terminates the stack walk (see topofstack()).
|
|
|
|
|
TEXT runtime·systemstack_switch(SB), NOSPLIT, $0-0
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
// We have several undefs here so that 16 bytes past
|
|
|
|
|
// $runtime·systemstack_switch lies within them whether or not the
|
|
|
|
|
// instructions that derive r2 from r12 are there.
|
|
|
|
|
UNDEF
|
|
|
|
|
UNDEF
|
2014-08-12 19:55:43 -06:00
|
|
|
UNDEF
|
|
|
|
|
BL (LR) // make sure this function is not leaf
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 19:55:43 -06:00
|
|
|
|
2014-11-18 13:50:36 -07:00
|
|
|
// func systemstack(fn func())
|
|
|
|
|
TEXT runtime·systemstack(SB), NOSPLIT, $0-8
|
2014-08-12 19:55:43 -06:00
|
|
|
MOVD fn+0(FP), R3 // R3 = fn
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVD R3, R11 // context
|
2014-08-12 19:55:43 -06:00
|
|
|
MOVD g_m(g), R4 // R4 = m
|
2014-10-27 15:27:03 -06:00
|
|
|
|
2014-11-18 13:50:36 -07:00
|
|
|
MOVD m_gsignal(R4), R5 // R5 = gsignal
|
|
|
|
|
CMP g, R5
|
|
|
|
|
BEQ noswitch
|
|
|
|
|
|
2014-08-12 19:55:43 -06:00
|
|
|
MOVD m_g0(R4), R5 // R5 = g0
|
|
|
|
|
CMP g, R5
|
2014-11-18 13:50:36 -07:00
|
|
|
BEQ noswitch
|
2014-08-12 19:55:43 -06:00
|
|
|
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVD m_curg(R4), R6
|
|
|
|
|
CMP g, R6
|
2014-11-18 13:50:36 -07:00
|
|
|
BEQ switch
|
2014-10-27 15:27:03 -06:00
|
|
|
|
2014-11-18 13:50:36 -07:00
|
|
|
// Bad: g is not gsignal, not g0, not curg. What is it?
|
2014-10-27 15:27:03 -06:00
|
|
|
// Hide call from linker nosplit analysis.
|
2015-10-08 03:49:39 -06:00
|
|
|
MOVD $runtime·badsystemstack(SB), R12
|
|
|
|
|
MOVD R12, CTR
|
2014-10-27 15:27:03 -06:00
|
|
|
BL (CTR)
|
|
|
|
|
|
2014-11-18 13:50:36 -07:00
|
|
|
switch:
|
2014-08-12 19:55:43 -06:00
|
|
|
// save our state in g->sched. Pretend to
|
2014-11-18 13:50:36 -07:00
|
|
|
// be systemstack_switch if the G stack is scanned.
|
|
|
|
|
MOVD $runtime·systemstack_switch(SB), R6
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
ADD $16, R6 // get past prologue (including r2-setting instructions when they're there)
|
2014-08-12 19:55:43 -06:00
|
|
|
MOVD R6, (g_sched+gobuf_pc)(g)
|
|
|
|
|
MOVD R1, (g_sched+gobuf_sp)(g)
|
|
|
|
|
MOVD R0, (g_sched+gobuf_lr)(g)
|
|
|
|
|
MOVD g, (g_sched+gobuf_g)(g)
|
|
|
|
|
|
|
|
|
|
// switch to g0
|
|
|
|
|
MOVD R5, g
|
2014-12-16 15:14:00 -07:00
|
|
|
BL runtime·save_g(SB)
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVD (g_sched+gobuf_sp)(g), R3
|
2014-11-18 13:50:36 -07:00
|
|
|
// make it look like mstart called systemstack on g0, to stop traceback
|
2015-10-08 03:34:29 -06:00
|
|
|
SUB $FIXED_FRAME, R3
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVD $runtime·mstart(SB), R4
|
|
|
|
|
MOVD R4, 0(R3)
|
|
|
|
|
MOVD R3, R1
|
2014-08-12 19:55:43 -06:00
|
|
|
|
|
|
|
|
// call target function
|
2015-10-08 03:49:39 -06:00
|
|
|
MOVD 0(R11), R12 // code pointer
|
|
|
|
|
MOVD R12, CTR
|
2014-08-12 19:55:43 -06:00
|
|
|
BL (CTR)
|
|
|
|
|
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
// restore TOC pointer. It seems unlikely that we will use systemstack
|
|
|
|
|
// to call a function defined in another module, but the results of
|
|
|
|
|
// doing so would be so confusing that it's worth doing this.
|
|
|
|
|
MOVD g_m(g), R3
|
|
|
|
|
MOVD m_curg(R3), g
|
|
|
|
|
MOVD (g_sched+gobuf_sp)(g), R3
|
|
|
|
|
MOVD 24(R3), R2
|
2014-08-12 19:55:43 -06:00
|
|
|
// switch back to g
|
|
|
|
|
MOVD g_m(g), R3
|
|
|
|
|
MOVD m_curg(R3), g
|
2014-12-16 15:14:00 -07:00
|
|
|
BL runtime·save_g(SB)
|
2014-08-12 19:55:43 -06:00
|
|
|
MOVD (g_sched+gobuf_sp)(g), R1
|
|
|
|
|
MOVD R0, (g_sched+gobuf_sp)(g)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 19:55:43 -06:00
|
|
|
|
2014-11-18 13:50:36 -07:00
|
|
|
noswitch:
|
2014-08-12 19:55:43 -06:00
|
|
|
// already on m stack, just call directly
|
2015-10-08 03:49:39 -06:00
|
|
|
MOVD 0(R11), R12 // code pointer
|
|
|
|
|
MOVD R12, CTR
|
2014-08-12 19:55:43 -06:00
|
|
|
BL (CTR)
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
MOVD 24(R1), R2
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 19:55:43 -06:00
|
|
|
|
2014-08-12 17:48:49 -06:00
|
|
|
/*
|
|
|
|
|
* support for morestack
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
// Called during function prolog when more stack is needed.
|
|
|
|
|
// Caller has already loaded:
|
|
|
|
|
// R3: framesize, R4: argsize, R5: LR
|
|
|
|
|
//
|
|
|
|
|
// The traceback routines see morestack on a g0 as being
|
|
|
|
|
// the top of a stack (for example, morestack calling newstack
|
|
|
|
|
// calling the scheduler calling newm calling gc), so we must
|
|
|
|
|
// record an argument size. For that purpose, it has no arguments.
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·morestack(SB),NOSPLIT|NOFRAME,$0-0
|
2014-08-12 17:48:49 -06:00
|
|
|
// Cannot grow scheduler stack (m->g0).
|
|
|
|
|
MOVD g_m(g), R7
|
|
|
|
|
MOVD m_g0(R7), R8
|
|
|
|
|
CMP g, R8
|
|
|
|
|
BNE 2(PC)
|
|
|
|
|
BL runtime·abort(SB)
|
|
|
|
|
|
2014-10-27 15:27:03 -06:00
|
|
|
// Cannot grow signal stack (m->gsignal).
|
|
|
|
|
MOVD m_gsignal(R7), R8
|
|
|
|
|
CMP g, R8
|
|
|
|
|
BNE 2(PC)
|
|
|
|
|
BL runtime·abort(SB)
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
// Called from f.
|
|
|
|
|
// Set g->sched to context in f.
|
|
|
|
|
MOVD R11, (g_sched+gobuf_ctxt)(g)
|
|
|
|
|
MOVD R1, (g_sched+gobuf_sp)(g)
|
|
|
|
|
MOVD LR, R8
|
|
|
|
|
MOVD R8, (g_sched+gobuf_pc)(g)
|
|
|
|
|
MOVD R5, (g_sched+gobuf_lr)(g)
|
|
|
|
|
|
|
|
|
|
// Called from f.
|
|
|
|
|
// Set m->morebuf to f's caller.
|
|
|
|
|
MOVD R5, (m_morebuf+gobuf_pc)(R7) // f's caller's PC
|
|
|
|
|
MOVD R1, (m_morebuf+gobuf_sp)(R7) // f's caller's SP
|
|
|
|
|
MOVD g, (m_morebuf+gobuf_g)(R7)
|
|
|
|
|
|
|
|
|
|
// Call newstack on m->g0's stack.
|
|
|
|
|
MOVD m_g0(R7), g
|
2014-12-16 15:14:00 -07:00
|
|
|
BL runtime·save_g(SB)
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD (g_sched+gobuf_sp)(g), R1
|
|
|
|
|
BL runtime·newstack(SB)
|
|
|
|
|
|
|
|
|
|
// Not reached, but make sure the return PC from the call to newstack
|
|
|
|
|
// is still in this function, and not the beginning of the next.
|
|
|
|
|
UNDEF
|
|
|
|
|
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·morestack_noctxt(SB),NOSPLIT|NOFRAME,$0-0
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD R0, R11
|
|
|
|
|
BR runtime·morestack(SB)
|
|
|
|
|
|
runtime: implement GC stack barriers
This commit implements stack barriers to minimize the amount of
stack re-scanning that must be done during mark termination.
Currently the GC scans stacks of active goroutines twice during every
GC cycle: once at the beginning during root discovery and once at the
end during mark termination. The second scan happens while the world
is stopped and guarantees that we've seen all of the roots (since
there are no write barriers on writes to local stack
variables). However, this means pause time is proportional to stack
size. In particularly recursive programs, this can drive pause time up
past our 10ms goal (e.g., it takes about 150ms to scan a 50MB heap).
Re-scanning the entire stack is rarely necessary, especially for large
stacks, because usually most of the frames on the stack were not
active between the first and second scans and hence any changes to
these frames (via non-escaping pointers passed down the stack) were
tracked by write barriers.
To efficiently track how far a stack has been unwound since the first
scan (and, hence, how much needs to be re-scanned), this commit
introduces stack barriers. During the first scan, at exponentially
spaced points in each stack, the scan overwrites return PCs with the
PC of the stack barrier function. When "returned" to, the stack
barrier function records how far the stack has unwound and jumps to
the original return PC for that point in the stack. Then the second
scan only needs to proceed as far as the lowest barrier that hasn't
been hit.
For deeply recursive programs, this substantially reduces mark
termination time (and hence pause time). For the goscheme example
linked in issue #10898, prior to this change, mark termination times
were typically between 100 and 500ms; with this change, mark
termination times are typically between 10 and 20ms. As a result of
the reduced stack scanning work, this reduces overall execution time
of the goscheme example by 20%.
Fixes #10898.
The effect of this on programs that are not deeply recursive is
minimal:
name old time/op new time/op delta
BinaryTree17 3.16s ± 2% 3.26s ± 1% +3.31% (p=0.000 n=19+19)
Fannkuch11 2.42s ± 1% 2.48s ± 1% +2.24% (p=0.000 n=17+19)
FmtFprintfEmpty 50.0ns ± 3% 49.8ns ± 1% ~ (p=0.534 n=20+19)
FmtFprintfString 173ns ± 0% 175ns ± 0% +1.49% (p=0.000 n=16+19)
FmtFprintfInt 170ns ± 1% 175ns ± 1% +2.97% (p=0.000 n=20+19)
FmtFprintfIntInt 288ns ± 0% 295ns ± 0% +2.73% (p=0.000 n=16+19)
FmtFprintfPrefixedInt 242ns ± 1% 252ns ± 1% +4.13% (p=0.000 n=18+18)
FmtFprintfFloat 324ns ± 0% 323ns ± 0% -0.36% (p=0.000 n=20+19)
FmtManyArgs 1.14µs ± 0% 1.12µs ± 1% -1.01% (p=0.000 n=18+19)
GobDecode 8.88ms ± 1% 8.87ms ± 0% ~ (p=0.480 n=19+18)
GobEncode 6.80ms ± 1% 6.85ms ± 0% +0.82% (p=0.000 n=20+18)
Gzip 363ms ± 1% 363ms ± 1% ~ (p=0.077 n=18+20)
Gunzip 90.6ms ± 0% 90.0ms ± 1% -0.71% (p=0.000 n=17+18)
HTTPClientServer 51.5µs ± 1% 50.8µs ± 1% -1.32% (p=0.000 n=18+18)
JSONEncode 17.0ms ± 0% 17.1ms ± 0% +0.40% (p=0.000 n=18+17)
JSONDecode 61.8ms ± 0% 63.8ms ± 1% +3.11% (p=0.000 n=18+17)
Mandelbrot200 3.84ms ± 0% 3.84ms ± 1% ~ (p=0.583 n=19+19)
GoParse 3.71ms ± 1% 3.72ms ± 1% ~ (p=0.159 n=18+19)
RegexpMatchEasy0_32 100ns ± 0% 100ns ± 1% -0.19% (p=0.033 n=17+19)
RegexpMatchEasy0_1K 342ns ± 1% 331ns ± 0% -3.41% (p=0.000 n=19+19)
RegexpMatchEasy1_32 82.5ns ± 0% 81.7ns ± 0% -0.98% (p=0.000 n=18+18)
RegexpMatchEasy1_1K 505ns ± 0% 494ns ± 1% -2.16% (p=0.000 n=18+18)
RegexpMatchMedium_32 137ns ± 1% 137ns ± 1% -0.24% (p=0.048 n=20+18)
RegexpMatchMedium_1K 41.6µs ± 0% 41.3µs ± 1% -0.57% (p=0.004 n=18+20)
RegexpMatchHard_32 2.11µs ± 0% 2.11µs ± 1% +0.20% (p=0.037 n=17+19)
RegexpMatchHard_1K 63.9µs ± 2% 63.3µs ± 0% -0.99% (p=0.000 n=20+17)
Revcomp 560ms ± 1% 522ms ± 0% -6.87% (p=0.000 n=18+16)
Template 75.0ms ± 0% 75.1ms ± 1% +0.18% (p=0.013 n=18+19)
TimeParse 358ns ± 1% 364ns ± 0% +1.74% (p=0.000 n=20+15)
TimeFormat 360ns ± 0% 372ns ± 0% +3.55% (p=0.000 n=20+18)
Change-Id: If8a9bfae6c128d15a4f405e02bcfa50129df82a2
Reviewed-on: https://go-review.googlesource.com/10314
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2015-05-20 14:30:49 -06:00
|
|
|
TEXT runtime·stackBarrier(SB),NOSPLIT,$0
|
|
|
|
|
// We came here via a RET to an overwritten LR.
|
|
|
|
|
// R3 may be live. Other registers are available.
|
|
|
|
|
|
|
|
|
|
// Get the original return PC, g.stkbar[g.stkbarPos].savedLRVal.
|
|
|
|
|
MOVD (g_stkbar+slice_array)(g), R4
|
|
|
|
|
MOVD g_stkbarPos(g), R5
|
|
|
|
|
MOVD $stkbar__size, R6
|
|
|
|
|
MULLD R5, R6
|
|
|
|
|
ADD R4, R6
|
|
|
|
|
MOVD stkbar_savedLRVal(R6), R6
|
|
|
|
|
// Record that this stack barrier was hit.
|
|
|
|
|
ADD $1, R5
|
|
|
|
|
MOVD R5, g_stkbarPos(g)
|
|
|
|
|
// Jump to the original return PC.
|
|
|
|
|
MOVD R6, CTR
|
|
|
|
|
BR (CTR)
|
|
|
|
|
|
2014-10-27 15:27:03 -06:00
|
|
|
// reflectcall: call a function with the given argument list
|
2014-12-30 11:59:55 -07:00
|
|
|
// func call(argtype *_type, f *FuncVal, arg *byte, argsize, retoffset uint32).
|
2014-08-12 17:48:49 -06:00
|
|
|
// we don't have variable-sized frames, so we use a small number
|
|
|
|
|
// of constant-sized-frame functions to encode a few bits of size in the pc.
|
|
|
|
|
// Caution: ugly multiline assembly macros in your future!
|
|
|
|
|
|
|
|
|
|
#define DISPATCH(NAME,MAXSIZE) \
|
|
|
|
|
MOVD $MAXSIZE, R31; \
|
|
|
|
|
CMP R3, R31; \
|
|
|
|
|
BGT 4(PC); \
|
2015-10-08 03:49:39 -06:00
|
|
|
MOVD $NAME(SB), R12; \
|
|
|
|
|
MOVD R12, CTR; \
|
2014-08-12 17:48:49 -06:00
|
|
|
BR (CTR)
|
2014-10-27 15:27:03 -06:00
|
|
|
// Note: can't just "BR NAME(SB)" - bad inlining results.
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2014-12-22 11:27:53 -07:00
|
|
|
TEXT reflect·call(SB), NOSPLIT, $0-0
|
|
|
|
|
BR ·reflectcall(SB)
|
|
|
|
|
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT ·reflectcall(SB), NOSPLIT|NOFRAME, $0-32
|
2014-12-30 11:59:55 -07:00
|
|
|
MOVWZ argsize+24(FP), R3
|
|
|
|
|
// NOTE(rsc): No call16, because CALLFN needs four words
|
|
|
|
|
// of argument space to invoke callwritebarrier.
|
2014-10-27 15:27:03 -06:00
|
|
|
DISPATCH(runtime·call32, 32)
|
|
|
|
|
DISPATCH(runtime·call64, 64)
|
|
|
|
|
DISPATCH(runtime·call128, 128)
|
|
|
|
|
DISPATCH(runtime·call256, 256)
|
|
|
|
|
DISPATCH(runtime·call512, 512)
|
|
|
|
|
DISPATCH(runtime·call1024, 1024)
|
|
|
|
|
DISPATCH(runtime·call2048, 2048)
|
|
|
|
|
DISPATCH(runtime·call4096, 4096)
|
|
|
|
|
DISPATCH(runtime·call8192, 8192)
|
|
|
|
|
DISPATCH(runtime·call16384, 16384)
|
|
|
|
|
DISPATCH(runtime·call32768, 32768)
|
|
|
|
|
DISPATCH(runtime·call65536, 65536)
|
|
|
|
|
DISPATCH(runtime·call131072, 131072)
|
|
|
|
|
DISPATCH(runtime·call262144, 262144)
|
|
|
|
|
DISPATCH(runtime·call524288, 524288)
|
|
|
|
|
DISPATCH(runtime·call1048576, 1048576)
|
|
|
|
|
DISPATCH(runtime·call2097152, 2097152)
|
|
|
|
|
DISPATCH(runtime·call4194304, 4194304)
|
|
|
|
|
DISPATCH(runtime·call8388608, 8388608)
|
|
|
|
|
DISPATCH(runtime·call16777216, 16777216)
|
|
|
|
|
DISPATCH(runtime·call33554432, 33554432)
|
|
|
|
|
DISPATCH(runtime·call67108864, 67108864)
|
|
|
|
|
DISPATCH(runtime·call134217728, 134217728)
|
|
|
|
|
DISPATCH(runtime·call268435456, 268435456)
|
|
|
|
|
DISPATCH(runtime·call536870912, 536870912)
|
|
|
|
|
DISPATCH(runtime·call1073741824, 1073741824)
|
2015-10-08 03:49:39 -06:00
|
|
|
MOVD $runtime·badreflectcall(SB), R12
|
|
|
|
|
MOVD R12, CTR
|
2014-08-12 17:48:49 -06:00
|
|
|
BR (CTR)
|
|
|
|
|
|
|
|
|
|
#define CALLFN(NAME,MAXSIZE) \
|
2014-10-27 15:27:03 -06:00
|
|
|
TEXT NAME(SB), WRAPPER, $MAXSIZE-24; \
|
|
|
|
|
NO_LOCAL_POINTERS; \
|
2014-08-12 17:48:49 -06:00
|
|
|
/* copy arguments to stack */ \
|
2014-12-30 11:59:55 -07:00
|
|
|
MOVD arg+16(FP), R3; \
|
|
|
|
|
MOVWZ argsize+24(FP), R4; \
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD R1, R5; \
|
2015-10-08 03:34:29 -06:00
|
|
|
ADD $(FIXED_FRAME-1), R5; \
|
2014-08-12 17:48:49 -06:00
|
|
|
SUB $1, R3; \
|
|
|
|
|
ADD R5, R4; \
|
|
|
|
|
CMP R5, R4; \
|
|
|
|
|
BEQ 4(PC); \
|
|
|
|
|
MOVBZU 1(R3), R6; \
|
|
|
|
|
MOVBZU R6, 1(R5); \
|
|
|
|
|
BR -4(PC); \
|
|
|
|
|
/* call function */ \
|
2014-12-30 11:59:55 -07:00
|
|
|
MOVD f+8(FP), R11; \
|
2015-10-08 03:49:39 -06:00
|
|
|
MOVD (R11), R12; \
|
|
|
|
|
MOVD R12, CTR; \
|
2014-08-12 17:48:49 -06:00
|
|
|
PCDATA $PCDATA_StackMapIndex, $0; \
|
|
|
|
|
BL (CTR); \
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
MOVD 24(R1), R2; \
|
2014-08-12 17:48:49 -06:00
|
|
|
/* copy return values back */ \
|
2014-12-30 11:59:55 -07:00
|
|
|
MOVD arg+16(FP), R3; \
|
|
|
|
|
MOVWZ n+24(FP), R4; \
|
|
|
|
|
MOVWZ retoffset+28(FP), R6; \
|
2014-08-13 14:26:43 -06:00
|
|
|
MOVD R1, R5; \
|
|
|
|
|
ADD R6, R5; \
|
|
|
|
|
ADD R6, R3; \
|
|
|
|
|
SUB R6, R4; \
|
2015-10-08 03:34:29 -06:00
|
|
|
ADD $(FIXED_FRAME-1), R5; \
|
2014-08-12 17:48:49 -06:00
|
|
|
SUB $1, R3; \
|
2014-08-13 14:26:43 -06:00
|
|
|
ADD R5, R4; \
|
2014-12-30 11:59:55 -07:00
|
|
|
loop: \
|
2014-08-13 14:26:43 -06:00
|
|
|
CMP R5, R4; \
|
2014-12-30 11:59:55 -07:00
|
|
|
BEQ end; \
|
2014-08-13 14:26:43 -06:00
|
|
|
MOVBZU 1(R5), R6; \
|
|
|
|
|
MOVBZU R6, 1(R3); \
|
2014-12-30 11:59:55 -07:00
|
|
|
BR loop; \
|
|
|
|
|
end: \
|
|
|
|
|
/* execute write barrier updates */ \
|
|
|
|
|
MOVD argtype+0(FP), R7; \
|
|
|
|
|
MOVD arg+16(FP), R3; \
|
|
|
|
|
MOVWZ n+24(FP), R4; \
|
|
|
|
|
MOVWZ retoffset+28(FP), R6; \
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD R7, FIXED_FRAME+0(R1); \
|
|
|
|
|
MOVD R3, FIXED_FRAME+8(R1); \
|
|
|
|
|
MOVD R4, FIXED_FRAME+16(R1); \
|
|
|
|
|
MOVD R6, FIXED_FRAME+24(R1); \
|
2014-12-30 11:59:55 -07:00
|
|
|
BL runtime·callwritebarrier(SB); \
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2014-10-27 15:27:03 -06:00
|
|
|
CALLFN(·call32, 32)
|
|
|
|
|
CALLFN(·call64, 64)
|
|
|
|
|
CALLFN(·call128, 128)
|
|
|
|
|
CALLFN(·call256, 256)
|
|
|
|
|
CALLFN(·call512, 512)
|
|
|
|
|
CALLFN(·call1024, 1024)
|
|
|
|
|
CALLFN(·call2048, 2048)
|
|
|
|
|
CALLFN(·call4096, 4096)
|
|
|
|
|
CALLFN(·call8192, 8192)
|
|
|
|
|
CALLFN(·call16384, 16384)
|
|
|
|
|
CALLFN(·call32768, 32768)
|
|
|
|
|
CALLFN(·call65536, 65536)
|
|
|
|
|
CALLFN(·call131072, 131072)
|
|
|
|
|
CALLFN(·call262144, 262144)
|
|
|
|
|
CALLFN(·call524288, 524288)
|
|
|
|
|
CALLFN(·call1048576, 1048576)
|
|
|
|
|
CALLFN(·call2097152, 2097152)
|
|
|
|
|
CALLFN(·call4194304, 4194304)
|
|
|
|
|
CALLFN(·call8388608, 8388608)
|
|
|
|
|
CALLFN(·call16777216, 16777216)
|
|
|
|
|
CALLFN(·call33554432, 33554432)
|
|
|
|
|
CALLFN(·call67108864, 67108864)
|
|
|
|
|
CALLFN(·call134217728, 134217728)
|
|
|
|
|
CALLFN(·call268435456, 268435456)
|
|
|
|
|
CALLFN(·call536870912, 536870912)
|
|
|
|
|
CALLFN(·call1073741824, 1073741824)
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
TEXT runtime·procyield(SB),NOSPLIT,$0-0
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2014-10-27 15:27:03 -06:00
|
|
|
// void jmpdefer(fv, sp);
|
2014-08-12 17:48:49 -06:00
|
|
|
// called from deferreturn.
|
|
|
|
|
// 1. grab stored LR for caller
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
// 2. sub 8 bytes to get back to either nop or toc reload before deferreturn
|
2014-08-12 17:48:49 -06:00
|
|
|
// 3. BR to fn
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
// When dynamically linking Go, it is not sufficient to rewind to the BL
|
|
|
|
|
// deferreturn -- we might be jumping between modules and so we need to reset
|
|
|
|
|
// the TOC pointer in r2. To do this, codegen inserts MOVD 24(R1), R2 *before*
|
|
|
|
|
// the BL deferreturn and jmpdefer rewinds to that.
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·jmpdefer(SB), NOSPLIT|NOFRAME, $0-16
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD 0(R1), R31
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
SUB $8, R31
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD R31, LR
|
|
|
|
|
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVD fv+0(FP), R11
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD argp+8(FP), R1
|
2015-10-08 03:34:29 -06:00
|
|
|
SUB $FIXED_FRAME, R1
|
2015-10-08 03:49:39 -06:00
|
|
|
MOVD 0(R11), R12
|
|
|
|
|
MOVD R12, CTR
|
2014-08-12 17:48:49 -06:00
|
|
|
BR (CTR)
|
|
|
|
|
|
|
|
|
|
// Save state of caller into g->sched. Smashes R31.
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT gosave<>(SB),NOSPLIT|NOFRAME,$0
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD LR, R31
|
|
|
|
|
MOVD R31, (g_sched+gobuf_pc)(g)
|
|
|
|
|
MOVD R1, (g_sched+gobuf_sp)(g)
|
|
|
|
|
MOVD R0, (g_sched+gobuf_lr)(g)
|
|
|
|
|
MOVD R0, (g_sched+gobuf_ret)(g)
|
|
|
|
|
MOVD R0, (g_sched+gobuf_ctxt)(g)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2015-04-27 01:32:23 -06:00
|
|
|
// func asmcgocall(fn, arg unsafe.Pointer) int32
|
2014-08-12 17:48:49 -06:00
|
|
|
// Call fn(arg) on the scheduler stack,
|
|
|
|
|
// aligned appropriately for the gcc ABI.
|
2015-04-27 01:32:23 -06:00
|
|
|
// See cgocall.go for more details.
|
|
|
|
|
TEXT ·asmcgocall(SB),NOSPLIT,$0-20
|
2014-12-16 16:34:55 -07:00
|
|
|
MOVD fn+0(FP), R3
|
|
|
|
|
MOVD arg+8(FP), R4
|
|
|
|
|
|
2015-10-08 03:49:39 -06:00
|
|
|
MOVD R1, R7 // save original stack pointer
|
2014-12-16 16:34:55 -07:00
|
|
|
MOVD g, R5
|
|
|
|
|
|
|
|
|
|
// Figure out if we need to switch to m->g0 stack.
|
|
|
|
|
// We get called to create new OS threads too, and those
|
|
|
|
|
// come in on the m->g0 stack already.
|
|
|
|
|
MOVD g_m(g), R6
|
|
|
|
|
MOVD m_g0(R6), R6
|
|
|
|
|
CMP R6, g
|
|
|
|
|
BEQ g0
|
|
|
|
|
BL gosave<>(SB)
|
|
|
|
|
MOVD R6, g
|
|
|
|
|
BL runtime·save_g(SB)
|
|
|
|
|
MOVD (g_sched+gobuf_sp)(g), R1
|
|
|
|
|
|
|
|
|
|
// Now on a scheduling stack (a pthread-created stack).
|
|
|
|
|
g0:
|
|
|
|
|
// Save room for two of our pointers, plus 32 bytes of callee
|
|
|
|
|
// save area that lives on the caller stack.
|
|
|
|
|
SUB $48, R1
|
|
|
|
|
RLDCR $0, R1, $~15, R1 // 16-byte alignment for gcc ABI
|
|
|
|
|
MOVD R5, 40(R1) // save old g on stack
|
|
|
|
|
MOVD (g_stack+stack_hi)(R5), R5
|
2015-10-08 03:49:39 -06:00
|
|
|
SUB R7, R5
|
2014-12-16 16:34:55 -07:00
|
|
|
MOVD R5, 32(R1) // save depth in old g stack (can't just save SP, as stack might be copied during a callback)
|
|
|
|
|
MOVD R0, 0(R1) // clear back chain pointer (TODO can we give it real back trace information?)
|
|
|
|
|
// This is a "global call", so put the global entry point in r12
|
|
|
|
|
MOVD R3, R12
|
|
|
|
|
MOVD R12, CTR
|
|
|
|
|
MOVD R4, R3 // arg in r3
|
|
|
|
|
BL (CTR)
|
|
|
|
|
|
|
|
|
|
// C code can clobber R0, so set it back to 0. F27-F31 are
|
|
|
|
|
// callee save, so we don't need to recover those.
|
|
|
|
|
XOR R0, R0
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
// Restore g, stack pointer, toc pointer.
|
|
|
|
|
// R3 is errno, so don't touch it
|
2014-12-16 16:34:55 -07:00
|
|
|
MOVD 40(R1), g
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
MOVD (g_stack+stack_hi)(g), R5
|
|
|
|
|
MOVD 32(R1), R6
|
|
|
|
|
SUB R6, R5
|
|
|
|
|
MOVD 24(R5), R2
|
2014-12-16 16:34:55 -07:00
|
|
|
BL runtime·save_g(SB)
|
|
|
|
|
MOVD (g_stack+stack_hi)(g), R5
|
|
|
|
|
MOVD 32(R1), R6
|
|
|
|
|
SUB R6, R5
|
|
|
|
|
MOVD R5, R1
|
2015-04-27 01:32:23 -06:00
|
|
|
|
|
|
|
|
MOVW R3, ret+16(FP)
|
2014-12-16 16:34:55 -07:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
// cgocallback(void (*fn)(void*), void *frame, uintptr framesize)
|
|
|
|
|
// Turn the fn into a Go func (by taking its address) and call
|
|
|
|
|
// cgocallback_gofunc.
|
|
|
|
|
TEXT runtime·cgocallback(SB),NOSPLIT,$24-24
|
2014-12-16 16:34:55 -07:00
|
|
|
MOVD $fn+0(FP), R3
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD R3, FIXED_FRAME+0(R1)
|
2014-12-16 16:34:55 -07:00
|
|
|
MOVD frame+8(FP), R3
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD R3, FIXED_FRAME+8(R1)
|
2014-12-16 16:34:55 -07:00
|
|
|
MOVD framesize+16(FP), R3
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD R3, FIXED_FRAME+16(R1)
|
2015-10-08 03:49:39 -06:00
|
|
|
MOVD $runtime·cgocallback_gofunc(SB), R12
|
|
|
|
|
MOVD R12, CTR
|
2014-12-16 16:34:55 -07:00
|
|
|
BL (CTR)
|
|
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
// cgocallback_gofunc(FuncVal*, void *frame, uintptr framesize)
|
2015-04-27 01:32:23 -06:00
|
|
|
// See cgocall.go for more details.
|
2014-12-16 16:34:55 -07:00
|
|
|
TEXT ·cgocallback_gofunc(SB),NOSPLIT,$16-24
|
|
|
|
|
NO_LOCAL_POINTERS
|
|
|
|
|
|
|
|
|
|
// Load m and g from thread-local storage.
|
|
|
|
|
MOVB runtime·iscgo(SB), R3
|
|
|
|
|
CMP R3, $0
|
|
|
|
|
BEQ nocgo
|
|
|
|
|
BL runtime·load_g(SB)
|
|
|
|
|
nocgo:
|
|
|
|
|
|
|
|
|
|
// If g is nil, Go did not create the current thread.
|
|
|
|
|
// Call needm to obtain one for temporary use.
|
|
|
|
|
// In this case, we're running on the thread stack, so there's
|
|
|
|
|
// lots of space, but the linker doesn't know. Hide the call from
|
|
|
|
|
// the linker analysis by using an indirect call.
|
|
|
|
|
CMP g, $0
|
|
|
|
|
BNE havem
|
|
|
|
|
MOVD g, savedm-8(SP) // g is zero, so is m.
|
2015-10-08 03:49:39 -06:00
|
|
|
MOVD $runtime·needm(SB), R12
|
|
|
|
|
MOVD R12, CTR
|
2014-12-16 16:34:55 -07:00
|
|
|
BL (CTR)
|
|
|
|
|
|
|
|
|
|
// Set m->sched.sp = SP, so that if a panic happens
|
|
|
|
|
// during the function we are about to execute, it will
|
|
|
|
|
// have a valid SP to run on the g0 stack.
|
|
|
|
|
// The next few lines (after the havem label)
|
|
|
|
|
// will save this SP onto the stack and then write
|
|
|
|
|
// the same SP back to m->sched.sp. That seems redundant,
|
|
|
|
|
// but if an unrecovered panic happens, unwindm will
|
|
|
|
|
// restore the g->sched.sp from the stack location
|
|
|
|
|
// and then systemstack will try to use it. If we don't set it here,
|
|
|
|
|
// that restored SP will be uninitialized (typically 0) and
|
|
|
|
|
// will not be usable.
|
|
|
|
|
MOVD g_m(g), R3
|
|
|
|
|
MOVD m_g0(R3), R3
|
|
|
|
|
MOVD R1, (g_sched+gobuf_sp)(R3)
|
|
|
|
|
|
|
|
|
|
havem:
|
|
|
|
|
MOVD g_m(g), R8
|
|
|
|
|
MOVD R8, savedm-8(SP)
|
|
|
|
|
// Now there's a valid m, and we're running on its m->g0.
|
|
|
|
|
// Save current m->g0->sched.sp on stack and then set it to SP.
|
|
|
|
|
// Save current sp in m->g0->sched.sp in preparation for
|
|
|
|
|
// switch back to m->curg stack.
|
|
|
|
|
// NOTE: unwindm knows that the saved g->sched.sp is at 8(R1) aka savedsp-16(SP).
|
|
|
|
|
MOVD m_g0(R8), R3
|
|
|
|
|
MOVD (g_sched+gobuf_sp)(R3), R4
|
|
|
|
|
MOVD R4, savedsp-16(SP)
|
|
|
|
|
MOVD R1, (g_sched+gobuf_sp)(R3)
|
|
|
|
|
|
|
|
|
|
// Switch to m->curg stack and call runtime.cgocallbackg.
|
|
|
|
|
// Because we are taking over the execution of m->curg
|
|
|
|
|
// but *not* resuming what had been running, we need to
|
|
|
|
|
// save that information (m->curg->sched) so we can restore it.
|
|
|
|
|
// We can restore m->curg->sched.sp easily, because calling
|
|
|
|
|
// runtime.cgocallbackg leaves SP unchanged upon return.
|
|
|
|
|
// To save m->curg->sched.pc, we push it onto the stack.
|
|
|
|
|
// This has the added benefit that it looks to the traceback
|
|
|
|
|
// routine like cgocallbackg is going to return to that
|
|
|
|
|
// PC (because the frame we allocate below has the same
|
|
|
|
|
// size as cgocallback_gofunc's frame declared above)
|
|
|
|
|
// so that the traceback will seamlessly trace back into
|
|
|
|
|
// the earlier calls.
|
|
|
|
|
//
|
|
|
|
|
// In the new goroutine, -16(SP) and -8(SP) are unused.
|
|
|
|
|
MOVD m_curg(R8), g
|
|
|
|
|
BL runtime·save_g(SB)
|
|
|
|
|
MOVD (g_sched+gobuf_sp)(g), R4 // prepare stack as R4
|
|
|
|
|
MOVD (g_sched+gobuf_pc)(g), R5
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD R5, -(FIXED_FRAME+16)(R4)
|
|
|
|
|
MOVD $-(FIXED_FRAME+16)(R4), R1
|
2014-12-16 16:34:55 -07:00
|
|
|
BL runtime·cgocallbackg(SB)
|
|
|
|
|
|
|
|
|
|
// Restore g->sched (== m->curg->sched) from saved values.
|
|
|
|
|
MOVD 0(R1), R5
|
|
|
|
|
MOVD R5, (g_sched+gobuf_pc)(g)
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD $(FIXED_FRAME+16)(R1), R4
|
2014-12-16 16:34:55 -07:00
|
|
|
MOVD R4, (g_sched+gobuf_sp)(g)
|
|
|
|
|
|
|
|
|
|
// Switch back to m->g0's stack and restore m->g0->sched.sp.
|
|
|
|
|
// (Unlike m->curg, the g0 goroutine never uses sched.pc,
|
|
|
|
|
// so we do not have to restore it.)
|
|
|
|
|
MOVD g_m(g), R8
|
|
|
|
|
MOVD m_g0(R8), g
|
|
|
|
|
BL runtime·save_g(SB)
|
|
|
|
|
MOVD (g_sched+gobuf_sp)(g), R1
|
|
|
|
|
MOVD savedsp-16(SP), R4
|
|
|
|
|
MOVD R4, (g_sched+gobuf_sp)(g)
|
|
|
|
|
|
|
|
|
|
// If the m on entry was nil, we called needm above to borrow an m
|
|
|
|
|
// for the duration of the call. Since the call is over, return it with dropm.
|
|
|
|
|
MOVD savedm-8(SP), R6
|
|
|
|
|
CMP R6, $0
|
|
|
|
|
BNE droppedm
|
2015-10-08 03:49:39 -06:00
|
|
|
MOVD $runtime·dropm(SB), R12
|
|
|
|
|
MOVD R12, CTR
|
2014-12-16 16:34:55 -07:00
|
|
|
BL (CTR)
|
|
|
|
|
droppedm:
|
|
|
|
|
|
|
|
|
|
// Done!
|
|
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
// void setg(G*); set g. for use by needm.
|
2014-10-27 15:27:03 -06:00
|
|
|
TEXT runtime·setg(SB), NOSPLIT, $0-8
|
2014-12-16 16:34:55 -07:00
|
|
|
MOVD gg+0(FP), g
|
|
|
|
|
// This only happens if iscgo, so jump straight to save_g
|
|
|
|
|
BL runtime·save_g(SB)
|
|
|
|
|
RET
|
|
|
|
|
|
|
|
|
|
// void setg_gcc(G*); set g in C TLS.
|
|
|
|
|
// Must obey the gcc calling convention.
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT setg_gcc<>(SB),NOSPLIT|NOFRAME,$0-0
|
2014-12-16 16:34:55 -07:00
|
|
|
// The standard prologue clobbers R31, which is callee-save in
|
|
|
|
|
// the C ABI, so we have to use $-8-0 and save LR ourselves.
|
|
|
|
|
MOVD LR, R4
|
|
|
|
|
// Also save g and R31, since they're callee-save in C ABI
|
|
|
|
|
MOVD R31, R5
|
|
|
|
|
MOVD g, R6
|
|
|
|
|
|
|
|
|
|
MOVD R3, g
|
|
|
|
|
BL runtime·save_g(SB)
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2014-12-16 16:34:55 -07:00
|
|
|
MOVD R6, g
|
|
|
|
|
MOVD R5, R31
|
|
|
|
|
MOVD R4, LR
|
|
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
runtime: implement GC stack barriers
This commit implements stack barriers to minimize the amount of
stack re-scanning that must be done during mark termination.
Currently the GC scans stacks of active goroutines twice during every
GC cycle: once at the beginning during root discovery and once at the
end during mark termination. The second scan happens while the world
is stopped and guarantees that we've seen all of the roots (since
there are no write barriers on writes to local stack
variables). However, this means pause time is proportional to stack
size. In particularly recursive programs, this can drive pause time up
past our 10ms goal (e.g., it takes about 150ms to scan a 50MB heap).
Re-scanning the entire stack is rarely necessary, especially for large
stacks, because usually most of the frames on the stack were not
active between the first and second scans and hence any changes to
these frames (via non-escaping pointers passed down the stack) were
tracked by write barriers.
To efficiently track how far a stack has been unwound since the first
scan (and, hence, how much needs to be re-scanned), this commit
introduces stack barriers. During the first scan, at exponentially
spaced points in each stack, the scan overwrites return PCs with the
PC of the stack barrier function. When "returned" to, the stack
barrier function records how far the stack has unwound and jumps to
the original return PC for that point in the stack. Then the second
scan only needs to proceed as far as the lowest barrier that hasn't
been hit.
For deeply recursive programs, this substantially reduces mark
termination time (and hence pause time). For the goscheme example
linked in issue #10898, prior to this change, mark termination times
were typically between 100 and 500ms; with this change, mark
termination times are typically between 10 and 20ms. As a result of
the reduced stack scanning work, this reduces overall execution time
of the goscheme example by 20%.
Fixes #10898.
The effect of this on programs that are not deeply recursive is
minimal:
name old time/op new time/op delta
BinaryTree17 3.16s ± 2% 3.26s ± 1% +3.31% (p=0.000 n=19+19)
Fannkuch11 2.42s ± 1% 2.48s ± 1% +2.24% (p=0.000 n=17+19)
FmtFprintfEmpty 50.0ns ± 3% 49.8ns ± 1% ~ (p=0.534 n=20+19)
FmtFprintfString 173ns ± 0% 175ns ± 0% +1.49% (p=0.000 n=16+19)
FmtFprintfInt 170ns ± 1% 175ns ± 1% +2.97% (p=0.000 n=20+19)
FmtFprintfIntInt 288ns ± 0% 295ns ± 0% +2.73% (p=0.000 n=16+19)
FmtFprintfPrefixedInt 242ns ± 1% 252ns ± 1% +4.13% (p=0.000 n=18+18)
FmtFprintfFloat 324ns ± 0% 323ns ± 0% -0.36% (p=0.000 n=20+19)
FmtManyArgs 1.14µs ± 0% 1.12µs ± 1% -1.01% (p=0.000 n=18+19)
GobDecode 8.88ms ± 1% 8.87ms ± 0% ~ (p=0.480 n=19+18)
GobEncode 6.80ms ± 1% 6.85ms ± 0% +0.82% (p=0.000 n=20+18)
Gzip 363ms ± 1% 363ms ± 1% ~ (p=0.077 n=18+20)
Gunzip 90.6ms ± 0% 90.0ms ± 1% -0.71% (p=0.000 n=17+18)
HTTPClientServer 51.5µs ± 1% 50.8µs ± 1% -1.32% (p=0.000 n=18+18)
JSONEncode 17.0ms ± 0% 17.1ms ± 0% +0.40% (p=0.000 n=18+17)
JSONDecode 61.8ms ± 0% 63.8ms ± 1% +3.11% (p=0.000 n=18+17)
Mandelbrot200 3.84ms ± 0% 3.84ms ± 1% ~ (p=0.583 n=19+19)
GoParse 3.71ms ± 1% 3.72ms ± 1% ~ (p=0.159 n=18+19)
RegexpMatchEasy0_32 100ns ± 0% 100ns ± 1% -0.19% (p=0.033 n=17+19)
RegexpMatchEasy0_1K 342ns ± 1% 331ns ± 0% -3.41% (p=0.000 n=19+19)
RegexpMatchEasy1_32 82.5ns ± 0% 81.7ns ± 0% -0.98% (p=0.000 n=18+18)
RegexpMatchEasy1_1K 505ns ± 0% 494ns ± 1% -2.16% (p=0.000 n=18+18)
RegexpMatchMedium_32 137ns ± 1% 137ns ± 1% -0.24% (p=0.048 n=20+18)
RegexpMatchMedium_1K 41.6µs ± 0% 41.3µs ± 1% -0.57% (p=0.004 n=18+20)
RegexpMatchHard_32 2.11µs ± 0% 2.11µs ± 1% +0.20% (p=0.037 n=17+19)
RegexpMatchHard_1K 63.9µs ± 2% 63.3µs ± 0% -0.99% (p=0.000 n=20+17)
Revcomp 560ms ± 1% 522ms ± 0% -6.87% (p=0.000 n=18+16)
Template 75.0ms ± 0% 75.1ms ± 1% +0.18% (p=0.013 n=18+19)
TimeParse 358ns ± 1% 364ns ± 0% +1.74% (p=0.000 n=20+15)
TimeFormat 360ns ± 0% 372ns ± 0% +3.55% (p=0.000 n=20+18)
Change-Id: If8a9bfae6c128d15a4f405e02bcfa50129df82a2
Reviewed-on: https://go-review.googlesource.com/10314
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2015-05-20 14:30:49 -06:00
|
|
|
TEXT runtime·getcallerpc(SB),NOSPLIT,$8-16
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD FIXED_FRAME+8(R1), R3 // LR saved by caller
|
runtime: implement GC stack barriers
This commit implements stack barriers to minimize the amount of
stack re-scanning that must be done during mark termination.
Currently the GC scans stacks of active goroutines twice during every
GC cycle: once at the beginning during root discovery and once at the
end during mark termination. The second scan happens while the world
is stopped and guarantees that we've seen all of the roots (since
there are no write barriers on writes to local stack
variables). However, this means pause time is proportional to stack
size. In particularly recursive programs, this can drive pause time up
past our 10ms goal (e.g., it takes about 150ms to scan a 50MB heap).
Re-scanning the entire stack is rarely necessary, especially for large
stacks, because usually most of the frames on the stack were not
active between the first and second scans and hence any changes to
these frames (via non-escaping pointers passed down the stack) were
tracked by write barriers.
To efficiently track how far a stack has been unwound since the first
scan (and, hence, how much needs to be re-scanned), this commit
introduces stack barriers. During the first scan, at exponentially
spaced points in each stack, the scan overwrites return PCs with the
PC of the stack barrier function. When "returned" to, the stack
barrier function records how far the stack has unwound and jumps to
the original return PC for that point in the stack. Then the second
scan only needs to proceed as far as the lowest barrier that hasn't
been hit.
For deeply recursive programs, this substantially reduces mark
termination time (and hence pause time). For the goscheme example
linked in issue #10898, prior to this change, mark termination times
were typically between 100 and 500ms; with this change, mark
termination times are typically between 10 and 20ms. As a result of
the reduced stack scanning work, this reduces overall execution time
of the goscheme example by 20%.
Fixes #10898.
The effect of this on programs that are not deeply recursive is
minimal:
name old time/op new time/op delta
BinaryTree17 3.16s ± 2% 3.26s ± 1% +3.31% (p=0.000 n=19+19)
Fannkuch11 2.42s ± 1% 2.48s ± 1% +2.24% (p=0.000 n=17+19)
FmtFprintfEmpty 50.0ns ± 3% 49.8ns ± 1% ~ (p=0.534 n=20+19)
FmtFprintfString 173ns ± 0% 175ns ± 0% +1.49% (p=0.000 n=16+19)
FmtFprintfInt 170ns ± 1% 175ns ± 1% +2.97% (p=0.000 n=20+19)
FmtFprintfIntInt 288ns ± 0% 295ns ± 0% +2.73% (p=0.000 n=16+19)
FmtFprintfPrefixedInt 242ns ± 1% 252ns ± 1% +4.13% (p=0.000 n=18+18)
FmtFprintfFloat 324ns ± 0% 323ns ± 0% -0.36% (p=0.000 n=20+19)
FmtManyArgs 1.14µs ± 0% 1.12µs ± 1% -1.01% (p=0.000 n=18+19)
GobDecode 8.88ms ± 1% 8.87ms ± 0% ~ (p=0.480 n=19+18)
GobEncode 6.80ms ± 1% 6.85ms ± 0% +0.82% (p=0.000 n=20+18)
Gzip 363ms ± 1% 363ms ± 1% ~ (p=0.077 n=18+20)
Gunzip 90.6ms ± 0% 90.0ms ± 1% -0.71% (p=0.000 n=17+18)
HTTPClientServer 51.5µs ± 1% 50.8µs ± 1% -1.32% (p=0.000 n=18+18)
JSONEncode 17.0ms ± 0% 17.1ms ± 0% +0.40% (p=0.000 n=18+17)
JSONDecode 61.8ms ± 0% 63.8ms ± 1% +3.11% (p=0.000 n=18+17)
Mandelbrot200 3.84ms ± 0% 3.84ms ± 1% ~ (p=0.583 n=19+19)
GoParse 3.71ms ± 1% 3.72ms ± 1% ~ (p=0.159 n=18+19)
RegexpMatchEasy0_32 100ns ± 0% 100ns ± 1% -0.19% (p=0.033 n=17+19)
RegexpMatchEasy0_1K 342ns ± 1% 331ns ± 0% -3.41% (p=0.000 n=19+19)
RegexpMatchEasy1_32 82.5ns ± 0% 81.7ns ± 0% -0.98% (p=0.000 n=18+18)
RegexpMatchEasy1_1K 505ns ± 0% 494ns ± 1% -2.16% (p=0.000 n=18+18)
RegexpMatchMedium_32 137ns ± 1% 137ns ± 1% -0.24% (p=0.048 n=20+18)
RegexpMatchMedium_1K 41.6µs ± 0% 41.3µs ± 1% -0.57% (p=0.004 n=18+20)
RegexpMatchHard_32 2.11µs ± 0% 2.11µs ± 1% +0.20% (p=0.037 n=17+19)
RegexpMatchHard_1K 63.9µs ± 2% 63.3µs ± 0% -0.99% (p=0.000 n=20+17)
Revcomp 560ms ± 1% 522ms ± 0% -6.87% (p=0.000 n=18+16)
Template 75.0ms ± 0% 75.1ms ± 1% +0.18% (p=0.013 n=18+19)
TimeParse 358ns ± 1% 364ns ± 0% +1.74% (p=0.000 n=20+15)
TimeFormat 360ns ± 0% 372ns ± 0% +3.55% (p=0.000 n=20+18)
Change-Id: If8a9bfae6c128d15a4f405e02bcfa50129df82a2
Reviewed-on: https://go-review.googlesource.com/10314
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2015-05-20 14:30:49 -06:00
|
|
|
MOVD runtime·stackBarrierPC(SB), R4
|
|
|
|
|
CMP R3, R4
|
|
|
|
|
BNE nobar
|
|
|
|
|
// Get original return PC.
|
|
|
|
|
BL runtime·nextBarrierPC(SB)
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD FIXED_FRAME+0(R1), R3
|
runtime: implement GC stack barriers
This commit implements stack barriers to minimize the amount of
stack re-scanning that must be done during mark termination.
Currently the GC scans stacks of active goroutines twice during every
GC cycle: once at the beginning during root discovery and once at the
end during mark termination. The second scan happens while the world
is stopped and guarantees that we've seen all of the roots (since
there are no write barriers on writes to local stack
variables). However, this means pause time is proportional to stack
size. In particularly recursive programs, this can drive pause time up
past our 10ms goal (e.g., it takes about 150ms to scan a 50MB heap).
Re-scanning the entire stack is rarely necessary, especially for large
stacks, because usually most of the frames on the stack were not
active between the first and second scans and hence any changes to
these frames (via non-escaping pointers passed down the stack) were
tracked by write barriers.
To efficiently track how far a stack has been unwound since the first
scan (and, hence, how much needs to be re-scanned), this commit
introduces stack barriers. During the first scan, at exponentially
spaced points in each stack, the scan overwrites return PCs with the
PC of the stack barrier function. When "returned" to, the stack
barrier function records how far the stack has unwound and jumps to
the original return PC for that point in the stack. Then the second
scan only needs to proceed as far as the lowest barrier that hasn't
been hit.
For deeply recursive programs, this substantially reduces mark
termination time (and hence pause time). For the goscheme example
linked in issue #10898, prior to this change, mark termination times
were typically between 100 and 500ms; with this change, mark
termination times are typically between 10 and 20ms. As a result of
the reduced stack scanning work, this reduces overall execution time
of the goscheme example by 20%.
Fixes #10898.
The effect of this on programs that are not deeply recursive is
minimal:
name old time/op new time/op delta
BinaryTree17 3.16s ± 2% 3.26s ± 1% +3.31% (p=0.000 n=19+19)
Fannkuch11 2.42s ± 1% 2.48s ± 1% +2.24% (p=0.000 n=17+19)
FmtFprintfEmpty 50.0ns ± 3% 49.8ns ± 1% ~ (p=0.534 n=20+19)
FmtFprintfString 173ns ± 0% 175ns ± 0% +1.49% (p=0.000 n=16+19)
FmtFprintfInt 170ns ± 1% 175ns ± 1% +2.97% (p=0.000 n=20+19)
FmtFprintfIntInt 288ns ± 0% 295ns ± 0% +2.73% (p=0.000 n=16+19)
FmtFprintfPrefixedInt 242ns ± 1% 252ns ± 1% +4.13% (p=0.000 n=18+18)
FmtFprintfFloat 324ns ± 0% 323ns ± 0% -0.36% (p=0.000 n=20+19)
FmtManyArgs 1.14µs ± 0% 1.12µs ± 1% -1.01% (p=0.000 n=18+19)
GobDecode 8.88ms ± 1% 8.87ms ± 0% ~ (p=0.480 n=19+18)
GobEncode 6.80ms ± 1% 6.85ms ± 0% +0.82% (p=0.000 n=20+18)
Gzip 363ms ± 1% 363ms ± 1% ~ (p=0.077 n=18+20)
Gunzip 90.6ms ± 0% 90.0ms ± 1% -0.71% (p=0.000 n=17+18)
HTTPClientServer 51.5µs ± 1% 50.8µs ± 1% -1.32% (p=0.000 n=18+18)
JSONEncode 17.0ms ± 0% 17.1ms ± 0% +0.40% (p=0.000 n=18+17)
JSONDecode 61.8ms ± 0% 63.8ms ± 1% +3.11% (p=0.000 n=18+17)
Mandelbrot200 3.84ms ± 0% 3.84ms ± 1% ~ (p=0.583 n=19+19)
GoParse 3.71ms ± 1% 3.72ms ± 1% ~ (p=0.159 n=18+19)
RegexpMatchEasy0_32 100ns ± 0% 100ns ± 1% -0.19% (p=0.033 n=17+19)
RegexpMatchEasy0_1K 342ns ± 1% 331ns ± 0% -3.41% (p=0.000 n=19+19)
RegexpMatchEasy1_32 82.5ns ± 0% 81.7ns ± 0% -0.98% (p=0.000 n=18+18)
RegexpMatchEasy1_1K 505ns ± 0% 494ns ± 1% -2.16% (p=0.000 n=18+18)
RegexpMatchMedium_32 137ns ± 1% 137ns ± 1% -0.24% (p=0.048 n=20+18)
RegexpMatchMedium_1K 41.6µs ± 0% 41.3µs ± 1% -0.57% (p=0.004 n=18+20)
RegexpMatchHard_32 2.11µs ± 0% 2.11µs ± 1% +0.20% (p=0.037 n=17+19)
RegexpMatchHard_1K 63.9µs ± 2% 63.3µs ± 0% -0.99% (p=0.000 n=20+17)
Revcomp 560ms ± 1% 522ms ± 0% -6.87% (p=0.000 n=18+16)
Template 75.0ms ± 0% 75.1ms ± 1% +0.18% (p=0.013 n=18+19)
TimeParse 358ns ± 1% 364ns ± 0% +1.74% (p=0.000 n=20+15)
TimeFormat 360ns ± 0% 372ns ± 0% +3.55% (p=0.000 n=20+18)
Change-Id: If8a9bfae6c128d15a4f405e02bcfa50129df82a2
Reviewed-on: https://go-review.googlesource.com/10314
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2015-05-20 14:30:49 -06:00
|
|
|
nobar:
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVD R3, ret+8(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
runtime: implement GC stack barriers
This commit implements stack barriers to minimize the amount of
stack re-scanning that must be done during mark termination.
Currently the GC scans stacks of active goroutines twice during every
GC cycle: once at the beginning during root discovery and once at the
end during mark termination. The second scan happens while the world
is stopped and guarantees that we've seen all of the roots (since
there are no write barriers on writes to local stack
variables). However, this means pause time is proportional to stack
size. In particularly recursive programs, this can drive pause time up
past our 10ms goal (e.g., it takes about 150ms to scan a 50MB heap).
Re-scanning the entire stack is rarely necessary, especially for large
stacks, because usually most of the frames on the stack were not
active between the first and second scans and hence any changes to
these frames (via non-escaping pointers passed down the stack) were
tracked by write barriers.
To efficiently track how far a stack has been unwound since the first
scan (and, hence, how much needs to be re-scanned), this commit
introduces stack barriers. During the first scan, at exponentially
spaced points in each stack, the scan overwrites return PCs with the
PC of the stack barrier function. When "returned" to, the stack
barrier function records how far the stack has unwound and jumps to
the original return PC for that point in the stack. Then the second
scan only needs to proceed as far as the lowest barrier that hasn't
been hit.
For deeply recursive programs, this substantially reduces mark
termination time (and hence pause time). For the goscheme example
linked in issue #10898, prior to this change, mark termination times
were typically between 100 and 500ms; with this change, mark
termination times are typically between 10 and 20ms. As a result of
the reduced stack scanning work, this reduces overall execution time
of the goscheme example by 20%.
Fixes #10898.
The effect of this on programs that are not deeply recursive is
minimal:
name old time/op new time/op delta
BinaryTree17 3.16s ± 2% 3.26s ± 1% +3.31% (p=0.000 n=19+19)
Fannkuch11 2.42s ± 1% 2.48s ± 1% +2.24% (p=0.000 n=17+19)
FmtFprintfEmpty 50.0ns ± 3% 49.8ns ± 1% ~ (p=0.534 n=20+19)
FmtFprintfString 173ns ± 0% 175ns ± 0% +1.49% (p=0.000 n=16+19)
FmtFprintfInt 170ns ± 1% 175ns ± 1% +2.97% (p=0.000 n=20+19)
FmtFprintfIntInt 288ns ± 0% 295ns ± 0% +2.73% (p=0.000 n=16+19)
FmtFprintfPrefixedInt 242ns ± 1% 252ns ± 1% +4.13% (p=0.000 n=18+18)
FmtFprintfFloat 324ns ± 0% 323ns ± 0% -0.36% (p=0.000 n=20+19)
FmtManyArgs 1.14µs ± 0% 1.12µs ± 1% -1.01% (p=0.000 n=18+19)
GobDecode 8.88ms ± 1% 8.87ms ± 0% ~ (p=0.480 n=19+18)
GobEncode 6.80ms ± 1% 6.85ms ± 0% +0.82% (p=0.000 n=20+18)
Gzip 363ms ± 1% 363ms ± 1% ~ (p=0.077 n=18+20)
Gunzip 90.6ms ± 0% 90.0ms ± 1% -0.71% (p=0.000 n=17+18)
HTTPClientServer 51.5µs ± 1% 50.8µs ± 1% -1.32% (p=0.000 n=18+18)
JSONEncode 17.0ms ± 0% 17.1ms ± 0% +0.40% (p=0.000 n=18+17)
JSONDecode 61.8ms ± 0% 63.8ms ± 1% +3.11% (p=0.000 n=18+17)
Mandelbrot200 3.84ms ± 0% 3.84ms ± 1% ~ (p=0.583 n=19+19)
GoParse 3.71ms ± 1% 3.72ms ± 1% ~ (p=0.159 n=18+19)
RegexpMatchEasy0_32 100ns ± 0% 100ns ± 1% -0.19% (p=0.033 n=17+19)
RegexpMatchEasy0_1K 342ns ± 1% 331ns ± 0% -3.41% (p=0.000 n=19+19)
RegexpMatchEasy1_32 82.5ns ± 0% 81.7ns ± 0% -0.98% (p=0.000 n=18+18)
RegexpMatchEasy1_1K 505ns ± 0% 494ns ± 1% -2.16% (p=0.000 n=18+18)
RegexpMatchMedium_32 137ns ± 1% 137ns ± 1% -0.24% (p=0.048 n=20+18)
RegexpMatchMedium_1K 41.6µs ± 0% 41.3µs ± 1% -0.57% (p=0.004 n=18+20)
RegexpMatchHard_32 2.11µs ± 0% 2.11µs ± 1% +0.20% (p=0.037 n=17+19)
RegexpMatchHard_1K 63.9µs ± 2% 63.3µs ± 0% -0.99% (p=0.000 n=20+17)
Revcomp 560ms ± 1% 522ms ± 0% -6.87% (p=0.000 n=18+16)
Template 75.0ms ± 0% 75.1ms ± 1% +0.18% (p=0.013 n=18+19)
TimeParse 358ns ± 1% 364ns ± 0% +1.74% (p=0.000 n=20+15)
TimeFormat 360ns ± 0% 372ns ± 0% +3.55% (p=0.000 n=20+18)
Change-Id: If8a9bfae6c128d15a4f405e02bcfa50129df82a2
Reviewed-on: https://go-review.googlesource.com/10314
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2015-05-20 14:30:49 -06:00
|
|
|
TEXT runtime·setcallerpc(SB),NOSPLIT,$8-16
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVD pc+8(FP), R3
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD FIXED_FRAME+8(R1), R4
|
runtime: implement GC stack barriers
This commit implements stack barriers to minimize the amount of
stack re-scanning that must be done during mark termination.
Currently the GC scans stacks of active goroutines twice during every
GC cycle: once at the beginning during root discovery and once at the
end during mark termination. The second scan happens while the world
is stopped and guarantees that we've seen all of the roots (since
there are no write barriers on writes to local stack
variables). However, this means pause time is proportional to stack
size. In particularly recursive programs, this can drive pause time up
past our 10ms goal (e.g., it takes about 150ms to scan a 50MB heap).
Re-scanning the entire stack is rarely necessary, especially for large
stacks, because usually most of the frames on the stack were not
active between the first and second scans and hence any changes to
these frames (via non-escaping pointers passed down the stack) were
tracked by write barriers.
To efficiently track how far a stack has been unwound since the first
scan (and, hence, how much needs to be re-scanned), this commit
introduces stack barriers. During the first scan, at exponentially
spaced points in each stack, the scan overwrites return PCs with the
PC of the stack barrier function. When "returned" to, the stack
barrier function records how far the stack has unwound and jumps to
the original return PC for that point in the stack. Then the second
scan only needs to proceed as far as the lowest barrier that hasn't
been hit.
For deeply recursive programs, this substantially reduces mark
termination time (and hence pause time). For the goscheme example
linked in issue #10898, prior to this change, mark termination times
were typically between 100 and 500ms; with this change, mark
termination times are typically between 10 and 20ms. As a result of
the reduced stack scanning work, this reduces overall execution time
of the goscheme example by 20%.
Fixes #10898.
The effect of this on programs that are not deeply recursive is
minimal:
name old time/op new time/op delta
BinaryTree17 3.16s ± 2% 3.26s ± 1% +3.31% (p=0.000 n=19+19)
Fannkuch11 2.42s ± 1% 2.48s ± 1% +2.24% (p=0.000 n=17+19)
FmtFprintfEmpty 50.0ns ± 3% 49.8ns ± 1% ~ (p=0.534 n=20+19)
FmtFprintfString 173ns ± 0% 175ns ± 0% +1.49% (p=0.000 n=16+19)
FmtFprintfInt 170ns ± 1% 175ns ± 1% +2.97% (p=0.000 n=20+19)
FmtFprintfIntInt 288ns ± 0% 295ns ± 0% +2.73% (p=0.000 n=16+19)
FmtFprintfPrefixedInt 242ns ± 1% 252ns ± 1% +4.13% (p=0.000 n=18+18)
FmtFprintfFloat 324ns ± 0% 323ns ± 0% -0.36% (p=0.000 n=20+19)
FmtManyArgs 1.14µs ± 0% 1.12µs ± 1% -1.01% (p=0.000 n=18+19)
GobDecode 8.88ms ± 1% 8.87ms ± 0% ~ (p=0.480 n=19+18)
GobEncode 6.80ms ± 1% 6.85ms ± 0% +0.82% (p=0.000 n=20+18)
Gzip 363ms ± 1% 363ms ± 1% ~ (p=0.077 n=18+20)
Gunzip 90.6ms ± 0% 90.0ms ± 1% -0.71% (p=0.000 n=17+18)
HTTPClientServer 51.5µs ± 1% 50.8µs ± 1% -1.32% (p=0.000 n=18+18)
JSONEncode 17.0ms ± 0% 17.1ms ± 0% +0.40% (p=0.000 n=18+17)
JSONDecode 61.8ms ± 0% 63.8ms ± 1% +3.11% (p=0.000 n=18+17)
Mandelbrot200 3.84ms ± 0% 3.84ms ± 1% ~ (p=0.583 n=19+19)
GoParse 3.71ms ± 1% 3.72ms ± 1% ~ (p=0.159 n=18+19)
RegexpMatchEasy0_32 100ns ± 0% 100ns ± 1% -0.19% (p=0.033 n=17+19)
RegexpMatchEasy0_1K 342ns ± 1% 331ns ± 0% -3.41% (p=0.000 n=19+19)
RegexpMatchEasy1_32 82.5ns ± 0% 81.7ns ± 0% -0.98% (p=0.000 n=18+18)
RegexpMatchEasy1_1K 505ns ± 0% 494ns ± 1% -2.16% (p=0.000 n=18+18)
RegexpMatchMedium_32 137ns ± 1% 137ns ± 1% -0.24% (p=0.048 n=20+18)
RegexpMatchMedium_1K 41.6µs ± 0% 41.3µs ± 1% -0.57% (p=0.004 n=18+20)
RegexpMatchHard_32 2.11µs ± 0% 2.11µs ± 1% +0.20% (p=0.037 n=17+19)
RegexpMatchHard_1K 63.9µs ± 2% 63.3µs ± 0% -0.99% (p=0.000 n=20+17)
Revcomp 560ms ± 1% 522ms ± 0% -6.87% (p=0.000 n=18+16)
Template 75.0ms ± 0% 75.1ms ± 1% +0.18% (p=0.013 n=18+19)
TimeParse 358ns ± 1% 364ns ± 0% +1.74% (p=0.000 n=20+15)
TimeFormat 360ns ± 0% 372ns ± 0% +3.55% (p=0.000 n=20+18)
Change-Id: If8a9bfae6c128d15a4f405e02bcfa50129df82a2
Reviewed-on: https://go-review.googlesource.com/10314
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2015-05-20 14:30:49 -06:00
|
|
|
MOVD runtime·stackBarrierPC(SB), R5
|
|
|
|
|
CMP R4, R5
|
|
|
|
|
BEQ setbar
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD R3, FIXED_FRAME+8(R1) // set LR in caller
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
runtime: implement GC stack barriers
This commit implements stack barriers to minimize the amount of
stack re-scanning that must be done during mark termination.
Currently the GC scans stacks of active goroutines twice during every
GC cycle: once at the beginning during root discovery and once at the
end during mark termination. The second scan happens while the world
is stopped and guarantees that we've seen all of the roots (since
there are no write barriers on writes to local stack
variables). However, this means pause time is proportional to stack
size. In particularly recursive programs, this can drive pause time up
past our 10ms goal (e.g., it takes about 150ms to scan a 50MB heap).
Re-scanning the entire stack is rarely necessary, especially for large
stacks, because usually most of the frames on the stack were not
active between the first and second scans and hence any changes to
these frames (via non-escaping pointers passed down the stack) were
tracked by write barriers.
To efficiently track how far a stack has been unwound since the first
scan (and, hence, how much needs to be re-scanned), this commit
introduces stack barriers. During the first scan, at exponentially
spaced points in each stack, the scan overwrites return PCs with the
PC of the stack barrier function. When "returned" to, the stack
barrier function records how far the stack has unwound and jumps to
the original return PC for that point in the stack. Then the second
scan only needs to proceed as far as the lowest barrier that hasn't
been hit.
For deeply recursive programs, this substantially reduces mark
termination time (and hence pause time). For the goscheme example
linked in issue #10898, prior to this change, mark termination times
were typically between 100 and 500ms; with this change, mark
termination times are typically between 10 and 20ms. As a result of
the reduced stack scanning work, this reduces overall execution time
of the goscheme example by 20%.
Fixes #10898.
The effect of this on programs that are not deeply recursive is
minimal:
name old time/op new time/op delta
BinaryTree17 3.16s ± 2% 3.26s ± 1% +3.31% (p=0.000 n=19+19)
Fannkuch11 2.42s ± 1% 2.48s ± 1% +2.24% (p=0.000 n=17+19)
FmtFprintfEmpty 50.0ns ± 3% 49.8ns ± 1% ~ (p=0.534 n=20+19)
FmtFprintfString 173ns ± 0% 175ns ± 0% +1.49% (p=0.000 n=16+19)
FmtFprintfInt 170ns ± 1% 175ns ± 1% +2.97% (p=0.000 n=20+19)
FmtFprintfIntInt 288ns ± 0% 295ns ± 0% +2.73% (p=0.000 n=16+19)
FmtFprintfPrefixedInt 242ns ± 1% 252ns ± 1% +4.13% (p=0.000 n=18+18)
FmtFprintfFloat 324ns ± 0% 323ns ± 0% -0.36% (p=0.000 n=20+19)
FmtManyArgs 1.14µs ± 0% 1.12µs ± 1% -1.01% (p=0.000 n=18+19)
GobDecode 8.88ms ± 1% 8.87ms ± 0% ~ (p=0.480 n=19+18)
GobEncode 6.80ms ± 1% 6.85ms ± 0% +0.82% (p=0.000 n=20+18)
Gzip 363ms ± 1% 363ms ± 1% ~ (p=0.077 n=18+20)
Gunzip 90.6ms ± 0% 90.0ms ± 1% -0.71% (p=0.000 n=17+18)
HTTPClientServer 51.5µs ± 1% 50.8µs ± 1% -1.32% (p=0.000 n=18+18)
JSONEncode 17.0ms ± 0% 17.1ms ± 0% +0.40% (p=0.000 n=18+17)
JSONDecode 61.8ms ± 0% 63.8ms ± 1% +3.11% (p=0.000 n=18+17)
Mandelbrot200 3.84ms ± 0% 3.84ms ± 1% ~ (p=0.583 n=19+19)
GoParse 3.71ms ± 1% 3.72ms ± 1% ~ (p=0.159 n=18+19)
RegexpMatchEasy0_32 100ns ± 0% 100ns ± 1% -0.19% (p=0.033 n=17+19)
RegexpMatchEasy0_1K 342ns ± 1% 331ns ± 0% -3.41% (p=0.000 n=19+19)
RegexpMatchEasy1_32 82.5ns ± 0% 81.7ns ± 0% -0.98% (p=0.000 n=18+18)
RegexpMatchEasy1_1K 505ns ± 0% 494ns ± 1% -2.16% (p=0.000 n=18+18)
RegexpMatchMedium_32 137ns ± 1% 137ns ± 1% -0.24% (p=0.048 n=20+18)
RegexpMatchMedium_1K 41.6µs ± 0% 41.3µs ± 1% -0.57% (p=0.004 n=18+20)
RegexpMatchHard_32 2.11µs ± 0% 2.11µs ± 1% +0.20% (p=0.037 n=17+19)
RegexpMatchHard_1K 63.9µs ± 2% 63.3µs ± 0% -0.99% (p=0.000 n=20+17)
Revcomp 560ms ± 1% 522ms ± 0% -6.87% (p=0.000 n=18+16)
Template 75.0ms ± 0% 75.1ms ± 1% +0.18% (p=0.013 n=18+19)
TimeParse 358ns ± 1% 364ns ± 0% +1.74% (p=0.000 n=20+15)
TimeFormat 360ns ± 0% 372ns ± 0% +3.55% (p=0.000 n=20+18)
Change-Id: If8a9bfae6c128d15a4f405e02bcfa50129df82a2
Reviewed-on: https://go-review.googlesource.com/10314
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2015-05-20 14:30:49 -06:00
|
|
|
setbar:
|
|
|
|
|
// Set the stack barrier return PC.
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD R3, FIXED_FRAME+0(R1)
|
runtime: implement GC stack barriers
This commit implements stack barriers to minimize the amount of
stack re-scanning that must be done during mark termination.
Currently the GC scans stacks of active goroutines twice during every
GC cycle: once at the beginning during root discovery and once at the
end during mark termination. The second scan happens while the world
is stopped and guarantees that we've seen all of the roots (since
there are no write barriers on writes to local stack
variables). However, this means pause time is proportional to stack
size. In particularly recursive programs, this can drive pause time up
past our 10ms goal (e.g., it takes about 150ms to scan a 50MB heap).
Re-scanning the entire stack is rarely necessary, especially for large
stacks, because usually most of the frames on the stack were not
active between the first and second scans and hence any changes to
these frames (via non-escaping pointers passed down the stack) were
tracked by write barriers.
To efficiently track how far a stack has been unwound since the first
scan (and, hence, how much needs to be re-scanned), this commit
introduces stack barriers. During the first scan, at exponentially
spaced points in each stack, the scan overwrites return PCs with the
PC of the stack barrier function. When "returned" to, the stack
barrier function records how far the stack has unwound and jumps to
the original return PC for that point in the stack. Then the second
scan only needs to proceed as far as the lowest barrier that hasn't
been hit.
For deeply recursive programs, this substantially reduces mark
termination time (and hence pause time). For the goscheme example
linked in issue #10898, prior to this change, mark termination times
were typically between 100 and 500ms; with this change, mark
termination times are typically between 10 and 20ms. As a result of
the reduced stack scanning work, this reduces overall execution time
of the goscheme example by 20%.
Fixes #10898.
The effect of this on programs that are not deeply recursive is
minimal:
name old time/op new time/op delta
BinaryTree17 3.16s ± 2% 3.26s ± 1% +3.31% (p=0.000 n=19+19)
Fannkuch11 2.42s ± 1% 2.48s ± 1% +2.24% (p=0.000 n=17+19)
FmtFprintfEmpty 50.0ns ± 3% 49.8ns ± 1% ~ (p=0.534 n=20+19)
FmtFprintfString 173ns ± 0% 175ns ± 0% +1.49% (p=0.000 n=16+19)
FmtFprintfInt 170ns ± 1% 175ns ± 1% +2.97% (p=0.000 n=20+19)
FmtFprintfIntInt 288ns ± 0% 295ns ± 0% +2.73% (p=0.000 n=16+19)
FmtFprintfPrefixedInt 242ns ± 1% 252ns ± 1% +4.13% (p=0.000 n=18+18)
FmtFprintfFloat 324ns ± 0% 323ns ± 0% -0.36% (p=0.000 n=20+19)
FmtManyArgs 1.14µs ± 0% 1.12µs ± 1% -1.01% (p=0.000 n=18+19)
GobDecode 8.88ms ± 1% 8.87ms ± 0% ~ (p=0.480 n=19+18)
GobEncode 6.80ms ± 1% 6.85ms ± 0% +0.82% (p=0.000 n=20+18)
Gzip 363ms ± 1% 363ms ± 1% ~ (p=0.077 n=18+20)
Gunzip 90.6ms ± 0% 90.0ms ± 1% -0.71% (p=0.000 n=17+18)
HTTPClientServer 51.5µs ± 1% 50.8µs ± 1% -1.32% (p=0.000 n=18+18)
JSONEncode 17.0ms ± 0% 17.1ms ± 0% +0.40% (p=0.000 n=18+17)
JSONDecode 61.8ms ± 0% 63.8ms ± 1% +3.11% (p=0.000 n=18+17)
Mandelbrot200 3.84ms ± 0% 3.84ms ± 1% ~ (p=0.583 n=19+19)
GoParse 3.71ms ± 1% 3.72ms ± 1% ~ (p=0.159 n=18+19)
RegexpMatchEasy0_32 100ns ± 0% 100ns ± 1% -0.19% (p=0.033 n=17+19)
RegexpMatchEasy0_1K 342ns ± 1% 331ns ± 0% -3.41% (p=0.000 n=19+19)
RegexpMatchEasy1_32 82.5ns ± 0% 81.7ns ± 0% -0.98% (p=0.000 n=18+18)
RegexpMatchEasy1_1K 505ns ± 0% 494ns ± 1% -2.16% (p=0.000 n=18+18)
RegexpMatchMedium_32 137ns ± 1% 137ns ± 1% -0.24% (p=0.048 n=20+18)
RegexpMatchMedium_1K 41.6µs ± 0% 41.3µs ± 1% -0.57% (p=0.004 n=18+20)
RegexpMatchHard_32 2.11µs ± 0% 2.11µs ± 1% +0.20% (p=0.037 n=17+19)
RegexpMatchHard_1K 63.9µs ± 2% 63.3µs ± 0% -0.99% (p=0.000 n=20+17)
Revcomp 560ms ± 1% 522ms ± 0% -6.87% (p=0.000 n=18+16)
Template 75.0ms ± 0% 75.1ms ± 1% +0.18% (p=0.013 n=18+19)
TimeParse 358ns ± 1% 364ns ± 0% +1.74% (p=0.000 n=20+15)
TimeFormat 360ns ± 0% 372ns ± 0% +3.55% (p=0.000 n=20+18)
Change-Id: If8a9bfae6c128d15a4f405e02bcfa50129df82a2
Reviewed-on: https://go-review.googlesource.com/10314
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2015-05-20 14:30:49 -06:00
|
|
|
BL runtime·setNextBarrierPC(SB)
|
|
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2014-10-27 15:27:03 -06:00
|
|
|
TEXT runtime·getcallersp(SB),NOSPLIT,$0-16
|
2014-10-30 13:58:30 -06:00
|
|
|
MOVD argp+0(FP), R3
|
2015-10-08 03:34:29 -06:00
|
|
|
SUB $FIXED_FRAME, R3
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVD R3, ret+8(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·abort(SB),NOSPLIT|NOFRAME,$0-0
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVW (R0), R0
|
|
|
|
|
UNDEF
|
|
|
|
|
|
|
|
|
|
#define TBRL 268
|
|
|
|
|
#define TBRU 269 /* Time base Upper/Lower */
|
|
|
|
|
|
|
|
|
|
// int64 runtime·cputicks(void)
|
2014-10-27 15:27:03 -06:00
|
|
|
TEXT runtime·cputicks(SB),NOSPLIT,$0-8
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVW SPR(TBRU), R4
|
|
|
|
|
MOVW SPR(TBRL), R3
|
|
|
|
|
MOVW SPR(TBRU), R5
|
|
|
|
|
CMPW R4, R5
|
|
|
|
|
BNE -4(PC)
|
|
|
|
|
SLD $32, R5
|
|
|
|
|
OR R5, R3
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVD R3, ret+0(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2015-01-06 17:42:48 -07:00
|
|
|
// memhash_varlen(p unsafe.Pointer, h seed) uintptr
|
|
|
|
|
// redirects to memhash(p, h, size) using the size
|
|
|
|
|
// stored in the closure.
|
|
|
|
|
TEXT runtime·memhash_varlen(SB),NOSPLIT,$40-24
|
|
|
|
|
GO_ARGS
|
|
|
|
|
NO_LOCAL_POINTERS
|
|
|
|
|
MOVD p+0(FP), R3
|
|
|
|
|
MOVD h+8(FP), R4
|
|
|
|
|
MOVD 8(R11), R5
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD R3, FIXED_FRAME+0(R1)
|
|
|
|
|
MOVD R4, FIXED_FRAME+8(R1)
|
|
|
|
|
MOVD R5, FIXED_FRAME+16(R1)
|
2015-01-06 17:42:48 -07:00
|
|
|
BL runtime·memhash(SB)
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD FIXED_FRAME+24(R1), R3
|
2015-01-06 17:42:48 -07:00
|
|
|
MOVD R3, ret+16(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2015-01-06 17:42:48 -07:00
|
|
|
|
2014-12-05 17:13:20 -07:00
|
|
|
// AES hashing not implemented for ppc64
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·aeshash(SB),NOSPLIT|NOFRAME,$0-0
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVW (R0), R1
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·aeshash32(SB),NOSPLIT|NOFRAME,$0-0
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVW (R0), R1
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·aeshash64(SB),NOSPLIT|NOFRAME,$0-0
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVW (R0), R1
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·aeshashstr(SB),NOSPLIT|NOFRAME,$0-0
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVW (R0), R1
|
|
|
|
|
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·memeq(SB),NOSPLIT|NOFRAME,$0-25
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD a+0(FP), R3
|
|
|
|
|
MOVD b+8(FP), R4
|
2014-10-30 13:58:30 -06:00
|
|
|
MOVD size+16(FP), R5
|
2014-08-12 17:48:49 -06:00
|
|
|
SUB $1, R3
|
|
|
|
|
SUB $1, R4
|
|
|
|
|
ADD R3, R5, R8
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
loop:
|
2014-08-12 17:48:49 -06:00
|
|
|
CMP R3, R8
|
2014-10-30 13:58:30 -06:00
|
|
|
BNE test
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD $1, R3
|
|
|
|
|
MOVB R3, ret+24(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-10-30 13:58:30 -06:00
|
|
|
test:
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVBZU 1(R3), R6
|
|
|
|
|
MOVBZU 1(R4), R7
|
|
|
|
|
CMP R6, R7
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
BEQ loop
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2014-10-30 14:44:42 -06:00
|
|
|
MOVB R0, ret+24(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2015-01-06 17:42:48 -07:00
|
|
|
// memequal_varlen(a, b unsafe.Pointer) bool
|
|
|
|
|
TEXT runtime·memequal_varlen(SB),NOSPLIT,$40-17
|
|
|
|
|
MOVD a+0(FP), R3
|
|
|
|
|
MOVD b+8(FP), R4
|
|
|
|
|
CMP R3, R4
|
|
|
|
|
BEQ eq
|
|
|
|
|
MOVD 8(R11), R5 // compiler stores size at offset 8 in the closure
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVD R3, FIXED_FRAME+0(R1)
|
|
|
|
|
MOVD R4, FIXED_FRAME+8(R1)
|
|
|
|
|
MOVD R5, FIXED_FRAME+16(R1)
|
2015-01-06 17:42:48 -07:00
|
|
|
BL runtime·memeq(SB)
|
2015-10-08 03:34:29 -06:00
|
|
|
MOVBZ FIXED_FRAME+24(R1), R3
|
2015-01-06 17:42:48 -07:00
|
|
|
MOVB R3, ret+16(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2015-01-06 17:42:48 -07:00
|
|
|
eq:
|
|
|
|
|
MOVD $1, R3
|
|
|
|
|
MOVB R3, ret+16(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2015-01-06 17:42:48 -07:00
|
|
|
|
2014-08-12 17:48:49 -06:00
|
|
|
// eqstring tests whether two strings are equal.
|
2015-02-05 09:23:14 -07:00
|
|
|
// The compiler guarantees that strings passed
|
|
|
|
|
// to eqstring have equal length.
|
2014-08-12 17:48:49 -06:00
|
|
|
// See runtime_test.go:eqstring_generic for
|
2014-10-27 15:27:03 -06:00
|
|
|
// equivalent Go code.
|
2014-08-12 17:48:49 -06:00
|
|
|
TEXT runtime·eqstring(SB),NOSPLIT,$0-33
|
|
|
|
|
MOVD s1str+0(FP), R3
|
|
|
|
|
MOVD s2str+16(FP), R4
|
2015-02-05 09:23:14 -07:00
|
|
|
MOVD $1, R5
|
|
|
|
|
MOVB R5, ret+32(FP)
|
|
|
|
|
CMP R3, R4
|
|
|
|
|
BNE 2(PC)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2015-02-05 09:23:14 -07:00
|
|
|
MOVD s1len+8(FP), R5
|
2014-08-12 17:48:49 -06:00
|
|
|
SUB $1, R3
|
|
|
|
|
SUB $1, R4
|
|
|
|
|
ADD R3, R5, R8
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
loop:
|
2014-08-12 17:48:49 -06:00
|
|
|
CMP R3, R8
|
2015-02-05 09:23:14 -07:00
|
|
|
BNE 2(PC)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVBZU 1(R3), R6
|
|
|
|
|
MOVBZU 1(R4), R7
|
|
|
|
|
CMP R6, R7
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
BEQ loop
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVB R0, ret+32(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
// TODO: share code with memeq?
|
|
|
|
|
TEXT bytes·Equal(SB),NOSPLIT,$0-49
|
|
|
|
|
MOVD a_len+8(FP), R3
|
|
|
|
|
MOVD b_len+32(FP), R4
|
|
|
|
|
|
|
|
|
|
CMP R3, R4 // unequal lengths are not equal
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
BNE noteq
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
MOVD a+0(FP), R5
|
|
|
|
|
MOVD b+24(FP), R6
|
|
|
|
|
SUB $1, R5
|
|
|
|
|
SUB $1, R6
|
|
|
|
|
ADD R5, R3 // end-1
|
|
|
|
|
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
loop:
|
2014-08-12 17:48:49 -06:00
|
|
|
CMP R5, R3
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
BEQ equal // reached the end
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVBZU 1(R5), R4
|
|
|
|
|
MOVBZU 1(R6), R7
|
|
|
|
|
CMP R4, R7
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
BEQ loop
|
2014-08-12 17:48:49 -06:00
|
|
|
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
noteq:
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVBZ R0, ret+48(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
equal:
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD $1, R3
|
|
|
|
|
MOVBZ R3, ret+48(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
TEXT bytes·IndexByte(SB),NOSPLIT,$0-40
|
|
|
|
|
MOVD s+0(FP), R3
|
|
|
|
|
MOVD s_len+8(FP), R4
|
|
|
|
|
MOVBZ c+24(FP), R5 // byte to find
|
2014-08-13 12:36:29 -06:00
|
|
|
MOVD R3, R6 // store base for later
|
2014-08-12 17:48:49 -06:00
|
|
|
SUB $1, R3
|
|
|
|
|
ADD R3, R4 // end-1
|
|
|
|
|
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
loop:
|
2014-08-12 17:48:49 -06:00
|
|
|
CMP R3, R4
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
BEQ notfound
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVBZU 1(R3), R7
|
|
|
|
|
CMP R7, R5
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
BNE loop
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
SUB R6, R3 // remove base
|
|
|
|
|
MOVD R3, ret+32(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
notfound:
|
2014-08-13 12:36:29 -06:00
|
|
|
MOVD $-1, R3
|
|
|
|
|
MOVD R3, ret+32(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2015-03-06 22:18:16 -07:00
|
|
|
TEXT strings·IndexByte(SB),NOSPLIT,$0-32
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD p+0(FP), R3
|
|
|
|
|
MOVD b_len+8(FP), R4
|
|
|
|
|
MOVBZ c+16(FP), R5 // byte to find
|
2014-08-13 12:36:29 -06:00
|
|
|
MOVD R3, R6 // store base for later
|
2014-08-12 17:48:49 -06:00
|
|
|
SUB $1, R3
|
|
|
|
|
ADD R3, R4 // end-1
|
|
|
|
|
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
loop:
|
2014-08-12 17:48:49 -06:00
|
|
|
CMP R3, R4
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
BEQ notfound
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVBZU 1(R3), R7
|
|
|
|
|
CMP R7, R5
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
BNE loop
|
2014-08-12 17:48:49 -06:00
|
|
|
|
|
|
|
|
SUB R6, R3 // remove base
|
|
|
|
|
MOVD R3, ret+24(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
[dev.power64] cmd/5a, cmd/6a, cmd/8a, cmd/9a: make labels function-scoped
I removed support for jumping between functions years ago,
as part of doing the instruction layout for each function separately.
Given that, it makes sense to treat labels as function-scoped.
This lets each function have its own 'loop' label, for example.
Makes the assembly much cleaner and removes the last
reason anyone would reach for the 123(PC) form instead.
Note that this is on the dev.power64 branch, but it changes all
the assemblers. The change will ship in Go 1.5 (perhaps after
being ported into the new assembler).
Came up as part of CL 167730043.
LGTM=r
R=r
CC=austin, dave, golang-codereviews, minux
https://golang.org/cl/159670043
2014-10-28 19:50:16 -06:00
|
|
|
notfound:
|
2014-08-13 12:36:29 -06:00
|
|
|
MOVD $-1, R3
|
|
|
|
|
MOVD R3, ret+24(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·cmpstring(SB),NOSPLIT|NOFRAME,$0-40
|
2015-08-28 21:59:04 -06:00
|
|
|
MOVD s1_base+0(FP), R5
|
|
|
|
|
MOVD s1_len+8(FP), R3
|
|
|
|
|
MOVD s2_base+16(FP), R6
|
|
|
|
|
MOVD s2_len+24(FP), R4
|
|
|
|
|
MOVD $ret+32(FP), R7
|
|
|
|
|
BR runtime·cmpbody<>(SB)
|
|
|
|
|
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT bytes·Compare(SB),NOSPLIT|NOFRAME,$0-56
|
2015-08-28 21:59:04 -06:00
|
|
|
MOVD s1+0(FP), R5
|
|
|
|
|
MOVD s1+8(FP), R3
|
|
|
|
|
MOVD s2+24(FP), R6
|
|
|
|
|
MOVD s2+32(FP), R4
|
|
|
|
|
MOVD $ret+48(FP), R7
|
|
|
|
|
BR runtime·cmpbody<>(SB)
|
|
|
|
|
|
|
|
|
|
// On entry:
|
|
|
|
|
// R3 is the length of s1
|
|
|
|
|
// R4 is the length of s2
|
|
|
|
|
// R5 points to the start of s1
|
|
|
|
|
// R6 points to the start of s2
|
|
|
|
|
// R7 points to return value (-1/0/1 will be written here)
|
|
|
|
|
//
|
|
|
|
|
// On exit:
|
|
|
|
|
// R5, R6, R8, R9 and R10 are clobbered
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·cmpbody<>(SB),NOSPLIT|NOFRAME,$0-0
|
2015-08-28 21:59:04 -06:00
|
|
|
CMP R5, R6
|
|
|
|
|
BEQ samebytes // same starting pointers; compare lengths
|
|
|
|
|
SUB $1, R5
|
|
|
|
|
SUB $1, R6
|
|
|
|
|
MOVD R4, R8
|
|
|
|
|
CMP R3, R4
|
|
|
|
|
BGE 2(PC)
|
|
|
|
|
MOVD R3, R8 // R8 is min(R3, R4)
|
|
|
|
|
ADD R5, R8 // R5 is current byte in s1, R8 is last byte in s1 to compare
|
|
|
|
|
loop:
|
|
|
|
|
CMP R5, R8
|
|
|
|
|
BEQ samebytes // all compared bytes were the same; compare lengths
|
|
|
|
|
MOVBZU 1(R5), R9
|
|
|
|
|
MOVBZU 1(R6), R10
|
|
|
|
|
CMP R9, R10
|
|
|
|
|
BEQ loop
|
|
|
|
|
// bytes differed
|
|
|
|
|
MOVD $1, R4
|
|
|
|
|
BGT 2(PC)
|
|
|
|
|
NEG R4
|
|
|
|
|
MOVD R4, (R7)
|
|
|
|
|
RET
|
|
|
|
|
samebytes:
|
|
|
|
|
MOVD $1, R8
|
|
|
|
|
CMP R3, R4
|
|
|
|
|
BNE 3(PC)
|
|
|
|
|
MOVD R0, (R7)
|
|
|
|
|
RET
|
|
|
|
|
BGT 2(PC)
|
|
|
|
|
NEG R8
|
|
|
|
|
MOVD R8, (R7)
|
|
|
|
|
RET
|
|
|
|
|
|
2014-10-27 15:27:03 -06:00
|
|
|
TEXT runtime·fastrand1(SB), NOSPLIT, $0-4
|
2014-08-12 17:48:49 -06:00
|
|
|
MOVD g_m(g), R4
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVWZ m_fastrand(R4), R3
|
2014-08-12 17:48:49 -06:00
|
|
|
ADD R3, R3
|
2014-10-31 11:39:36 -06:00
|
|
|
CMPW R3, $0
|
2014-08-12 17:48:49 -06:00
|
|
|
BGE 2(PC)
|
|
|
|
|
XOR $0x88888eef, R3
|
2014-10-27 15:27:03 -06:00
|
|
|
MOVW R3, m_fastrand(R4)
|
|
|
|
|
MOVW R3, ret+0(FP)
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2014-10-27 15:27:03 -06:00
|
|
|
TEXT runtime·return0(SB), NOSPLIT, $0
|
|
|
|
|
MOVW $0, R3
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-08-12 17:48:49 -06:00
|
|
|
|
2014-10-27 15:27:03 -06:00
|
|
|
// Called from cgo wrappers, this function returns g->m->curg.stack.hi.
|
|
|
|
|
// Must obey the gcc calling convention.
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT _cgo_topofstack(SB),NOSPLIT|NOFRAME,$0
|
2014-12-16 16:34:55 -07:00
|
|
|
// g (R30) and R31 are callee-save in the C ABI, so save them
|
|
|
|
|
MOVD g, R4
|
|
|
|
|
MOVD R31, R5
|
|
|
|
|
MOVD LR, R6
|
|
|
|
|
|
|
|
|
|
BL runtime·load_g(SB) // clobbers g (R30), R31
|
|
|
|
|
MOVD g_m(g), R3
|
|
|
|
|
MOVD m_curg(R3), R3
|
|
|
|
|
MOVD (g_stack+stack_hi)(R3), R3
|
|
|
|
|
|
|
|
|
|
MOVD R4, g
|
|
|
|
|
MOVD R5, R31
|
|
|
|
|
MOVD R6, LR
|
|
|
|
|
RET
|
2014-11-03 08:53:11 -07:00
|
|
|
|
|
|
|
|
// The top-most function running on a goroutine
|
|
|
|
|
// returns to goexit+PCQuantum.
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
//
|
|
|
|
|
// When dynamically linking Go, it can be returned to from a function
|
|
|
|
|
// implemented in a different module and so needs to reload the TOC pointer
|
|
|
|
|
// from the stack (although this function declares that it does not set up x-a
|
|
|
|
|
// frame, newproc1 does in fact allocate one for goexit and saves the TOC
|
|
|
|
|
// pointer in the correct place).
|
|
|
|
|
// goexit+_PCQuantum is halfway through the usual global entry point prologue
|
|
|
|
|
// that derives r2 from r12 which is a bit silly, but not harmful.
|
2015-10-08 17:44:27 -06:00
|
|
|
TEXT runtime·goexit(SB),NOSPLIT|NOFRAME,$0-0
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
MOVD 24(R1), R2
|
2014-11-03 08:53:11 -07:00
|
|
|
BL runtime·goexit1(SB) // does not return
|
2015-02-20 10:07:02 -07:00
|
|
|
// traceback from goexit1 must hit code range of goexit
|
|
|
|
|
MOVD R0, R0 // NOP
|
2014-11-18 13:19:26 -07:00
|
|
|
|
2014-11-21 13:57:10 -07:00
|
|
|
TEXT runtime·prefetcht0(SB),NOSPLIT,$0-8
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-11-21 13:57:10 -07:00
|
|
|
|
|
|
|
|
TEXT runtime·prefetcht1(SB),NOSPLIT,$0-8
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-11-21 13:57:10 -07:00
|
|
|
|
|
|
|
|
TEXT runtime·prefetcht2(SB),NOSPLIT,$0-8
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2014-11-21 13:57:10 -07:00
|
|
|
|
|
|
|
|
TEXT runtime·prefetchnta(SB),NOSPLIT,$0-8
|
2015-06-03 12:59:27 -06:00
|
|
|
RET
|
2015-10-11 18:20:10 -06:00
|
|
|
|
|
|
|
|
TEXT runtime·sigreturn(SB),NOSPLIT,$0-8
|
|
|
|
|
RET
|
cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC
The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:
* Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
* A call to a function that *might* have to go via a PLT stub must be followed
by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
to restore the TOC pointer (only when dynamically linking Go code).
* When calling a function via a function pointer, the address of the function
must be in r12, and the first couple of instructions (the "global entry
point") of the called function use this to derive the address of the TOC
for the module it is in.
* When calling a function that is implemented in the same module, the system
linker adjusts the call to skip over the instructions mentioned above (the
"local entry point"), assuming that r2 is already correctly set.
So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.
Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967
Reviewed-by: Russ Cox <rsc@golang.org>
2015-10-15 20:42:09 -06:00
|
|
|
|
|
|
|
|
// prepGoExitFrame saves the current TOC pointer (i.e. the TOC pointer for the
|
|
|
|
|
// module containing runtime) to the frame that goexit will execute in when
|
|
|
|
|
// the goroutine exits. It's implemented in assembly mainly because that's the
|
|
|
|
|
// easiest way to get access to R2.
|
|
|
|
|
TEXT runtime·prepGoExitFrame(SB),NOSPLIT,$0-8
|
|
|
|
|
MOVD sp+0(FP), R3
|
|
|
|
|
MOVD R2, 24(R3)
|
|
|
|
|
RET
|
2015-10-16 02:22:20 -06:00
|
|
|
|
|
|
|
|
TEXT runtime·addmoduledata(SB),NOSPLIT|NOFRAME,$0-0
|
|
|
|
|
ADD $-8, R1
|
|
|
|
|
MOVD R31, 0(R1)
|
|
|
|
|
MOVD runtime·lastmoduledatap(SB), R4
|
|
|
|
|
MOVD R3, moduledata_next(R4)
|
|
|
|
|
MOVD R3, runtime·lastmoduledatap(SB)
|
|
|
|
|
MOVD 0(R1), R31
|
|
|
|
|
ADD $8, R1
|
|
|
|
|
RET
|
