runtime: add 64-bit atomics

This is factored out part of:
https://golang.org/cl/5279048/
(Parallel GC)

R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/5985047

src/pkg/runtime/arch_386.h

@@ -2,3 +2,7 @@ enum {
 	thechar = '8',
 	CacheLineSize = 64
 };
+
+// prefetches *addr into processor's cache
+#define PREFETCH(addr) runtime·prefetch(addr)
+void	runtime·prefetch(void*);

src/pkg/runtime/arch_amd64.h

@@ -2,3 +2,7 @@ enum {
 	thechar = '6',
 	CacheLineSize = 64
 };
+
+// prefetches *addr into processor's cache
+#define PREFETCH(addr) runtime·prefetch(addr)
+void	runtime·prefetch(void*);

src/pkg/runtime/arch_arm.h

@@ -2,3 +2,5 @@ enum {
 	thechar = '5',
 	CacheLineSize = 32
 };
+
+#define PREFETCH(addr) USED(addr)

src/pkg/runtime/asm_386.s

@@ -299,6 +299,33 @@ TEXT runtime·cas(SB), 7, $0
 	MOVL	$1, AX
 	RET
 
+// bool runtime·cas64(uint64 *val, uint64 *old, uint64 new)
+// Atomically:
+//	if(*val == *old){
+//		*val = new;
+//		return 1;
+//	} else {
+//		*old = *val
+//		return 0;
+//	}
+TEXT runtime·cas64(SB), 7, $0
+	MOVL	4(SP), BP
+	MOVL	8(SP), SI
+	MOVL	0(SI), AX
+	MOVL	4(SI), DX
+	MOVL	12(SP), BX
+	MOVL	16(SP), CX
+	LOCK
+	CMPXCHG8B	0(BP)
+	JNZ	cas64_fail
+	MOVL	$1, AX
+	RET
+cas64_fail:
+	MOVL	AX, 0(SI)
+	MOVL	DX, 4(SI)
+	XORL	AX, AX
+	RET
+
 // bool casp(void **p, void *old, void *new)
 // Atomically:
 //	if(*p == old){
@@ -357,6 +384,43 @@ TEXT runtime·atomicstore(SB), 7, $0
 	XCHGL	AX, 0(BX)
 	RET
 
+// uint64 atomicload64(uint64 volatile* addr);
+// so actually
+// void atomicload64(uint64 *res, uint64 volatile *addr);
+TEXT runtime·atomicload64(SB), 7, $0
+	MOVL    4(SP), BX
+	MOVL	8(SP), AX
+	// MOVQ (%EAX), %MM0
+	BYTE $0x0f; BYTE $0x6f; BYTE $0x00
+	// MOVQ %MM0, 0(%EBX)
+	BYTE $0x0f; BYTE $0x7f; BYTE $0x03
+	// EMMS
+	BYTE $0x0F; BYTE $0x77
+	RET
+
+// void runtime·atomicstore64(uint64 volatile* addr, uint64 v);
+TEXT runtime·atomicstore64(SB), 7, $0
+	MOVL	4(SP), AX
+	// MOVQ and EMMS were introduced on the Pentium MMX.
+	// MOVQ 0x8(%ESP), %MM0
+	BYTE $0x0f; BYTE $0x6f; BYTE $0x44; BYTE $0x24; BYTE $0x08
+	// MOVQ %MM0, (%EAX)
+	BYTE $0x0f; BYTE $0x7f; BYTE $0x00 
+	// EMMS
+	BYTE $0x0F; BYTE $0x77
+	// This is essentially a no-op, but it provides required memory fencing.
+	// It can be replaced with MFENCE, but MFENCE was introduced only on the Pentium4 (SSE2).
+	MOVL	$0, AX
+	LOCK
+	XADDL	AX, (SP)
+	RET
+
+TEXT runtime·prefetch(SB), 7, $0
+	MOVL    4(SP), AX
+	// PREFETCHNTA (AX)
+	BYTE $0x0f; BYTE $0x18; BYTE $0x00
+	RET
+
 // void jmpdefer(fn, sp);
 // called from deferreturn.
 // 1. pop the caller

src/pkg/runtime/asm_amd64.s

@@ -344,6 +344,30 @@ TEXT runtime·cas(SB), 7, $0
 	MOVL	$1, AX
 	RET
 
+// bool	runtime·cas64(uint64 *val, uint64 *old, uint64 new)
+// Atomically:
+//	if(*val == *old){
+//		*val = new;
+//		return 1;
+//	} else {
+//		*old = *val
+//		return 0;
+//	}
+TEXT runtime·cas64(SB), 7, $0
+	MOVQ	8(SP), BX
+	MOVQ	16(SP), BP
+	MOVQ	0(BP), AX
+	MOVQ	24(SP), CX
+	LOCK
+	CMPXCHGQ	CX, 0(BX)
+	JNZ	cas64_fail
+	MOVL	$1, AX
+	RET
+cas64_fail:
+	MOVQ	AX, 0(BP)
+	MOVL	$0, AX
+	RET
+
 // bool casp(void **val, void *old, void *new)
 // Atomically:
 //	if(*val == old){
@@ -376,6 +400,15 @@ TEXT runtime·xadd(SB), 7, $0
 	ADDL	CX, AX
 	RET
 
+TEXT runtime·xadd64(SB), 7, $0
+	MOVQ	8(SP), BX
+	MOVQ	16(SP), AX
+	MOVQ	AX, CX
+	LOCK
+	XADDQ	AX, 0(BX)
+	ADDQ	CX, AX
+	RET
+
 TEXT runtime·xchg(SB), 7, $0
 	MOVQ	8(SP), BX
 	MOVL	16(SP), AX
@@ -402,6 +435,18 @@ TEXT runtime·atomicstore(SB), 7, $0
 	XCHGL	AX, 0(BX)
 	RET
 
+TEXT runtime·atomicstore64(SB), 7, $0
+	MOVQ	8(SP), BX
+	MOVQ	16(SP), AX
+	XCHGQ	AX, 0(BX)
+	RET
+
+TEXT runtime·prefetch(SB), 7, $0
+	MOVQ    8(SP), AX
+	// PREFETCHNTA (AX)
+	BYTE $0x0f; BYTE $0x18; BYTE $0x00
+	RET
+
 // void jmpdefer(fn, sp);
 // called from deferreturn.
 // 1. pop the caller

src/pkg/runtime/atomic_386.c

@@ -17,3 +17,16 @@ runtime·atomicloadp(void* volatile* addr)
 {
 	return *addr;
 }
+
+#pragma textflag 7
+uint64
+runtime·xadd64(uint64 volatile* addr, int64 v)
+{
+	uint64 old;
+
+	old = *addr;
+	while(!runtime·cas64(addr, &old, old+v)) {
+		// nothing
+	}
+	return old+v;
+}

src/pkg/runtime/atomic_amd64.c

@@ -11,6 +11,13 @@ runtime·atomicload(uint32 volatile* addr)
 	return *addr;
 }
 
+#pragma textflag 7
+uint64
+runtime·atomicload64(uint64 volatile* addr)
+{
+	return *addr;
+}
+
 #pragma textflag 7
 void*
 runtime·atomicloadp(void* volatile* addr)

src/pkg/runtime/atomic_arm.c

@@ -3,6 +3,14 @@
 // license that can be found in the LICENSE file.
 
 #include "runtime.h"
+#include "arch_GOARCH.h"
+
+static union {
+	Lock l;
+	byte pad [CacheLineSize];
+} locktab[57];
+
+#define LOCK(addr) (&locktab[((uintptr)(addr)>>3)%nelem(locktab)].l)
 
 // Atomic add and return new value.
 #pragma textflag 7
@@ -81,3 +89,55 @@ runtime·atomicstore(uint32 volatile* addr, uint32 v)
 			return;
 	}
 }
+
+#pragma textflag 7
+bool
+runtime·cas64(uint64 volatile *addr, uint64 *old, uint64 new)
+{
+	bool res;
+	
+	runtime·lock(LOCK(addr));
+	if(*addr == *old) {
+		*addr = new;
+		res = true;
+	} else {
+		*old = *addr;
+		res = false;
+	}
+	runtime·unlock(LOCK(addr));
+	return res;
+}
+
+#pragma textflag 7
+uint64
+runtime·xadd64(uint64 volatile *addr, int64 delta)
+{
+	uint64 res;
+	
+	runtime·lock(LOCK(addr));
+	res = *addr + delta;
+	*addr = res;
+	runtime·unlock(LOCK(addr));
+	return res;
+}
+
+#pragma textflag 7
+uint64
+runtime·atomicload64(uint64 volatile *addr)
+{
+	uint64 res;
+	
+	runtime·lock(LOCK(addr));
+	res = *addr;
+	runtime·unlock(LOCK(addr));
+	return res;
+}
+
+#pragma textflag 7
+void
+runtime·atomicstore64(uint64 volatile *addr, uint64 v)
+{
+	runtime·lock(LOCK(addr));
+	*addr = v;
+	runtime·unlock(LOCK(addr));
+}

src/pkg/runtime/runtime.c

@@ -4,6 +4,7 @@
 
 #include "runtime.h"
 #include "stack.h"
+#include "arch_GOARCH.h"
 
 enum {
 	maxround = sizeof(uintptr),
@@ -267,6 +268,33 @@ runtime·atoi(byte *p)
 	return n;
 }
 
+static void
+TestAtomic64(void)
+{
+	uint64 z64, x64;
+
+	z64 = 42;
+	x64 = 0;
+	PREFETCH(&z64);
+	if(runtime·cas64(&z64, &x64, 1))
+		runtime·throw("cas64 failed");
+	if(x64 != 42)
+		runtime·throw("cas64 failed");
+	if(!runtime·cas64(&z64, &x64, 1))
+		runtime·throw("cas64 failed");
+	if(x64 != 42 || z64 != 1)
+		runtime·throw("cas64 failed");
+	if(runtime·atomicload64(&z64) != 1)
+		runtime·throw("load64 failed");
+	runtime·atomicstore64(&z64, (1ull<<40)+1);
+	if(runtime·atomicload64(&z64) != (1ull<<40)+1)
+		runtime·throw("store64 failed");
+	if(runtime·xadd64(&z64, (1ull<<40)+1) != (2ull<<40)+2)
+		runtime·throw("xadd64 failed");
+	if(runtime·atomicload64(&z64) != (2ull<<40)+2)
+		runtime·throw("xadd64 failed");
+}
+
 void
 runtime·check(void)
 {
@@ -342,6 +370,8 @@ runtime·check(void)
 		runtime·throw("float32nan2");
 	if(!(i != i1))
 		runtime·throw("float32nan3");
+
+	TestAtomic64();
 }
 
 void

src/pkg/runtime/runtime.h

@@ -512,13 +512,17 @@ void	runtime·tracebackothers(G*);
 int32	runtime·write(int32, void*, int32);
 int32	runtime·mincore(void*, uintptr, byte*);
 bool	runtime·cas(uint32*, uint32, uint32);
+bool	runtime·cas64(uint64*, uint64*, uint64);
 bool	runtime·casp(void**, void*, void*);
 // Don't confuse with XADD x86 instruction,
 // this one is actually 'addx', that is, add-and-fetch.
 uint32	runtime·xadd(uint32 volatile*, int32);
+uint64	runtime·xadd64(uint64 volatile*, int64);
 uint32	runtime·xchg(uint32 volatile*, uint32);
 uint32	runtime·atomicload(uint32 volatile*);
 void	runtime·atomicstore(uint32 volatile*, uint32);
+void	runtime·atomicstore64(uint64 volatile*, uint64);
+uint64	runtime·atomicload64(uint64 volatile*);
 void*	runtime·atomicloadp(void* volatile*);
 void	runtime·atomicstorep(void* volatile*, void*);
 void	runtime·jmpdefer(byte*, void*);