runtime: on 386, fix FP control word on all threads, not just initial thread

It is possible that Linux and Windows copy the FP control word
from the parent thread when creating a new thread.  Empirically,
Darwin does not.  Reset the FP control world in all cases.

Enable the floating-point strconv test.

Fixes #2917 (again).

R=golang-dev, r, iant
CC=golang-dev
https://golang.org/cl/5660047

src/pkg/runtime/asm_386.s

@@ -5,14 +5,6 @@
 #include "zasm_GOOS_GOARCH.h"
 
 TEXT _rt0_386(SB),7,$0
-	// Linux, Windows start the FPU in extended double precision.
-	// Other operating systems use double precision.
-	// Change to double precision to match them,
-	// and to match other hardware that only has double.
-	PUSHL $0x27F
-	FLDCW	0(SP)
-	POPL AX
-
 	// copy arguments forward on an even stack
 	MOVL	0(SP), AX		// argc
 	LEAL	4(SP), BX		// argv
@@ -99,6 +91,16 @@ TEXT runtime·breakpoint(SB),7,$0
 	INT $3
 	RET
 
+TEXT runtime·asminit(SB),7,$0
+	// Linux, Windows start the FPU in extended double precision.
+	// Other operating systems use double precision.
+	// Change to double precision to match them,
+	// and to match other hardware that only has double.
+	PUSHL $0x27F
+	FLDCW	0(SP)
+	POPL AX
+	RET
+
 /*
  *  go-routine
  */

src/pkg/runtime/asm_amd64.s

@@ -80,6 +80,10 @@ TEXT runtime·breakpoint(SB),7,$0
 	BYTE	$0xcc
 	RET
 
+TEXT runtime·asminit(SB),7,$0
+	// No per-thread init.
+	RET
+
 /*
  *  go-routine
  */

src/pkg/runtime/asm_arm.s

@@ -74,6 +74,10 @@ TEXT runtime·breakpoint(SB),7,$0
 	// no breakpoint yet; let program exit
 	RET
 
+TEXT runtime·asminit(SB),7,$0
+	// No per-thread init.
+	RET
+
 /*
  *  go-routine
  */

src/pkg/runtime/proc.c

@@ -720,6 +720,7 @@ runtime·mstart(void)
 	runtime·gosave(&m->g0->sched);
 	m->g0->sched.pc = (void*)-1;  // make sure it is never used
 
+	runtime·asminit();
 	runtime·minit();
 	schedule(nil);
 }

src/pkg/runtime/runtime.h

@@ -525,6 +525,7 @@ int32	runtime·atoi(byte*);
 void	runtime·newosproc(M *m, G *g, void *stk, void (*fn)(void));
 void	runtime·signalstack(byte*, int32);
 G*	runtime·malg(int32);
+void	runtime·asminit(void);
 void	runtime·minit(void);
 Func*	runtime·findfunc(uintptr);
 int32	runtime·funcline(Func*, uintptr);

src/pkg/strconv/atof_test.go

@@ -8,7 +8,6 @@ import (
 	"math"
 	"math/rand"
 	"reflect"
-	"runtime"
 	. "strconv"
 	"strings"
 	"testing"
@@ -232,16 +231,16 @@ var roundTripCases = []struct {
 	s string
 }{
 	// Issue 2917.
-	// A Darwin/386 builder failed on AtofRandom with this case.
+	// This test will break the optimized conversion if the
+	// FPU is using 80-bit registers instead of 64-bit registers,
+	// usually because the operating system initialized the
+	// thread with 80-bit precision and the Go runtime didn't
+	// fix the FP control word.
 	{8865794286000691 << 39, "4.87402195346389e+27"},
 	{8865794286000692 << 39, "4.8740219534638903e+27"},
 }
 
 func TestRoundTrip(t *testing.T) {
-	if runtime.GOOS == "darwin" && runtime.GOARCH == "386" {
-		t.Logf("skipping round-trip test on darwin/386 - known failure, issue 2917")
-		return
-	}
 	for _, tt := range roundTripCases {
 		old := SetOptimize(false)
 		s := FormatFloat(tt.f, 'g', -1, 64)