[dev.garbage] runtime: add write barrier to casp

Also rewrite some casp that don't use real pointers
to use casuintptr instead.

LGTM=rlh
R=rlh
CC=golang-codereviews
https://golang.org/cl/166440044

src/runtime/asm_386.s

@@ -502,7 +502,7 @@ fail:
 //		return 1;
 //	}else
 //		return 0;
-TEXT runtime·casp(SB), NOSPLIT, $0-13
+TEXT runtime·casp1(SB), NOSPLIT, $0-13
 	MOVL	ptr+0(FP), BX
 	MOVL	old+4(FP), AX
 	MOVL	new+8(FP), CX
@@ -537,7 +537,7 @@ TEXT runtime·xchg(SB), NOSPLIT, $0-12
 	MOVL	AX, ret+8(FP)
 	RET
 
-TEXT runtime·xchgp(SB), NOSPLIT, $0-12
+TEXT runtime·xchgp1(SB), NOSPLIT, $0-12
 	MOVL	ptr+0(FP), BX
 	MOVL	new+4(FP), AX
 	XCHGL	AX, 0(BX)
@@ -555,7 +555,7 @@ again:
 	JNZ	again
 	RET
 
-TEXT runtime·atomicstorep(SB), NOSPLIT, $0-8
+TEXT runtime·atomicstorep1(SB), NOSPLIT, $0-8
 	MOVL	ptr+0(FP), BX
 	MOVL	val+4(FP), AX
 	XCHGL	AX, 0(BX)

src/runtime/asm_amd64.s

@@ -489,7 +489,7 @@ TEXT runtime·atomicstoreuintptr(SB), NOSPLIT, $0-16
 //		return 1;
 //	} else
 //		return 0;
-TEXT runtime·casp(SB), NOSPLIT, $0-25
+TEXT runtime·casp1(SB), NOSPLIT, $0-25
 	MOVQ	ptr+0(FP), BX
 	MOVQ	old+8(FP), AX
 	MOVQ	new+16(FP), CX
@@ -541,7 +541,7 @@ TEXT runtime·xchg64(SB), NOSPLIT, $0-24
 	MOVQ	AX, ret+16(FP)
 	RET
 
-TEXT runtime·xchgp(SB), NOSPLIT, $0-24
+TEXT runtime·xchgp1(SB), NOSPLIT, $0-24
 	MOVQ	ptr+0(FP), BX
 	MOVQ	new+8(FP), AX
 	XCHGQ	AX, 0(BX)
@@ -559,7 +559,7 @@ again:
 	JNZ	again
 	RET
 
-TEXT runtime·atomicstorep(SB), NOSPLIT, $0-16
+TEXT runtime·atomicstorep1(SB), NOSPLIT, $0-16
 	MOVQ	ptr+0(FP), BX
 	MOVQ	val+8(FP), AX
 	XCHGQ	AX, 0(BX)

src/runtime/asm_amd64p32.s

@@ -460,7 +460,7 @@ fail:
 //		return 1;
 //	} else
 //		return 0;
-TEXT runtime·casp(SB), NOSPLIT, $0-17
+TEXT runtime·casp1(SB), NOSPLIT, $0-17
 	MOVL	ptr+0(FP), BX
 	MOVL	old+4(FP), AX
 	MOVL	new+8(FP), CX
@@ -512,7 +512,7 @@ TEXT runtime·xchg64(SB), NOSPLIT, $0-24
 	MOVQ	AX, ret+16(FP)
 	RET
 
-TEXT runtime·xchgp(SB), NOSPLIT, $0-12
+TEXT runtime·xchgp1(SB), NOSPLIT, $0-12
 	MOVL	ptr+0(FP), BX
 	MOVL	new+4(FP), AX
 	XCHGL	AX, 0(BX)
@@ -530,7 +530,7 @@ again:
 	JNZ	again
 	RET
 
-TEXT runtime·atomicstorep(SB), NOSPLIT, $0-8
+TEXT runtime·atomicstorep1(SB), NOSPLIT, $0-8
 	MOVL	ptr+0(FP), BX
 	MOVL	val+4(FP), AX
 	XCHGL	AX, 0(BX)

src/runtime/asm_power64x.s

@@ -472,7 +472,7 @@ TEXT runtime·atomicstoreuintptr(SB), NOSPLIT, $0-16
 //		return 1;
 //	} else
 //		return 0;
-TEXT runtime·casp(SB), NOSPLIT, $0-25
+TEXT runtime·casp1(SB), NOSPLIT, $0-25
 	BR runtime·cas64(SB)
 
 // uint32 xadd(uint32 volatile *val, int32 delta)
@@ -529,7 +529,7 @@ TEXT runtime·xchg64(SB), NOSPLIT, $0-24
 	MOVD	R3, ret+16(FP)
 	RETURN
 
-TEXT runtime·xchgp(SB), NOSPLIT, $0-24
+TEXT runtime·xchgp1(SB), NOSPLIT, $0-24
 	BR	runtime·xchg64(SB)
 
 TEXT runtime·xchguintptr(SB), NOSPLIT, $0-24
@@ -538,7 +538,7 @@ TEXT runtime·xchguintptr(SB), NOSPLIT, $0-24
 TEXT runtime·procyield(SB),NOSPLIT,$0-0
 	RETURN
 
-TEXT runtime·atomicstorep(SB), NOSPLIT, $0-16
+TEXT runtime·atomicstorep1(SB), NOSPLIT, $0-16
 	BR	runtime·atomicstore64(SB)
 
 TEXT runtime·atomicstore(SB), NOSPLIT, $0-12

src/runtime/atomic.go

@@ -20,8 +20,16 @@ func xchg(ptr *uint32, new uint32) uint32
 //go:noescape
 func xchg64(ptr *uint64, new uint64) uint64
 
-//go:noescape
-func xchgp(ptr unsafe.Pointer, new unsafe.Pointer) unsafe.Pointer
+// Cannot use noescape here: ptr does not but new does escape.
+// Instead use noescape(ptr) in wrapper below.
+func xchgp1(ptr unsafe.Pointer, new unsafe.Pointer) unsafe.Pointer
+
+//go:nosplit
+func xchgp(ptr unsafe.Pointer, new unsafe.Pointer) unsafe.Pointer {
+	old := xchgp1(noescape(ptr), new)
+	writebarrierptr_nostore((*uintptr)(ptr), uintptr(new))
+	return old
+}
 
 //go:noescape
 func xchguintptr(ptr *uintptr, new uintptr) uintptr
@@ -47,5 +55,27 @@ func atomicstore(ptr *uint32, val uint32)
 //go:noescape
 func atomicstore64(ptr *uint64, val uint64)
 
-//go:noescape
-func atomicstorep(ptr unsafe.Pointer, val unsafe.Pointer)
+// Cannot use noescape here: ptr does not but val does escape.
+// Instead use noescape(ptr) in wrapper below.
+func atomicstorep1(ptr unsafe.Pointer, val unsafe.Pointer)
+
+//go:nosplit
+func atomicstorep(ptr unsafe.Pointer, val unsafe.Pointer) {
+	atomicstorep1(noescape(ptr), val)
+	// TODO(rsc): Why does the compiler think writebarrierptr_nostore's dst argument escapes?
+	writebarrierptr_nostore((*uintptr)(noescape(ptr)), uintptr(val))
+}
+
+// Cannot use noescape here: ptr does not but new does escape.
+// Instead use noescape(ptr) in wrapper below.
+func casp1(ptr *unsafe.Pointer, old, new unsafe.Pointer) bool
+
+//go:nosplit
+func casp(ptr *unsafe.Pointer, old, new unsafe.Pointer) bool {
+	ok := casp1((*unsafe.Pointer)(noescape(unsafe.Pointer(ptr))), old, new)
+	if !ok {
+		return false
+	}
+	writebarrierptr_nostore((*uintptr)(unsafe.Pointer(ptr)), uintptr(new))
+	return true
+}

src/runtime/mgc0.c

@@ -1098,7 +1098,6 @@ runtime·gcmarkwb_m()
 	slot = (byte**)g->m->scalararg[0];
 	ptr = (byte*)g->m->scalararg[1];
 
-	*slot = ptr;
 	switch(runtime·gcphase) {
 	default:
 		runtime·throw("gcphasework in bad gcphase");

src/runtime/mgc0.go

@@ -92,13 +92,24 @@ const (
 // but if we do that, Go inserts a write barrier on *dst = src.
 //go:nosplit
 func writebarrierptr(dst *uintptr, src uintptr) {
+	*dst = src
+	writebarrierptr_nostore(dst, src)
+}
+
+// Like writebarrierptr, but the store has already been applied.
+// Do not reapply.
+//go:nosplit
+func writebarrierptr_nostore(dst *uintptr, src uintptr) {
+	if getg() == nil { // very low-level startup
+		return
+	}
+
 	if src != 0 && (src < _PageSize || src == _PoisonGC || src == _PoisonStack) {
 		onM(func() { gothrow("bad pointer in write barrier") })
 	}
 
 	mp := acquirem()
 	if mp.inwb {
-		*dst = src
 		releasem(mp)
 		return
 	}
@@ -112,7 +123,6 @@ func writebarrierptr(dst *uintptr, src uintptr) {
 	mp.scalararg[1] = oldscalar1
 	mp.inwb = false
 	releasem(mp)
-	//	*dst = src is done inside of the write barrier.
 }
 
 //go:nosplit

src/runtime/proc.c

@@ -1060,7 +1060,7 @@ runtime·dropm(void)
 	unlockextra(mp);
 }
 
-#define MLOCKED ((M*)1)
+#define MLOCKED 1
 
 // lockextra locks the extra list and returns the list head.
 // The caller must unlock the list by storing a new list head
@@ -1071,28 +1071,28 @@ runtime·dropm(void)
 static M*
 lockextra(bool nilokay)
 {
-	M *mp;
+	uintptr mpx;
 	void (*yield)(void);
 
 	for(;;) {
-		mp = runtime·atomicloadp(&runtime·extram);
-		if(mp == MLOCKED) {
+		mpx = runtime·atomicloaduintptr((uintptr*)&runtime·extram);
+		if(mpx == MLOCKED) {
 			yield = runtime·osyield;
 			yield();
 			continue;
 		}
-		if(mp == nil && !nilokay) {
+		if(mpx == 0 && !nilokay) {
 			runtime·usleep(1);
 			continue;
 		}
-		if(!runtime·casp(&runtime·extram, mp, MLOCKED)) {
+		if(!runtime·casuintptr((uintptr*)&runtime·extram, mpx, MLOCKED)) {
 			yield = runtime·osyield;
 			yield();
 			continue;
 		}
 		break;
 	}
-	return mp;
+	return (M*)mpx;
 }
 
 #pragma textflag NOSPLIT

src/runtime/runtime.h

@@ -894,6 +894,7 @@ int32	runtime·round2(int32 x); // round x up to a power of 2.
 bool	runtime·cas(uint32*, uint32, uint32);
 bool	runtime·cas64(uint64*, uint64, uint64);
 bool	runtime·casp(void**, void*, void*);
+bool	runtime·casuintptr(uintptr*, uintptr, uintptr);
 // Don't confuse with XADD x86 instruction,
 // this one is actually 'addx', that is, add-and-fetch.
 uint32	runtime·xadd(uint32 volatile*, int32);

src/runtime/string.c

@@ -48,7 +48,7 @@ runtime·gostringnocopy(byte *str)
 	s.len = runtime·findnull(str);
 	while(true) {
 		ms = runtime·maxstring;
-		if(s.len <= ms || runtime·casp((void**)&runtime·maxstring, (void*)ms, (void*)s.len))
+		if(s.len <= ms || runtime·casuintptr(&runtime·maxstring, ms, s.len))
 			return s;
 	}
 }

src/runtime/stubs.go

@@ -213,9 +213,6 @@ func write(fd uintptr, p unsafe.Pointer, n int32) int32
 //go:noescape
 func cas(ptr *uint32, old, new uint32) bool
 
-//go:noescape
-func casp(ptr *unsafe.Pointer, old, new unsafe.Pointer) bool
-
 //go:noescape
 func casuintptr(ptr *uintptr, old, new uintptr) bool