[dev.ssa] cmd/compile: implement GO386=387

Last part of the 386 SSA port.

Modify the x86 backend to simulate SSE registers and
instructions with 387 registers and instructions.
The simulation isn't terribly performant, but it works,
and the old implementation wasn't very performant either.
Leaving to people who care about 387 to optimize if they want.

Turn on SSA backend for 386 by default.

Fixes #16358

Change-Id: I678fb59132620b2c47e993c1c10c4c21135f70c0
Reviewed-on: https://go-review.googlesource.com/25271
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>

src/cmd/compile/internal/gc/ssa.go

@@ -26,6 +26,9 @@ func initssa() *ssa.Config {
 	ssaExp.mustImplement = true
 	if ssaConfig == nil {
 		ssaConfig = ssa.NewConfig(Thearch.LinkArch.Name, &ssaExp, Ctxt, Debug['N'] == 0)
+		if Thearch.LinkArch.Name == "386" {
+			ssaConfig.Set387(Thearch.Use387)
+		}
 	}
 	return ssaConfig
 }
@@ -37,7 +40,7 @@ func shouldssa(fn *Node) bool {
 		if os.Getenv("SSATEST") == "" {
 			return false
 		}
-	case "amd64", "amd64p32", "arm":
+	case "amd64", "amd64p32", "arm", "386":
 		// Generally available.
 	}
 	if !ssaEnabled {
@@ -3948,6 +3951,10 @@ type SSAGenState struct {
 
 	// bstart remembers where each block starts (indexed by block ID)
 	bstart []*obj.Prog
+
+	// 387 port: maps from SSE registers (REG_X?) to 387 registers (REG_F?)
+	SSEto387   map[int16]int16
+	Scratch387 *Node
 }
 
 // Pc returns the current Prog.
@@ -3984,6 +3991,11 @@ func genssa(f *ssa.Func, ptxt *obj.Prog, gcargs, gclocals *Sym) {
 		blockProgs[Pc] = f.Blocks[0]
 	}
 
+	if Thearch.Use387 {
+		s.SSEto387 = map[int16]int16{}
+		s.Scratch387 = temp(Types[TUINT64])
+	}
+
 	// Emit basic blocks
 	for i, b := range f.Blocks {
 		s.bstart[b.ID] = Pc

src/cmd/compile/internal/ssa/config.go

@@ -30,6 +30,7 @@ type Config struct {
 	optimize        bool                       // Do optimization
 	noDuffDevice    bool                       // Don't use Duff's device
 	nacl            bool                       // GOOS=nacl
+	use387          bool                       // GO386=387
 	sparsePhiCutoff uint64                     // Sparse phi location algorithm used above this #blocks*#variables score
 	curFunc         *Func
 
@@ -243,6 +244,10 @@ func NewConfig(arch string, fe Frontend, ctxt *obj.Link, optimize bool) *Config
 	return c
 }
 
+func (c *Config) Set387(b bool) {
+	c.use387 = b
+}
+
 func (c *Config) Frontend() Frontend      { return c.fe }
 func (c *Config) SparsePhiCutoff() uint64 { return c.sparsePhiCutoff }
 

src/cmd/compile/internal/ssa/gen/386Ops.go

@@ -49,6 +49,17 @@ var regNames386 = []string{
 	"SB",
 }
 
+// Notes on 387 support.
+//  - The 387 has a weird stack-register setup for floating-point registers.
+//    We use these registers when SSE registers are not available (when GO386=387).
+//  - We use the same register names (X0-X7) but they refer to the 387
+//    floating-point registers. That way, most of the SSA backend is unchanged.
+//  - The instruction generation pass maintains an SSE->387 register mapping.
+//    This mapping is updated whenever the FP stack is pushed or popped so that
+//    we can always find a given SSE register even when the TOS pointer has changed.
+//  - To facilitate the mapping from SSE to 387, we enforce that
+//    every basic block starts and ends with an empty floating-point stack.
+
 func init() {
 	// Make map from reg names to reg integers.
 	if len(regNames386) > 64 {

src/cmd/compile/internal/ssa/regalloc.go

@@ -507,6 +507,9 @@ func (s *regAllocState) init(f *Func) {
 			s.allocatable &^= 1 << 15 // R15 - reserved for nacl
 		}
 	}
+	if s.f.Config.use387 {
+		s.allocatable &^= 1 << 15 // X7 disallowed (one 387 register is used as scratch space during SSE->387 generation in ../x86/387.go)
+	}
 
 	s.regs = make([]regState, s.numRegs)
 	s.values = make([]valState, f.NumValues())
@@ -834,6 +837,9 @@ func (s *regAllocState) regalloc(f *Func) {
 				if phiRegs[i] != noRegister {
 					continue
 				}
+				if s.f.Config.use387 && v.Type.IsFloat() {
+					continue // 387 can't handle floats in registers between blocks
+				}
 				m := s.compatRegs(v.Type) &^ phiUsed &^ s.used
 				if m != 0 {
 					r := pickReg(m)
@@ -1300,6 +1306,11 @@ func (s *regAllocState) regalloc(f *Func) {
 			s.freeUseRecords = u
 		}
 
+		// Spill any values that can't live across basic block boundaries.
+		if s.f.Config.use387 {
+			s.freeRegs(s.f.Config.fpRegMask)
+		}
+
 		// If we are approaching a merge point and we are the primary
 		// predecessor of it, find live values that we use soon after
 		// the merge point and promote them to registers now.
@@ -1323,6 +1334,9 @@ func (s *regAllocState) regalloc(f *Func) {
 					continue
 				}
 				v := s.orig[vid]
+				if s.f.Config.use387 && v.Type.IsFloat() {
+					continue // 387 can't handle floats in registers between blocks
+				}
 				m := s.compatRegs(v.Type) &^ s.used
 				if m&^desired.avoid != 0 {
 					m &^= desired.avoid

src/cmd/compile/internal/x86/387.go

+// Copyright 2016 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 x86
+
+import (
+	"cmd/compile/internal/gc"
+	"cmd/compile/internal/ssa"
+	"cmd/internal/obj"
+	"cmd/internal/obj/x86"
+	"math"
+)
+
+// Generates code for v using 387 instructions.  Reports whether
+// the instruction was handled by this routine.
+func ssaGenValue387(s *gc.SSAGenState, v *ssa.Value) bool {
+	// The SSA compiler pretends that it has an SSE backend.
+	// If we don't have one of those, we need to translate
+	// all the SSE ops to equivalent 387 ops. That's what this
+	// function does.
+
+	switch v.Op {
+	case ssa.Op386MOVSSconst, ssa.Op386MOVSDconst:
+		p := gc.Prog(loadPush(v.Type))
+		p.From.Type = obj.TYPE_FCONST
+		p.From.Val = math.Float64frombits(uint64(v.AuxInt))
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = x86.REG_F0
+		popAndSave(s, v)
+		return true
+
+	case ssa.Op386MOVSSload, ssa.Op386MOVSDload, ssa.Op386MOVSSloadidx1, ssa.Op386MOVSDloadidx1, ssa.Op386MOVSSloadidx4, ssa.Op386MOVSDloadidx8:
+		p := gc.Prog(loadPush(v.Type))
+		p.From.Type = obj.TYPE_MEM
+		p.From.Reg = gc.SSARegNum(v.Args[0])
+		gc.AddAux(&p.From, v)
+		switch v.Op {
+		case ssa.Op386MOVSSloadidx1, ssa.Op386MOVSDloadidx1:
+			p.From.Scale = 1
+			p.From.Index = gc.SSARegNum(v.Args[1])
+		case ssa.Op386MOVSSloadidx4:
+			p.From.Scale = 4
+			p.From.Index = gc.SSARegNum(v.Args[1])
+		case ssa.Op386MOVSDloadidx8:
+			p.From.Scale = 8
+			p.From.Index = gc.SSARegNum(v.Args[1])
+		}
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = x86.REG_F0
+		popAndSave(s, v)
+		return true
+
+	case ssa.Op386MOVSSstore, ssa.Op386MOVSDstore:
+		// Push to-be-stored value on top of stack.
+		push(s, v.Args[1])
+
+		// Pop and store value.
+		var op obj.As
+		switch v.Op {
+		case ssa.Op386MOVSSstore:
+			op = x86.AFMOVFP
+		case ssa.Op386MOVSDstore:
+			op = x86.AFMOVDP
+		}
+		p := gc.Prog(op)
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = x86.REG_F0
+		p.To.Type = obj.TYPE_MEM
+		p.To.Reg = gc.SSARegNum(v.Args[0])
+		gc.AddAux(&p.To, v)
+		return true
+
+	case ssa.Op386MOVSSstoreidx1, ssa.Op386MOVSDstoreidx1, ssa.Op386MOVSSstoreidx4, ssa.Op386MOVSDstoreidx8:
+		push(s, v.Args[2])
+		var op obj.As
+		switch v.Op {
+		case ssa.Op386MOVSSstoreidx1, ssa.Op386MOVSSstoreidx4:
+			op = x86.AFMOVFP
+		case ssa.Op386MOVSDstoreidx1, ssa.Op386MOVSDstoreidx8:
+			op = x86.AFMOVDP
+		}
+		p := gc.Prog(op)
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = x86.REG_F0
+		p.To.Type = obj.TYPE_MEM
+		p.To.Reg = gc.SSARegNum(v.Args[0])
+		gc.AddAux(&p.To, v)
+		switch v.Op {
+		case ssa.Op386MOVSSstoreidx1, ssa.Op386MOVSDstoreidx1:
+			p.To.Scale = 1
+			p.To.Index = gc.SSARegNum(v.Args[1])
+		case ssa.Op386MOVSSstoreidx4:
+			p.To.Scale = 4
+			p.To.Index = gc.SSARegNum(v.Args[1])
+		case ssa.Op386MOVSDstoreidx8:
+			p.To.Scale = 8
+			p.To.Index = gc.SSARegNum(v.Args[1])
+		}
+		return true
+
+	case ssa.Op386ADDSS, ssa.Op386ADDSD, ssa.Op386SUBSS, ssa.Op386SUBSD,
+		ssa.Op386MULSS, ssa.Op386MULSD, ssa.Op386DIVSS, ssa.Op386DIVSD:
+		if gc.SSARegNum(v) != gc.SSARegNum(v.Args[0]) {
+			v.Fatalf("input[0] and output not in same register %s", v.LongString())
+		}
+
+		// Push arg1 on top of stack
+		push(s, v.Args[1])
+
+		// Set precision if needed.  64 bits is the default.
+		switch v.Op {
+		case ssa.Op386ADDSS, ssa.Op386SUBSS, ssa.Op386MULSS, ssa.Op386DIVSS:
+			p := gc.Prog(x86.AFSTCW)
+			scratch387(s, &p.To)
+			p = gc.Prog(x86.AFLDCW)
+			p.From.Type = obj.TYPE_MEM
+			p.From.Name = obj.NAME_EXTERN
+			p.From.Sym = gc.Linksym(gc.Pkglookup("controlWord32", gc.Runtimepkg))
+		}
+
+		var op obj.As
+		switch v.Op {
+		case ssa.Op386ADDSS, ssa.Op386ADDSD:
+			op = x86.AFADDDP
+		case ssa.Op386SUBSS, ssa.Op386SUBSD:
+			op = x86.AFSUBDP
+		case ssa.Op386MULSS, ssa.Op386MULSD:
+			op = x86.AFMULDP
+		case ssa.Op386DIVSS, ssa.Op386DIVSD:
+			op = x86.AFDIVDP
+		}
+		p := gc.Prog(op)
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = x86.REG_F0
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = s.SSEto387[gc.SSARegNum(v)] + 1
+
+		// Restore precision if needed.
+		switch v.Op {
+		case ssa.Op386ADDSS, ssa.Op386SUBSS, ssa.Op386MULSS, ssa.Op386DIVSS:
+			p := gc.Prog(x86.AFLDCW)
+			scratch387(s, &p.From)
+		}
+
+		return true
+
+	case ssa.Op386UCOMISS, ssa.Op386UCOMISD:
+		push(s, v.Args[0])
+
+		// Compare.
+		p := gc.Prog(x86.AFUCOMP)
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = x86.REG_F0
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = s.SSEto387[gc.SSARegNum(v.Args[1])] + 1
+
+		// Save AX.
+		p = gc.Prog(x86.AMOVL)
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = x86.REG_AX
+		scratch387(s, &p.To)
+
+		// Move status word into AX.
+		p = gc.Prog(x86.AFSTSW)
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = x86.REG_AX
+
+		// Then move the flags we need to the integer flags.
+		gc.Prog(x86.ASAHF)
+
+		// Restore AX.
+		p = gc.Prog(x86.AMOVL)
+		scratch387(s, &p.From)
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = x86.REG_AX
+
+		return true
+
+	case ssa.Op386SQRTSD:
+		push(s, v.Args[0])
+		gc.Prog(x86.AFSQRT)
+		popAndSave(s, v)
+		return true
+
+	case ssa.Op386PXOR:
+		a0 := v.Args[0]
+		a1 := v.Args[1]
+		for a0.Op == ssa.OpCopy {
+			a0 = a0.Args[0]
+		}
+		for a1.Op == ssa.OpCopy {
+			a1 = a1.Args[0]
+		}
+		if (a0.Op == ssa.Op386MOVSSconst || a0.Op == ssa.Op386MOVSDconst) && a0.AuxInt == -0x8000000000000000 {
+			push(s, v.Args[1])
+			gc.Prog(x86.AFCHS)
+			popAndSave(s, v)
+			return true
+		}
+		if (a1.Op == ssa.Op386MOVSSconst || a1.Op == ssa.Op386MOVSDconst) && a1.AuxInt == -0x8000000000000000 {
+			push(s, v.Args[0])
+			gc.Prog(x86.AFCHS)
+			popAndSave(s, v)
+			return true
+		}
+		v.Fatalf("PXOR not used to change sign %s", v.LongString())
+
+	case ssa.Op386CVTSL2SS, ssa.Op386CVTSL2SD:
+		p := gc.Prog(x86.AMOVL)
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = gc.SSARegNum(v.Args[0])
+		scratch387(s, &p.To)
+		p = gc.Prog(x86.AFMOVL)
+		scratch387(s, &p.From)
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = x86.REG_F0
+		popAndSave(s, v)
+		return true
+
+	case ssa.Op386CVTTSD2SL, ssa.Op386CVTTSS2SL:
+		push(s, v.Args[0])
+
+		// Save control word.
+		p := gc.Prog(x86.AFSTCW)
+		scratch387(s, &p.To)
+		p.To.Offset += 4
+
+		// Load control word which truncates (rounds towards zero).
+		p = gc.Prog(x86.AFLDCW)
+		p.From.Type = obj.TYPE_MEM
+		p.From.Name = obj.NAME_EXTERN
+		p.From.Sym = gc.Linksym(gc.Pkglookup("controlWord64trunc", gc.Runtimepkg))
+
+		// Now do the conversion.
+		p = gc.Prog(x86.AFMOVLP)
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = x86.REG_F0
+		scratch387(s, &p.To)
+		p = gc.Prog(x86.AMOVL)
+		scratch387(s, &p.From)
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = gc.SSARegNum(v)
+
+		// Restore control word.
+		p = gc.Prog(x86.AFLDCW)
+		scratch387(s, &p.From)
+		p.From.Offset += 4
+		return true
+
+	case ssa.Op386CVTSS2SD:
+		// float32 -> float64 is a nop
+		push(s, v.Args[0])
+		popAndSave(s, v)
+		return true
+
+	case ssa.Op386CVTSD2SS:
+		// Round to nearest float32.
+		push(s, v.Args[0])
+		p := gc.Prog(x86.AFMOVFP)
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = x86.REG_F0
+		scratch387(s, &p.To)
+		p = gc.Prog(x86.AFMOVF)
+		scratch387(s, &p.From)
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = x86.REG_F0
+		popAndSave(s, v)
+		return true
+
+	case ssa.OpLoadReg:
+		if !v.Type.IsFloat() {
+			return false
+		}
+		// Load+push the value we need.
+		p := gc.Prog(loadPush(v.Type))
+		n, off := gc.AutoVar(v.Args[0])
+		p.From.Type = obj.TYPE_MEM
+		p.From.Node = n
+		p.From.Sym = gc.Linksym(n.Sym)
+		p.From.Offset = off
+		if n.Class == gc.PPARAM || n.Class == gc.PPARAMOUT {
+			p.From.Name = obj.NAME_PARAM
+			p.From.Offset += n.Xoffset
+		} else {
+			p.From.Name = obj.NAME_AUTO
+		}
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = x86.REG_F0
+		// Move the value to its assigned register.
+		popAndSave(s, v)
+		return true
+
+	case ssa.OpStoreReg:
+		if !v.Type.IsFloat() {
+			return false
+		}
+		push(s, v.Args[0])
+		var op obj.As
+		switch v.Type.Size() {
+		case 4:
+			op = x86.AFMOVFP
+		case 8:
+			op = x86.AFMOVDP
+		}
+		p := gc.Prog(op)
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = x86.REG_F0
+		n, off := gc.AutoVar(v)
+		p.To.Type = obj.TYPE_MEM
+		p.To.Node = n
+		p.To.Sym = gc.Linksym(n.Sym)
+		p.To.Offset = off
+		if n.Class == gc.PPARAM || n.Class == gc.PPARAMOUT {
+			p.To.Name = obj.NAME_PARAM
+			p.To.Offset += n.Xoffset
+		} else {
+			p.To.Name = obj.NAME_AUTO
+		}
+		return true
+
+	case ssa.OpCopy:
+		if !v.Type.IsFloat() {
+			return false
+		}
+		push(s, v.Args[0])
+		popAndSave(s, v)
+		return true
+
+	case ssa.Op386CALLstatic, ssa.Op386CALLclosure, ssa.Op386CALLdefer, ssa.Op386CALLgo, ssa.Op386CALLinter:
+		flush387(s)  // Calls must empty the the FP stack.
+		return false // then issue the call as normal
+	}
+	return false
+}
+
+// push pushes v onto the floating-point stack.  v must be in a register.
+func push(s *gc.SSAGenState, v *ssa.Value) {
+	p := gc.Prog(x86.AFMOVD)
+	p.From.Type = obj.TYPE_REG
+	p.From.Reg = s.SSEto387[gc.SSARegNum(v)]
+	p.To.Type = obj.TYPE_REG
+	p.To.Reg = x86.REG_F0
+}
+
+// popAndSave pops a value off of the floating-point stack and stores
+// it in the reigster assigned to v.
+func popAndSave(s *gc.SSAGenState, v *ssa.Value) {
+	r := gc.SSARegNum(v)
+	if _, ok := s.SSEto387[r]; ok {
+		// Pop value, write to correct register.
+		p := gc.Prog(x86.AFMOVDP)
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = x86.REG_F0
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = s.SSEto387[gc.SSARegNum(v)] + 1
+	} else {
+		// Don't actually pop value. This 387 register is now the
+		// new home for the not-yet-assigned-a-home SSE register.
+		// Increase the register mapping of all other registers by one.
+		for rSSE, r387 := range s.SSEto387 {
+			s.SSEto387[rSSE] = r387 + 1
+		}
+		s.SSEto387[r] = x86.REG_F0
+	}
+}
+
+// loadPush returns the opcode for load+push of the given type.
+func loadPush(t ssa.Type) obj.As {
+	if t.Size() == 4 {
+		return x86.AFMOVF
+	}
+	return x86.AFMOVD
+}
+
+// flush387 removes all entries from the 387 floating-point stack.
+func flush387(s *gc.SSAGenState) {
+	for k := range s.SSEto387 {
+		p := gc.Prog(x86.AFMOVDP)
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = x86.REG_F0
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = x86.REG_F0
+		delete(s.SSEto387, k)
+	}
+}
+
+// scratch387 initializes a to the scratch location used by some 387 rewrites.
+func scratch387(s *gc.SSAGenState, a *obj.Addr) {
+	a.Type = obj.TYPE_MEM
+	a.Name = obj.NAME_AUTO
+	a.Node = s.Scratch387
+	a.Sym = gc.Linksym(s.Scratch387.Sym)
+	a.Reg = x86.REG_SP
+}

src/cmd/compile/internal/x86/ssa.go

@@ -139,6 +139,13 @@ func opregreg(op obj.As, dest, src int16) *obj.Prog {
 
 func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 	s.SetLineno(v.Line)
+
+	if gc.Thearch.Use387 {
+		if ssaGenValue387(s, v) {
+			return // v was handled by 387 generation.
+		}
+	}
+
 	switch v.Op {
 	case ssa.Op386ADDL:
 		r := gc.SSARegNum(v)
@@ -899,6 +906,11 @@ var nefJumps = [2][2]gc.FloatingEQNEJump{
 func ssaGenBlock(s *gc.SSAGenState, b, next *ssa.Block) {
 	s.SetLineno(b.Line)
 
+	if gc.Thearch.Use387 {
+		// Empty the 387's FP stack before the block ends.
+		flush387(s)
+	}
+
 	switch b.Kind {
 	case ssa.BlockPlain, ssa.BlockCall, ssa.BlockCheck:
 		if b.Succs[0].Block() != next {

src/runtime/asm_386.s

@@ -193,9 +193,7 @@ TEXT runtime·asminit(SB),NOSPLIT,$0-0
 	// 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
+	FLDCW	runtime·controlWord64(SB)
 	RET
 
 /*
@@ -1638,47 +1636,20 @@ TEXT runtime·addmoduledata(SB),NOSPLIT,$0-0
        MOVL    AX, runtime·lastmoduledatap(SB)
        RET
 
-TEXT runtime·uint32tofloat64(SB),NOSPLIT,$0-12
-	// TODO: condition on GO386 env var.
+TEXT runtime·uint32tofloat64(SB),NOSPLIT,$8-12
 	MOVL	a+0(FP), AX
-
-	// Check size.
-	CMPL	AX, $0x80000000
-	JAE	large
-
-	// Less than 2**31, convert directly.
-	CVTSL2SD	AX, X0
-	MOVSD	X0, ret+4(FP)
-	RET
-large:
-	// >= 2**31.  Subtract 2**31 (uint32), convert, then add 2**31 (float64).
-	SUBL	$0x80000000, AX
-	CVTSL2SD	AX, X0
-	ADDSD	twotothe31<>(SB), X0
-	MOVSD	X0, ret+4(FP)
+	MOVL	AX, 0(SP)
+	MOVL	$0, 4(SP)
+	FMOVV	0(SP), F0
+	FMOVDP	F0, ret+4(FP)
 	RET
 
-TEXT runtime·float64touint32(SB),NOSPLIT,$0-12
-	// TODO: condition on GO386 env var.
-	MOVSD	a+0(FP), X0
-
-	// Check size.
-	MOVSD	twotothe31<>(SB), X1
-	UCOMISD	X1, X0 //note: args swapped relative to CMPL
-	JAE	large
-
-	// Less than 2**31, convert directly.
-	CVTTSD2SL X0, AX
-	MOVL	AX, ret+8(FP)
-	RET
-large:
-	// >= 2**31.  Subtract 2**31 (float64), convert, then add 2**31 (uint32).
-	SUBSD	X1, X0
-	CVTTSD2SL	X0, AX
-	ADDL	$0x80000000, AX
+TEXT runtime·float64touint32(SB),NOSPLIT,$12-12
+	FMOVD	a+0(FP), F0
+	FSTCW	0(SP)
+	FLDCW	runtime·controlWord64trunc(SB)
+	FMOVVP	F0, 4(SP)
+	FLDCW	0(SP)
+	MOVL	4(SP), AX
 	MOVL	AX, ret+8(FP)
 	RET
-
-// 2**31 as a float64.
-DATA	twotothe31<>+0x00(SB)/8, $0x41e0000000000000
-GLOBL	twotothe31<>(SB),RODATA,$8

src/runtime/vlrt.go

@@ -255,3 +255,17 @@ func slowdodiv(n, d uint64) (q, r uint64) {
 	}
 	return q, n
 }
+
+// Floating point control word values for GOARCH=386 GO386=387.
+// Bits 0-5 are bits to disable floating-point exceptions.
+// Bits 8-9 are the precision control:
+//   0 = single precision a.k.a. float32
+//   2 = double precision a.k.a. float64
+// Bits 10-11 are the rounding mode:
+//   0 = round to nearest (even on a tie)
+//   3 = round toward zero
+var (
+	controlWord64      uint16 = 0x3f + 2<<8 + 0<<10
+	controlWord32             = 0x3f + 0<<8 + 0<<10
+	controlWord64trunc        = 0x3f + 2<<8 + 3<<10
+)

test/live.go

-// +build !amd64,!arm,!amd64p32
+// +build !amd64,!arm,!amd64p32,!386
 // errorcheck -0 -l -live -wb=0
 
 // Copyright 2014 The Go Authors. All rights reserved.

test/live_ssa.go

-// +build amd64 arm amd64p32
+// +build amd64 arm amd64p32 386
 // errorcheck -0 -l -live -wb=0
 
 // Copyright 2014 The Go Authors. All rights reserved.

test/nilptr3.go

@@ -2,7 +2,7 @@
 // Fails on ppc64x because of incomplete optimization.
 // See issues 9058.
 // Same reason for mips64x and s390x.
-// +build !ppc64,!ppc64le,!mips64,!mips64le,!amd64,!s390x,!arm,!amd64p32
+// +build !ppc64,!ppc64le,!mips64,!mips64le,!amd64,!s390x,!arm,!amd64p32,!386
 
 // Copyright 2013 The Go Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style

test/nilptr3_ssa.go

 // errorcheck -0 -d=nil
-// +build amd64 arm amd64p32
+// +build amd64 arm amd64p32 386
 
 // Copyright 2013 The Go Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style

test/sliceopt.go

-// +build !amd64,!arm,!amd64p32
+// +build !amd64,!arm,!amd64p32,!386
 // errorcheck -0 -d=append,slice
 
 // Copyright 2015 The Go Authors. All rights reserved.