all: replace runtime SSE2 detection with GO386 setting

When GO386=sse2 we can assume sse2 to be present without
a runtime check. If GO386=softfloat is set we can avoid
the usage of SSE2 even if detected.

This might cause a memcpy, memclr and bytealg slowdown of Go
binaries compiled with softfloat on machines that support
SSE2. Such setups are rare and should use GO386=sse2 instead
if performance matters.

On targets that support SSE2 we avoid the runtime overhead of
dynamic cpu feature dispatch.

The removal of runtime sse2 checks also allows to simplify
internal/cpu further by removing handling of the required
feature option as a followup after this CL.

Change-Id: I90a853a8853a405cb665497c6d1a86556947ba17
Reviewed-on: https://go-review.googlesource.com/c/go/+/344350
Trust: Martin Möhrmann <martin@golang.org>
Run-TryBot: Martin Möhrmann <martin@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>

15 files changed, 31 insertions, 60 deletions

diff --git a/src/cmd/go/internal/work/gc.go b/src/cmd/go/internal/work/gc.go
index 74e14d0065..eee8adca94 100644
--- a/src/cmd/go/internal/work/gc.go
+++ b/src/cmd/go/internal/work/gc.go
@@ -374,6 +374,11 @@ func asmArgs(a *Action, p *load.Package) []interface{} {
 		args = append(args, "-compiling-runtime")
 	}
 
+	if cfg.Goarch == "386" {
+		// Define GO386_value from cfg.GO386.
+		args = append(args, "-D", "GO386_"+cfg.GO386)
+	}
+
 	if cfg.Goarch == "mips" || cfg.Goarch == "mipsle" {
 		// Define GOMIPS_value from cfg.GOMIPS.
 		args = append(args, "-D", "GOMIPS_"+cfg.GOMIPS)
diff --git a/src/internal/bytealg/bytealg.go b/src/internal/bytealg/bytealg.go
index 6b2b540acc..ebebce75fe 100644
--- a/src/internal/bytealg/bytealg.go
+++ b/src/internal/bytealg/bytealg.go
@@ -11,7 +11,6 @@ import (
 
 // Offsets into internal/cpu records for use in assembly.
 const (
-	offsetX86HasSSE2   = unsafe.Offsetof(cpu.X86.HasSSE2)
 	offsetX86HasSSE42  = unsafe.Offsetof(cpu.X86.HasSSE42)
 	offsetX86HasAVX2   = unsafe.Offsetof(cpu.X86.HasAVX2)
 	offsetX86HasPOPCNT = unsafe.Offsetof(cpu.X86.HasPOPCNT)
diff --git a/src/internal/bytealg/compare_386.s b/src/internal/bytealg/compare_386.s
index 0981983d20..27b660ccf7 100644
--- a/src/internal/bytealg/compare_386.s
+++ b/src/internal/bytealg/compare_386.s
@@ -36,8 +36,9 @@ TEXT cmpbody<>(SB),NOSPLIT,$0-0
 	JEQ	allsame
 	CMPL	BP, $4
 	JB	small
-	CMPB	internal∕cpu·X86+const_offsetX86HasSSE2(SB), $1
-	JNE	mediumloop
+#ifdef GO386_softfloat
+	JMP	mediumloop
+#endif
 largeloop:
 	CMPL	BP, $16
 	JB	mediumloop
diff --git a/src/internal/bytealg/equal_386.s b/src/internal/bytealg/equal_386.s
index 87233635a9..58b3cbe3d0 100644
--- a/src/internal/bytealg/equal_386.s
+++ b/src/internal/bytealg/equal_386.s
@@ -43,8 +43,9 @@ TEXT memeqbody<>(SB),NOSPLIT,$0-0
 hugeloop:
 	CMPL	BX, $64
 	JB	bigloop
-	CMPB	internal∕cpu·X86+const_offsetX86HasSSE2(SB), $1
-	JNE	bigloop
+#ifdef GO386_softfloat
+	JMP	bigloop
+#endif
 	MOVOU	(SI), X0
 	MOVOU	(DI), X1
 	MOVOU	16(SI), X2
diff --git a/src/internal/cpu/cpu.go b/src/internal/cpu/cpu.go
index a87d8a2314..4f0c5d2896 100644
--- a/src/internal/cpu/cpu.go
+++ b/src/internal/cpu/cpu.go
@@ -37,7 +37,6 @@ var X86 struct {
 	HasPCLMULQDQ bool
 	HasPOPCNT    bool
 	HasRDTSCP    bool
-	HasSSE2      bool
 	HasSSE3      bool
 	HasSSSE3     bool
 	HasSSE41     bool
diff --git a/src/internal/cpu/cpu_386.go b/src/internal/cpu/cpu_386.go
deleted file mode 100644
index 561c81f808..0000000000
--- a/src/internal/cpu/cpu_386.go
+++ /dev/null
@@ -1,7 +0,0 @@
-// Copyright 2018 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.
-
-package cpu
-
-const GOARCH = "386"
diff --git a/src/internal/cpu/cpu_amd64.go b/src/internal/cpu/cpu_amd64.go
deleted file mode 100644
index 9b0015362d..0000000000
--- a/src/internal/cpu/cpu_amd64.go
+++ /dev/null
@@ -1,7 +0,0 @@
-// Copyright 2018 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.
-
-package cpu
-
-const GOARCH = "amd64"
diff --git a/src/internal/cpu/cpu_x86.go b/src/internal/cpu/cpu_x86.go
index a3f1fb809a..1582e832a4 100644
--- a/src/internal/cpu/cpu_x86.go
+++ b/src/internal/cpu/cpu_x86.go
@@ -61,9 +61,6 @@ func doinit() {
 		{Name: "sse41", Feature: &X86.HasSSE41},
 		{Name: "sse42", Feature: &X86.HasSSE42},
 		{Name: "ssse3", Feature: &X86.HasSSSE3},
-
-		// These capabilities should always be enabled on amd64:
-		{Name: "sse2", Feature: &X86.HasSSE2, Required: GOARCH == "amd64"},
 	}
 
 	maxID, _, _, _ := cpuid(0, 0)
@@ -74,8 +71,7 @@ func doinit() {
 
 	maxExtendedFunctionInformation, _, _, _ = cpuid(0x80000000, 0)
 
-	_, _, ecx1, edx1 := cpuid(1, 0)
-	X86.HasSSE2 = isSet(edx1, cpuid_SSE2)
+	_, _, ecx1, _ := cpuid(1, 0)
 
 	X86.HasSSE3 = isSet(ecx1, cpuid_SSE3)
 	X86.HasPCLMULQDQ = isSet(ecx1, cpuid_PCLMULQDQ)
diff --git a/src/internal/cpu/cpu_x86_test.go b/src/internal/cpu/cpu_x86_test.go
index e3e16cc161..de1c5fb3b0 100644
--- a/src/internal/cpu/cpu_x86_test.go
+++ b/src/internal/cpu/cpu_x86_test.go
@@ -10,7 +10,6 @@ package cpu_test
 import (
 	. "internal/cpu"
 	"os"
-	"runtime"
 	"testing"
 )
 
@@ -20,23 +19,6 @@ func TestX86ifAVX2hasAVX(t *testing.T) {
 	}
 }
 
-func TestDisableSSE2(t *testing.T) {
-	runDebugOptionsTest(t, "TestSSE2DebugOption", "cpu.sse2=off")
-}
-
-func TestSSE2DebugOption(t *testing.T) {
-	MustHaveDebugOptionsSupport(t)
-
-	if os.Getenv("GODEBUG") != "cpu.sse2=off" {
-		t.Skipf("skipping test: GODEBUG=cpu.sse2=off not set")
-	}
-
-	want := runtime.GOARCH != "386" // SSE2 can only be disabled on 386.
-	if got := X86.HasSSE2; got != want {
-		t.Errorf("X86.HasSSE2 on %s expected %v, got %v", runtime.GOARCH, want, got)
-	}
-}
-
 func TestDisableSSE3(t *testing.T) {
 	runDebugOptionsTest(t, "TestSSE3DebugOption", "cpu.sse3=off")
 }
diff --git a/src/runtime/asm_386.s b/src/runtime/asm_386.s
index b711356822..594cd5ed0d 100644
--- a/src/runtime/asm_386.s
+++ b/src/runtime/asm_386.s
@@ -838,8 +838,9 @@ TEXT runtime·cputicks(SB),NOSPLIT,$0-8
 	// When no SSE2 support is present do not enforce any serialization
 	// since using CPUID to serialize the instruction stream is
 	// very costly.
-	CMPB	internal∕cpu·X86+const_offsetX86HasSSE2(SB), $1
-	JNE	rdtsc
+#ifdef GO386_softfloat
+	JMP	rdtsc  // no fence instructions available
+#endif
 	CMPB	internal∕cpu·X86+const_offsetX86HasRDTSCP(SB), $1
 	JNE	fences
 	// Instruction stream serializing RDTSCP is supported.
diff --git a/src/runtime/cpuflags.go b/src/runtime/cpuflags.go
index c5291ce4ee..bbe93c5bea 100644
--- a/src/runtime/cpuflags.go
+++ b/src/runtime/cpuflags.go
@@ -15,7 +15,6 @@ const (
 	offsetX86HasAVX2   = unsafe.Offsetof(cpu.X86.HasAVX2)
 	offsetX86HasERMS   = unsafe.Offsetof(cpu.X86.HasERMS)
 	offsetX86HasRDTSCP = unsafe.Offsetof(cpu.X86.HasRDTSCP)
-	offsetX86HasSSE2   = unsafe.Offsetof(cpu.X86.HasSSE2)
 
 	offsetARMHasIDIVA = unsafe.Offsetof(cpu.ARM.HasIDIVA)
 
diff --git a/src/runtime/memclr_386.s b/src/runtime/memclr_386.s
index 046c344119..2627792ced 100644
--- a/src/runtime/memclr_386.s
+++ b/src/runtime/memclr_386.s
@@ -30,8 +30,9 @@ tail:
 	JBE	_5through8
 	CMPL	BX, $16
 	JBE	_9through16
-	CMPB	internal∕cpu·X86+const_offsetX86HasSSE2(SB), $1
-	JNE	nosse2
+#ifdef GO386_softfloat
+	JMP	nosse2
+#endif
 	PXOR	X0, X0
 	CMPL	BX, $32
 	JBE	_17through32
diff --git a/src/runtime/memmove_386.s b/src/runtime/memmove_386.s
index 1a43a1f724..389ef88477 100644
--- a/src/runtime/memmove_386.s
+++ b/src/runtime/memmove_386.s
@@ -55,8 +55,9 @@ tail:
 	JBE	move_5through8
 	CMPL	BX, $16
 	JBE	move_9through16
-	CMPB	internal∕cpu·X86+const_offsetX86HasSSE2(SB), $1
-	JNE	nosse2
+#ifdef GO386_softfloat
+	JMP	nosse2
+#endif
 	CMPL	BX, $32
 	JBE	move_17through32
 	CMPL	BX, $64
diff --git a/src/runtime/mkpreempt.go b/src/runtime/mkpreempt.go
index f2b90307ca..d87446d036 100644
--- a/src/runtime/mkpreempt.go
+++ b/src/runtime/mkpreempt.go
@@ -200,6 +200,8 @@ func gen386() {
 		l.add("MOVL", reg, 4)
 	}
 
+	softfloat := "GO386_softfloat"
+
 	// Save SSE state only if supported.
 	lSSE := layout{stack: l.stack, sp: "SP"}
 	for i := 0; i < 8; i++ {
@@ -209,13 +211,13 @@ func gen386() {
 	p("ADJSP $%d", lSSE.stack)
 	p("NOP SP")
 	l.save()
-	p("CMPB internal∕cpu·X86+const_offsetX86HasSSE2(SB), $1\nJNE nosse")
+	p("#ifndef %s", softfloat)
 	lSSE.save()
-	label("nosse:")
+	p("#endif")
 	p("CALL ·asyncPreempt2(SB)")
-	p("CMPB internal∕cpu·X86+const_offsetX86HasSSE2(SB), $1\nJNE nosse2")
+	p("#ifndef %s", softfloat)
 	lSSE.restore()
-	label("nosse2:")
+	p("#endif")
 	l.restore()
 	p("ADJSP $%d", -lSSE.stack)
 
diff --git a/src/runtime/preempt_386.s b/src/runtime/preempt_386.s
index c3a5fa1f36..d57bc3d37c 100644
--- a/src/runtime/preempt_386.s
+++ b/src/runtime/preempt_386.s
@@ -14,8 +14,7 @@ TEXT ·asyncPreempt(SB),NOSPLIT|NOFRAME,$0-0
 	MOVL BP, 16(SP)
 	MOVL SI, 20(SP)
 	MOVL DI, 24(SP)
-	CMPB internal∕cpu·X86+const_offsetX86HasSSE2(SB), $1
-	JNE nosse
+	#ifndef GO386_softfloat
 	MOVUPS X0, 28(SP)
 	MOVUPS X1, 44(SP)
 	MOVUPS X2, 60(SP)
@@ -24,10 +23,9 @@ TEXT ·asyncPreempt(SB),NOSPLIT|NOFRAME,$0-0
 	MOVUPS X5, 108(SP)
 	MOVUPS X6, 124(SP)
 	MOVUPS X7, 140(SP)
-nosse:
+	#endif
 	CALL ·asyncPreempt2(SB)
-	CMPB internal∕cpu·X86+const_offsetX86HasSSE2(SB), $1
-	JNE nosse2
+	#ifndef GO386_softfloat
 	MOVUPS 140(SP), X7
 	MOVUPS 124(SP), X6
 	MOVUPS 108(SP), X5
@@ -36,7 +34,7 @@ nosse:
 	MOVUPS 60(SP), X2
 	MOVUPS 44(SP), X1
 	MOVUPS 28(SP), X0
-nosse2:
+	#endif
 	MOVL 24(SP), DI
 	MOVL 20(SP), SI
 	MOVL 16(SP), BP