cmd/6g, liblink, runtime: support saving base pointers

This adds a "framepointer" GOEXPERIMENT that that makes the amd64
toolchain maintain base pointer chains in the same way that gcc
-fno-omit-frame-pointer does.  Go doesn't use these saved base
pointers, but this does enable external tools like Linux perf and
VTune to unwind Go stacks when collecting system-wide profiles.

This requires support in the compilers to not clobber BP, support in
liblink for generating the BP-saving function prologue and unwinding
epilogue, and support in the runtime to save BPs across preemption, to
skip saved BPs during stack unwinding and, and to adjust saved BPs
during stack moving.

As with other GOEXPERIMENTs, everything from the toolchain to the
runtime must be compiled with this experiment enabled.  To do this,
run make.bash (or all.bash) with GOEXPERIMENT=framepointer.

Change-Id: I4024853beefb9539949e5ca381adfdd9cfada544
Reviewed-on: https://go-review.googlesource.com/2992
Reviewed-by: Russ Cox <rsc@golang.org>

src/cmd/6g/gsubr.c

@@ -285,6 +285,9 @@ ginit(void)
 	if(nacl) {
 		reg[REG_BP]++;
 		reg[REG_R15]++;
+	} else if(framepointer_enabled) {
+		// BP is part of the calling convention of framepointer_enabled.
+		reg[REG_BP]++;
 	}
 }
 
@@ -298,6 +301,8 @@ gclean(void)
 	if(nacl) {
 		reg[REG_BP]--;
 		reg[REG_R15]--;
+	} else if(framepointer_enabled) {
+		reg[REG_BP]--;
 	}
 
 

src/cmd/6g/reg.c

@@ -1186,6 +1186,9 @@ BtoR(uint32 b)
 	b &= 0xffffL;
 	if(nacl)
 		b &= ~((1<<(REG_BP-REG_AX)) | (1<<(REG_R15-REG_AX)));
+	else if(framepointer_enabled)
+		// BP is part of the calling convention if framepointer_enabled.
+		b &= ~(1<<(REG_BP-REG_AX));
 	if(b == 0)
 		return 0;
 	return bitno(b) + REG_AX;

src/liblink/go.c

@@ -21,7 +21,7 @@ static struct {
 	int *val;
 } exper[] = {
 	{"fieldtrack", &fieldtrack_enabled},
-	{"basepointer", &framepointer_enabled}, 
+	{"framepointer", &framepointer_enabled},
 };
 
 static void

src/liblink/obj6.c

@@ -379,7 +379,7 @@ preprocess(Link *ctxt, LSym *cursym)
 {
 	Prog *p, *q, *p1, *p2;
 	int32 autoffset, deltasp;
-	int a, pcsize;
+	int a, pcsize, bpsize;
 	vlong textstksiz, textarg;
 
 	if(ctxt->tlsg == nil)
@@ -403,6 +403,18 @@ preprocess(Link *ctxt, LSym *cursym)
 	if(autoffset < 0)
 		autoffset = 0;
 	
+	if(framepointer_enabled && autoffset > 0) {
+		// Make room for to save a base pointer.  If autoffset == 0,
+		// this might do something special like a tail jump to
+		// another function, so in that case we omit this.
+		bpsize = ctxt->arch->ptrsize;
+		autoffset += bpsize;
+		textstksiz += bpsize;
+		p->to.offset = ((uint64)p->to.offset & (0xffffffffull<<32)) | (uint32)autoffset;
+	} else {
+		bpsize = 0;
+	}
+
 	cursym->args = p->to.offset>>32;
 	cursym->locals = textstksiz;
 
@@ -448,6 +460,28 @@ preprocess(Link *ctxt, LSym *cursym)
 		q->pcond = p;
 	deltasp = autoffset;
 
+	if(bpsize > 0) {
+		// Save caller's BP
+		p = appendp(ctxt, p);
+		p->as = AMOVQ;
+		p->from.type = TYPE_REG;
+		p->from.reg = REG_BP;
+		p->to.type = TYPE_MEM;
+		p->to.reg = REG_SP;
+		p->to.scale = 1;
+		p->to.offset = autoffset - bpsize;
+
+		// Move current frame to BP
+		p = appendp(ctxt, p);
+		p->as = ALEAQ;
+		p->from.type = TYPE_MEM;
+		p->from.reg = REG_SP;
+		p->from.scale = 1;
+		p->from.offset = autoffset - bpsize;
+		p->to.type = TYPE_REG;
+		p->to.reg = REG_BP;
+	}
+	
 	if(cursym->text->from.scale & WRAPPER) {
 		// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
 		//
@@ -580,12 +614,12 @@ preprocess(Link *ctxt, LSym *cursym)
 		pcsize = p->mode/8;
 		a = p->from.name;
 		if(a == NAME_AUTO)
-			p->from.offset += deltasp;
+			p->from.offset += deltasp - bpsize;
 		if(a == NAME_PARAM)
 			p->from.offset += deltasp + pcsize;
 		a = p->to.name;
 		if(a == NAME_AUTO)
-			p->to.offset += deltasp;
+			p->to.offset += deltasp - bpsize;
 		if(a == NAME_PARAM)
 			p->to.offset += deltasp + pcsize;
 
@@ -630,6 +664,18 @@ preprocess(Link *ctxt, LSym *cursym)
 			ctxt->diag("unbalanced PUSH/POP");
 
 		if(autoffset) {
+			if(bpsize > 0) {
+				// Restore caller's BP
+				p->as = AMOVQ;
+				p->from.type = TYPE_MEM;
+				p->from.reg = REG_SP;
+				p->from.scale = 1;
+				p->from.offset = autoffset - bpsize;
+				p->to.type = TYPE_REG;
+				p->to.reg = REG_BP;
+				p = appendp(ctxt, p);
+			}
+
 			p->as = AADJSP;
 			p->from.type = TYPE_CONST;
 			p->from.offset = -autoffset;

src/runtime/arch1_amd64.go

@@ -8,7 +8,7 @@ const (
 	thechar           = '6'
 	_BigEndian        = 0
 	_CacheLineSize    = 64
-	_RuntimeGogoBytes = 64 + (goos_plan9|goos_solaris|goos_windows)*16
+	_RuntimeGogoBytes = 80 + (goos_plan9|goos_solaris|goos_windows)*16
 	_PhysPageSize     = 4096
 	_PCQuantum        = 1
 	_Int64Align       = 8

src/runtime/asm_amd64.s

@@ -134,6 +134,7 @@ TEXT runtime·gosave(SB), NOSPLIT, $0-8
 	MOVQ	BX, gobuf_pc(AX)
 	MOVQ	$0, gobuf_ret(AX)
 	MOVQ	$0, gobuf_ctxt(AX)
+	MOVQ	BP, gobuf_bp(AX)
 	get_tls(CX)
 	MOVQ	g(CX), BX
 	MOVQ	BX, gobuf_g(AX)
@@ -150,9 +151,11 @@ TEXT runtime·gogo(SB), NOSPLIT, $0-8
 	MOVQ	gobuf_sp(BX), SP	// restore SP
 	MOVQ	gobuf_ret(BX), AX
 	MOVQ	gobuf_ctxt(BX), DX
+	MOVQ	gobuf_bp(BX), BP
 	MOVQ	$0, gobuf_sp(BX)	// clear to help garbage collector
 	MOVQ	$0, gobuf_ret(BX)
 	MOVQ	$0, gobuf_ctxt(BX)
+	MOVQ	$0, gobuf_bp(BX)
 	MOVQ	gobuf_pc(BX), BX
 	JMP	BX
 
@@ -170,6 +173,7 @@ TEXT runtime·mcall(SB), NOSPLIT, $0-8
 	LEAQ	fn+0(FP), BX	// caller's SP
 	MOVQ	BX, (g_sched+gobuf_sp)(AX)
 	MOVQ	AX, (g_sched+gobuf_g)(AX)
+	MOVQ	BP, (g_sched+gobuf_bp)(AX)
 
 	// switch to m->g0 & its stack, call fn
 	MOVQ	g(CX), BX
@@ -228,6 +232,7 @@ switch:
 	MOVQ	SI, (g_sched+gobuf_pc)(AX)
 	MOVQ	SP, (g_sched+gobuf_sp)(AX)
 	MOVQ	AX, (g_sched+gobuf_g)(AX)
+	MOVQ	BP, (g_sched+gobuf_bp)(AX)
 
 	// switch to g0
 	MOVQ	DX, g(CX)
@@ -303,6 +308,7 @@ TEXT runtime·morestack(SB),NOSPLIT,$0-0
 	LEAQ	8(SP), AX // f's SP
 	MOVQ	AX, (g_sched+gobuf_sp)(SI)
 	MOVQ	DX, (g_sched+gobuf_ctxt)(SI)
+	MOVQ	BP, (g_sched+gobuf_bp)(SI)
 
 	// Call newstack on m->g0's stack.
 	MOVQ	m_g0(BX), BX
@@ -592,6 +598,7 @@ TEXT gosave<>(SB),NOSPLIT,$0
 	MOVQ	R9, (g_sched+gobuf_sp)(R8)
 	MOVQ	$0, (g_sched+gobuf_ret)(R8)
 	MOVQ	$0, (g_sched+gobuf_ctxt)(R8)
+	MOVQ	BP, (g_sched+gobuf_bp)(R8)
 	RET
 
 // asmcgocall(void(*fn)(void*), void *arg)
@@ -747,17 +754,30 @@ havem:
 	MOVQ	(g_sched+gobuf_sp)(SI), DI  // prepare stack as DI
 	MOVQ	(g_sched+gobuf_pc)(SI), BX
 	MOVQ	BX, -8(DI)
-	LEAQ	-(8+8)(DI), SP
+	// Compute the size of the frame, including return PC and, if
+	// GOEXPERIMENT=framepointer, the saved based pointer
+	LEAQ	x+0(FP), AX
+	SUBQ	SP, AX
+	SUBQ	AX, DI
+	MOVQ	DI, SP
+
 	MOVQ	R8, 0(SP)
 	CALL	runtime·cgocallbackg(SB)
 	MOVQ	0(SP), R8
 
+	// Compute the size of the frame again.  FP and SP have
+	// completely different values here than they did above,
+	// but only their difference matters.
+	LEAQ	x+0(FP), AX
+	SUBQ	SP, AX
+
 	// Restore g->sched (== m->curg->sched) from saved values.
 	get_tls(CX)
 	MOVQ	g(CX), SI
-	MOVQ	8(SP), BX
+	MOVQ	SP, DI
+	ADDQ	AX, DI
+	MOVQ	-8(DI), BX
 	MOVQ	BX, (g_sched+gobuf_pc)(SI)
-	LEAQ	(8+8)(SP), DI
 	MOVQ	DI, (g_sched+gobuf_sp)(SI)
 
 	// Switch back to m->g0's stack and restore m->g0->sched.sp.

src/runtime/cgocall.go

@@ -225,6 +225,11 @@ func cgocallbackg1() {
 		cb = (*args)(unsafe.Pointer(sp + 4*ptrSize))
 	case "amd64":
 		// On amd64, stack frame is one word, plus caller PC.
+		if framepointer_enabled {
+			// In this case, there's also saved BP.
+			cb = (*args)(unsafe.Pointer(sp + 3*ptrSize))
+			break
+		}
 		cb = (*args)(unsafe.Pointer(sp + 2*ptrSize))
 	case "386":
 		// On 386, stack frame is three words, plus caller PC.

src/runtime/proc1.go

@@ -113,6 +113,9 @@ func schedinit() {
 
 	sched.maxmcount = 10000
 
+	// Cache the framepointer experiment.  This affects stack unwinding.
+	framepointer_enabled = haveexperiment("framepointer")
+
 	tracebackinit()
 	symtabinit()
 	stackinit()

src/runtime/runtime2.go

@@ -125,6 +125,7 @@ type gobuf struct {
 	ctxt unsafe.Pointer // this has to be a pointer so that gc scans it
 	ret  uintreg
 	lr   uintptr
+	bp   uintptr // for GOEXPERIMENT=framepointer
 }
 
 // Known to compiler.

src/runtime/stack1.go

@@ -46,6 +46,9 @@ var stackpoolmu mutex
 
 var stackfreequeue stack
 
+// Cached value of haveexperiment("framepointer")
+var framepointer_enabled bool
+
 func stackinit() {
 	if _StackCacheSize&_PageMask != 0 {
 		throw("cache size must be a multiple of page size")
@@ -308,6 +311,8 @@ var mapnames = []string{
 // | args from caller |
 // +------------------+ <- frame->argp
 // |  return address  |
+// +------------------+
+// |  caller's BP (*) | (*) if framepointer_enabled && varp < sp
 // +------------------+ <- frame->varp
 // |     locals       |
 // +------------------+
@@ -460,6 +465,18 @@ func adjustframe(frame *stkframe, arg unsafe.Pointer) bool {
 		adjustpointers(unsafe.Pointer(frame.varp-size), &bv, adjinfo, f)
 	}
 
+	// Adjust saved base pointer if there is one.
+	if thechar == '6' && frame.argp-frame.varp == 2*ptrSize {
+		if !framepointer_enabled {
+			print("runtime: found space for saved base pointer, but no framepointer experiment")
+			throw("bad frame layout")
+		}
+		if stackDebug >= 3 {
+			print("      saved bp\n")
+		}
+		adjustpointer(adjinfo, unsafe.Pointer(frame.varp))
+	}
+
 	// Adjust arguments.
 	if frame.arglen > 0 {
 		var bv bitvector

src/runtime/traceback.go

@@ -232,6 +232,12 @@ func gentraceback(pc0 uintptr, sp0 uintptr, lr0 uintptr, gp *g, skip int, pcbuf
 			frame.varp -= regSize
 		}
 
+		// If framepointer_enabled and there's a frame, then
+		// there's a saved bp here.
+		if GOARCH == "amd64" && frame.varp > frame.sp && framepointer_enabled {
+			frame.varp -= ptrSize
+		}
+
 		// Derive size of arguments.
 		// Most functions have a fixed-size argument block,
 		// so we can use metadata about the function f.

test/nosplit.go

@@ -185,6 +185,26 @@ func main() {
 		goarch = runtime.GOARCH
 	}
 
+	thechar := ""
+	if gochar, err := exec.Command("go", "env", "GOCHAR").Output(); err != nil {
+		bug()
+		fmt.Printf("running go env GOCHAR: %v\n", err)
+		return
+	} else {
+		thechar = strings.TrimSpace(string(gochar))
+	}
+
+	version, err := exec.Command("go", "tool", thechar+"g", "-V").Output()
+	if err != nil {
+		bug()
+		fmt.Printf("running go tool %sg -V: %v\n", thechar, err)
+		return
+	}
+	if strings.Contains(string(version), "framepointer") {
+		// Skip this test if GOEXPERIMENT=framepointer
+		return
+	}
+
 	dir, err := ioutil.TempDir("", "go-test-nosplit")
 	if err != nil {
 		bug()