[dev.ssa] cmd/compile: PPC64, FP to/from int conversions.

Passes ssa_test.

Requires a few new instructions and some scratchpad
memory to move data between G and F registers.

Also fixed comparisons to be correct in case of NaN.
Added missing instructions for run.bash.
Removed some FP registers that are apparently "reserved"
(but that are also apparently also unused except for a
gratuitous multiplication by two when y = x+x would work
just as well).

Currently failing stack splits.

Updates #16010.

Change-Id: I73b161bfff54445d72bd7b813b1479f89fc72602
Reviewed-on: https://go-review.googlesource.com/26813
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>

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

@@ -3953,8 +3953,9 @@ type SSAGenState struct {
 	bstart []*obj.Prog
 
 	// 387 port: maps from SSE registers (REG_X?) to 387 registers (REG_F?)
-	SSEto387   map[int16]int16
-	Scratch387 *Node
+	SSEto387 map[int16]int16
+	// Some architectures require a 64-bit temporary for FP-related register shuffling. Examples include x86-387, PPC, and Sparc V8.
+	ScratchFpMem *Node
 }
 
 // Pc returns the current Prog.
@@ -3993,7 +3994,9 @@ func genssa(f *ssa.Func, ptxt *obj.Prog, gcargs, gclocals *Sym) {
 
 	if Thearch.Use387 {
 		s.SSEto387 = map[int16]int16{}
-		s.Scratch387 = temp(Types[TUINT64])
+	}
+	if f.Config.NeedsFpScratch {
+		s.ScratchFpMem = temp(Types[TUINT64])
 	}
 
 	// Emit basic blocks

src/cmd/compile/internal/ppc64/prog.go

@@ -82,6 +82,7 @@ var progtable = [ppc64.ALAST & obj.AMask]obj.ProgInfo{
 	ppc64.AFDIV & obj.AMask:   {Flags: gc.SizeD | gc.LeftRead | gc.RegRead | gc.RightWrite},
 	ppc64.AFDIVS & obj.AMask:  {Flags: gc.SizeF | gc.LeftRead | gc.RegRead | gc.RightWrite},
 	ppc64.AFCTIDZ & obj.AMask: {Flags: gc.SizeF | gc.LeftRead | gc.RegRead | gc.RightWrite},
+	ppc64.AFCTIWZ & obj.AMask: {Flags: gc.SizeF | gc.LeftRead | gc.RegRead | gc.RightWrite},
 	ppc64.AFCFID & obj.AMask:  {Flags: gc.SizeF | gc.LeftRead | gc.RegRead | gc.RightWrite},
 	ppc64.AFCFIDU & obj.AMask: {Flags: gc.SizeF | gc.LeftRead | gc.RegRead | gc.RightWrite},
 	ppc64.AFCMPU & obj.AMask:  {Flags: gc.SizeD | gc.LeftRead | gc.RightRead},
@@ -104,6 +105,8 @@ var progtable = [ppc64.ALAST & obj.AMask]obj.ProgInfo{
 	ppc64.AMOVD & obj.AMask:   {Flags: gc.SizeQ | gc.LeftRead | gc.RightWrite | gc.Move},
 	ppc64.AMOVDU & obj.AMask:  {Flags: gc.SizeQ | gc.LeftRead | gc.RightWrite | gc.Move | gc.PostInc},
 	ppc64.AFMOVS & obj.AMask:  {Flags: gc.SizeF | gc.LeftRead | gc.RightWrite | gc.Move | gc.Conv},
+	ppc64.AFMOVSX & obj.AMask: {Flags: gc.SizeF | gc.LeftRead | gc.RightWrite | gc.Move | gc.Conv},
+	ppc64.AFMOVSZ & obj.AMask: {Flags: gc.SizeF | gc.LeftRead | gc.RightWrite | gc.Move | gc.Conv},
 	ppc64.AFMOVD & obj.AMask:  {Flags: gc.SizeD | gc.LeftRead | gc.RightWrite | gc.Move},
 
 	// Jumps

src/cmd/compile/internal/ppc64/ssa.go

@@ -73,11 +73,11 @@ var ssaRegToReg = []int16{
 	ppc64.REG_F24,
 	ppc64.REG_F25,
 	ppc64.REG_F26,
-	ppc64.REG_F27,
-	ppc64.REG_F28,
-	ppc64.REG_F29,
-	ppc64.REG_F30,
-	ppc64.REG_F31,
+	// ppc64.REG_F27, // reserved for "floating conversion constant"
+	// ppc64.REG_F28, // 0.0
+	// ppc64.REG_F29, // 0.5
+	// ppc64.REG_F30, // 1.0
+	// ppc64.REG_F31, // 2.0
 
 	// ppc64.REG_CR0,
 	// ppc64.REG_CR1,
@@ -88,21 +88,12 @@ var ssaRegToReg = []int16{
 	// ppc64.REG_CR6,
 	// ppc64.REG_CR7,
 
-	ppc64.REG_CR,
+	// ppc64.REG_CR,
 	// ppc64.REG_XER,
 	// ppc64.REG_LR,
 	// ppc64.REG_CTR,
 }
 
-// Associated condition bit
-var condBits = map[ssa.Op]uint8{
-	ssa.OpPPC64Equal:        ppc64.C_COND_EQ,
-	ssa.OpPPC64NotEqual:     ppc64.C_COND_EQ,
-	ssa.OpPPC64LessThan:     ppc64.C_COND_LT,
-	ssa.OpPPC64GreaterEqual: ppc64.C_COND_LT,
-	ssa.OpPPC64GreaterThan:  ppc64.C_COND_GT,
-	ssa.OpPPC64LessEqual:    ppc64.C_COND_GT,
-}
 var condOps = map[ssa.Op]obj.As{
 	ssa.OpPPC64Equal:        ppc64.ABEQ,
 	ssa.OpPPC64NotEqual:     ppc64.ABNE,
@@ -110,16 +101,11 @@ var condOps = map[ssa.Op]obj.As{
 	ssa.OpPPC64GreaterEqual: ppc64.ABGE,
 	ssa.OpPPC64GreaterThan:  ppc64.ABGT,
 	ssa.OpPPC64LessEqual:    ppc64.ABLE,
-}
 
-// Is the condition bit set? 1=yes 0=no
-var condBitSet = map[ssa.Op]uint8{
-	ssa.OpPPC64Equal:        1,
-	ssa.OpPPC64NotEqual:     0,
-	ssa.OpPPC64LessThan:     1,
-	ssa.OpPPC64GreaterEqual: 0,
-	ssa.OpPPC64GreaterThan:  1,
-	ssa.OpPPC64LessEqual:    0,
+	ssa.OpPPC64FLessThan:     ppc64.ABLT, // 1 branch for FCMP
+	ssa.OpPPC64FGreaterThan:  ppc64.ABGT, // 1 branch for FCMP
+	ssa.OpPPC64FLessEqual:    ppc64.ABLT, // 2 branches for FCMP <=, second is BEQ
+	ssa.OpPPC64FGreaterEqual: ppc64.ABGT, // 2 branches for FCMP >=, second is BEQ
 }
 
 // markMoves marks any MOVXconst ops that need to avoid clobbering flags.
@@ -205,6 +191,17 @@ func storeByType(t ssa.Type) obj.As {
 	panic("bad store type")
 }
 
+// scratchFpMem initializes an Addr (field of a Prog)
+// to reference the scratchpad memory for movement between
+// F and G registers for FP conversions.
+func scratchFpMem(s *gc.SSAGenState, a *obj.Addr) {
+	a.Type = obj.TYPE_MEM
+	a.Name = obj.NAME_AUTO
+	a.Node = s.ScratchFpMem
+	a.Sym = gc.Linksym(s.ScratchFpMem.Sym)
+	a.Reg = ppc64.REGSP
+}
+
 func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 	s.SetLineno(v.Line)
 	switch v.Op {
@@ -212,22 +209,55 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 		// memory arg needs no code
 	case ssa.OpArg:
 		// input args need no code
-	case ssa.OpSP, ssa.OpSB:
+	case ssa.OpSP, ssa.OpSB, ssa.OpGetG:
 		// nothing to do
 
 	case ssa.OpCopy, ssa.OpPPC64MOVDconvert:
-		// TODO: copy of floats
-		if v.Type.IsMemory() {
+		t := v.Type
+		if t.IsMemory() {
 			return
 		}
 		x := gc.SSARegNum(v.Args[0])
 		y := gc.SSARegNum(v)
 		if x != y {
-			p := gc.Prog(ppc64.AMOVD)
+			rt := obj.TYPE_REG
+			op := ppc64.AMOVD
+
+			if t.IsFloat() {
+				op = ppc64.AFMOVD
+			}
+			p := gc.Prog(op)
+			p.From.Type = rt
+			p.From.Reg = x
+			p.To.Type = rt
+			p.To.Reg = y
+		}
+
+	case ssa.OpPPC64Xf2i64:
+		{
+			x := gc.SSARegNum(v.Args[0])
+			y := gc.SSARegNum(v)
+			p := gc.Prog(ppc64.AFMOVD)
 			p.From.Type = obj.TYPE_REG
 			p.From.Reg = x
+			scratchFpMem(s, &p.To)
+			p = gc.Prog(ppc64.AMOVD)
+			p.To.Type = obj.TYPE_REG
 			p.To.Reg = y
+			scratchFpMem(s, &p.From)
+		}
+	case ssa.OpPPC64Xi2f64:
+		{
+			x := gc.SSARegNum(v.Args[0])
+			y := gc.SSARegNum(v)
+			p := gc.Prog(ppc64.AMOVD)
+			p.From.Type = obj.TYPE_REG
+			p.From.Reg = x
+			scratchFpMem(s, &p.To)
+			p = gc.Prog(ppc64.AFMOVD)
 			p.To.Type = obj.TYPE_REG
+			p.To.Reg = y
+			scratchFpMem(s, &p.From)
 		}
 
 	case ssa.OpPPC64LoweredGetClosurePtr:
@@ -235,8 +265,9 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 		gc.CheckLoweredGetClosurePtr(v)
 
 	case ssa.OpLoadReg:
-		p := gc.Prog(loadByType(v.Type))
+		loadOp := loadByType(v.Type)
 		n, off := gc.AutoVar(v.Args[0])
+		p := gc.Prog(loadOp)
 		p.From.Type = obj.TYPE_MEM
 		p.From.Node = n
 		p.From.Sym = gc.Linksym(n.Sym)
@@ -251,10 +282,11 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 		p.To.Reg = gc.SSARegNum(v)
 
 	case ssa.OpStoreReg:
-		p := gc.Prog(storeByType(v.Type))
+		storeOp := storeByType(v.Type)
+		n, off := gc.AutoVar(v)
+		p := gc.Prog(storeOp)
 		p.From.Type = obj.TYPE_REG
 		p.From.Reg = gc.SSARegNum(v.Args[0])
-		n, off := gc.AutoVar(v)
 		p.To.Type = obj.TYPE_MEM
 		p.To.Node = n
 		p.To.Sym = gc.Linksym(n.Sym)
@@ -376,7 +408,7 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 		p.To.Type = obj.TYPE_REG
 		p.To.Reg = ppc64.REGTMP // Ignored; this is for the carry effect.
 
-	case ssa.OpPPC64NEG, ssa.OpPPC64FNEG:
+	case ssa.OpPPC64NEG, ssa.OpPPC64FNEG, ssa.OpPPC64FSQRT, ssa.OpPPC64FSQRTS, ssa.OpPPC64FCTIDZ, ssa.OpPPC64FCTIWZ, ssa.OpPPC64FCFID, ssa.OpPPC64FRSP:
 		r := gc.SSARegNum(v)
 		p := gc.Prog(v.Op.Asm())
 		p.To.Type = obj.TYPE_REG
@@ -510,8 +542,10 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 	case ssa.OpPPC64Equal,
 		ssa.OpPPC64NotEqual,
 		ssa.OpPPC64LessThan,
+		ssa.OpPPC64FLessThan,
 		ssa.OpPPC64LessEqual,
 		ssa.OpPPC64GreaterThan,
+		ssa.OpPPC64FGreaterThan,
 		ssa.OpPPC64GreaterEqual:
 		// On Power7 or later, can use isel instruction:
 		// for a < b, a > b, a = b:
@@ -549,6 +583,30 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 		p = gc.Prog(obj.ANOP)
 		gc.Patch(pb, p)
 
+	case ssa.OpPPC64FLessEqual, // These include a second branch for EQ -- dealing with NaN prevents REL= to !REL conversion
+		ssa.OpPPC64FGreaterEqual:
+
+		p := gc.Prog(ppc64.AMOVW)
+		p.From.Type = obj.TYPE_CONST
+		p.From.Offset = 1
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = gc.SSARegNum(v)
+
+		pb0 := gc.Prog(condOps[v.Op])
+		pb0.To.Type = obj.TYPE_BRANCH
+		pb1 := gc.Prog(ppc64.ABEQ)
+		pb1.To.Type = obj.TYPE_BRANCH
+
+		p = gc.Prog(ppc64.AMOVW)
+		p.From.Type = obj.TYPE_CONST
+		p.From.Offset = 0
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = gc.SSARegNum(v)
+
+		p = gc.Prog(obj.ANOP)
+		gc.Patch(pb0, p)
+		gc.Patch(pb1, p)
+
 	case ssa.OpPPC64LoweredZero:
 		// Similar to how this is done on ARM,
 		// except that PPC MOVDU x,off(y) is *(y+off) = x; y=y+off
@@ -843,20 +901,22 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 }
 
 var blockJump = [...]struct {
-	asm, invasm obj.As
+	asm, invasm     obj.As
+	asmeq, invasmeq bool
 }{
-	ssa.BlockPPC64EQ: {ppc64.ABEQ, ppc64.ABNE},
-	ssa.BlockPPC64NE: {ppc64.ABNE, ppc64.ABEQ},
-
-	ssa.BlockPPC64LT: {ppc64.ABLT, ppc64.ABGE},
-	ssa.BlockPPC64GE: {ppc64.ABGE, ppc64.ABLT},
-	ssa.BlockPPC64LE: {ppc64.ABLE, ppc64.ABGT},
-	ssa.BlockPPC64GT: {ppc64.ABGT, ppc64.ABLE},
-
-	ssa.BlockPPC64ULT: {ppc64.ABLT, ppc64.ABGE},
-	ssa.BlockPPC64UGE: {ppc64.ABGE, ppc64.ABLT},
-	ssa.BlockPPC64ULE: {ppc64.ABLE, ppc64.ABGT},
-	ssa.BlockPPC64UGT: {ppc64.ABGT, ppc64.ABLE},
+	ssa.BlockPPC64EQ: {ppc64.ABEQ, ppc64.ABNE, false, false},
+	ssa.BlockPPC64NE: {ppc64.ABNE, ppc64.ABEQ, false, false},
+
+	ssa.BlockPPC64LT: {ppc64.ABLT, ppc64.ABGE, false, false},
+	ssa.BlockPPC64GE: {ppc64.ABGE, ppc64.ABLT, false, false},
+	ssa.BlockPPC64LE: {ppc64.ABLE, ppc64.ABGT, false, false},
+	ssa.BlockPPC64GT: {ppc64.ABGT, ppc64.ABLE, false, false},
+
+	// TODO: need to work FP comparisons into block jumps
+	ssa.BlockPPC64FLT: {ppc64.ABLT, ppc64.ABGT, false, true},
+	ssa.BlockPPC64FGE: {ppc64.ABGT, ppc64.ABLT, true, false},
+	ssa.BlockPPC64FLE: {ppc64.ABLT, ppc64.ABGT, true, false},
+	ssa.BlockPPC64FGT: {ppc64.ABGT, ppc64.ABLT, false, true},
 }
 
 func ssaGenBlock(s *gc.SSAGenState, b, next *ssa.Block) {
@@ -893,12 +953,17 @@ func ssaGenBlock(s *gc.SSAGenState, b, next *ssa.Block) {
 		gc.Prog(obj.AUNDEF) // tell plive.go that we never reach here
 	case ssa.BlockRet:
 		gc.Prog(obj.ARET)
+	case ssa.BlockRetJmp:
+		p := gc.Prog(obj.AJMP)
+		p.To.Type = obj.TYPE_MEM
+		p.To.Name = obj.NAME_EXTERN
+		p.To.Sym = gc.Linksym(b.Aux.(*gc.Sym))
 
 	case ssa.BlockPPC64EQ, ssa.BlockPPC64NE,
 		ssa.BlockPPC64LT, ssa.BlockPPC64GE,
 		ssa.BlockPPC64LE, ssa.BlockPPC64GT,
-		ssa.BlockPPC64ULT, ssa.BlockPPC64UGT,
-		ssa.BlockPPC64ULE, ssa.BlockPPC64UGE:
+		ssa.BlockPPC64FLT, ssa.BlockPPC64FGE,
+		ssa.BlockPPC64FLE, ssa.BlockPPC64FGT:
 		jmp := blockJump[b.Kind]
 		likely := b.Likely
 		var p *obj.Prog
@@ -908,14 +973,30 @@ func ssaGenBlock(s *gc.SSAGenState, b, next *ssa.Block) {
 			likely *= -1
 			p.To.Type = obj.TYPE_BRANCH
 			s.Branches = append(s.Branches, gc.Branch{P: p, B: b.Succs[1].Block()})
+			if jmp.invasmeq {
+				// TODO: The second branch is probably predict-not-taken since it is for FP equality
+				q := gc.Prog(ppc64.ABEQ)
+				q.To.Type = obj.TYPE_BRANCH
+				s.Branches = append(s.Branches, gc.Branch{P: q, B: b.Succs[1].Block()})
+			}
 		case b.Succs[1].Block():
 			p = gc.Prog(jmp.asm)
 			p.To.Type = obj.TYPE_BRANCH
 			s.Branches = append(s.Branches, gc.Branch{P: p, B: b.Succs[0].Block()})
+			if jmp.asmeq {
+				q := gc.Prog(ppc64.ABEQ)
+				q.To.Type = obj.TYPE_BRANCH
+				s.Branches = append(s.Branches, gc.Branch{P: q, B: b.Succs[0].Block()})
+			}
 		default:
 			p = gc.Prog(jmp.asm)
 			p.To.Type = obj.TYPE_BRANCH
 			s.Branches = append(s.Branches, gc.Branch{P: p, B: b.Succs[0].Block()})
+			if jmp.asmeq {
+				q := gc.Prog(ppc64.ABEQ)
+				q.To.Type = obj.TYPE_BRANCH
+				s.Branches = append(s.Branches, gc.Branch{P: q, B: b.Succs[0].Block()})
+			}
 			q := gc.Prog(obj.AJMP)
 			q.To.Type = obj.TYPE_BRANCH
 			s.Branches = append(s.Branches, gc.Branch{P: q, B: b.Succs[1].Block()})

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

@@ -31,6 +31,7 @@ type Config struct {
 	noDuffDevice    bool                       // Don't use Duff's device
 	nacl            bool                       // GOOS=nacl
 	use387          bool                       // GO386=387
+	NeedsFpScratch  bool                       // No direct move between GP and FP register sets
 	sparsePhiCutoff uint64                     // Sparse phi location algorithm used above this #blocks*#variables score
 	curFunc         *Func
 
@@ -190,6 +191,7 @@ func NewConfig(arch string, fe Frontend, ctxt *obj.Link, optimize bool) *Config
 		c.fpRegMask = fpRegMaskPPC64
 		c.FPReg = framepointerRegPPC64
 		c.noDuffDevice = true // TODO: Resolve PPC64 DuffDevice (has zero, but not copy)
+		c.NeedsFpScratch = true
 		c.hasGReg = true
 	default:
 		fe.Unimplementedf(0, "arch %s not implemented", arch)
@@ -245,6 +247,7 @@ func NewConfig(arch string, fe Frontend, ctxt *obj.Link, optimize bool) *Config
 }
 
 func (c *Config) Set387(b bool) {
+	c.NeedsFpScratch = b
 	c.use387 = b
 }
 

src/cmd/compile/internal/ssa/gen/PPC64.rules

@@ -28,6 +28,8 @@
 (Mod32 x y) -> (SUB x (MULLW y (DIVW x y)))
 (Mod32u x y) -> (SUB x (MULLW y (DIVWU x y)))
 
+(Avg64u <t> x y) -> (ADD (ADD <t> (SRD <t> x (MOVDconst <t> [1])) (SRD <t> y (MOVDconst <t> [1]))) (ANDconst <t> (AND <t> x y) [1]))
+
 (Mul64  x y) -> (MULLD  x y)
 (Mul32  x y) -> (MULLW  x y)
 (Mul16  x y) -> (MULLW x y)
@@ -57,6 +59,22 @@
 (Div32F x y) -> (FDIVS x y)
 (Div64F x y) -> (FDIV x y)
 
+// Lowering float <-> int
+(Cvt32to32F x) -> (FRSP (FCFID (Xi2f64 (SignExt32to64 x))))
+(Cvt32to64F x) -> (FCFID (Xi2f64 (SignExt32to64 x)))
+(Cvt64to32F x) -> (FRSP (FCFID (Xi2f64 x)))
+(Cvt64to64F x) -> (FCFID (Xi2f64 x))
+
+(Cvt32Fto32 x) -> (Xf2i64 (FCTIWZ x))
+(Cvt32Fto64 x) -> (Xf2i64 (FCTIDZ x))
+(Cvt64Fto32 x) -> (Xf2i64 (FCTIWZ x))
+(Cvt64Fto64 x) -> (Xf2i64 (FCTIDZ x))
+
+(Cvt32Fto64F x) -> x // Note x will have the wrong type for patterns dependent on Float32/Float64
+(Cvt64Fto32F x) -> (FRSP x)
+
+(Sqrt x) -> (FSQRT x)
+
 (Rsh64x64 x y)  -> (SRAD x (ORN y <config.fe.TypeInt64()> (MaskIfNotCarry (ADDconstForCarry [-64] y))))
 (Rsh64Ux64 x y) -> (SRD  x (ORN y <config.fe.TypeInt64()> (MaskIfNotCarry (ADDconstForCarry [-64] y))))
 (Lsh64x64 x y)  -> (SLD  x (ORN y <config.fe.TypeInt64()> (MaskIfNotCarry (ADDconstForCarry [-64] y))))
@@ -199,8 +217,8 @@
 (Less16 x y) -> (LessThan (CMPW (SignExt16to32 x) (SignExt16to32 y)))
 (Less32 x y) -> (LessThan (CMPW x y))
 (Less64 x y) -> (LessThan (CMP x y))
-(Less32F x y) -> (LessThan (FCMPU x y))
-(Less64F x y) -> (LessThan (FCMPU x y))
+(Less32F x y) -> (FLessThan (FCMPU x y))
+(Less64F x y) -> (FLessThan (FCMPU x y))
 
 (Less8U x y)  -> (LessThan (CMPWU (ZeroExt8to32 x) (ZeroExt8to32 y)))
 (Less16U x y) -> (LessThan (CMPWU (ZeroExt16to32 x) (ZeroExt16to32 y)))
@@ -211,8 +229,8 @@
 (Leq16 x y) -> (LessEqual (CMPW (SignExt16to32 x) (SignExt16to32 y)))
 (Leq32 x y) -> (LessEqual (CMPW x y))
 (Leq64 x y) -> (LessEqual (CMP x y))
-(Leq32F x y) -> (LessEqual (FCMPU x y))
-(Leq64F x y) -> (LessEqual (FCMPU x y))
+(Leq32F x y) -> (FLessEqual (FCMPU x y))
+(Leq64F x y) -> (FLessEqual (FCMPU x y))
 
 (Leq8U x y)  -> (LessEqual (CMPWU (ZeroExt8to32 x) (ZeroExt8to32 y)))
 (Leq16U x y) -> (LessEqual (CMPWU (ZeroExt16to32 x) (ZeroExt16to32 y)))
@@ -223,8 +241,8 @@
 (Greater16 x y) -> (GreaterThan (CMPW (SignExt16to32 x) (SignExt16to32 y)))
 (Greater32 x y) -> (GreaterThan (CMPW x y))
 (Greater64 x y) -> (GreaterThan (CMP x y))
-(Greater32F x y) -> (GreaterThan (FCMPU x y))
-(Greater64F x y) -> (GreaterThan (FCMPU x y))
+(Greater32F x y) -> (FGreaterThan (FCMPU x y))
+(Greater64F x y) -> (FGreaterThan (FCMPU x y))
 
 (Greater8U x y)  -> (GreaterThan (CMPWU (ZeroExt8to32 x) (ZeroExt8to32 y)))
 (Greater16U x y) -> (GreaterThan (CMPWU (ZeroExt16to32 x) (ZeroExt16to32 y)))
@@ -235,8 +253,8 @@
 (Geq16 x y) -> (GreaterEqual (CMPW (SignExt16to32 x) (SignExt16to32 y)))
 (Geq32 x y) -> (GreaterEqual (CMPW x y))
 (Geq64 x y) -> (GreaterEqual (CMP x y))
-(Geq32F x y) -> (GreaterEqual (FCMPU x y))
-(Geq64F x y) -> (GreaterEqual (FCMPU x y))
+(Geq32F x y) -> (FGreaterEqual (FCMPU x y))
+(Geq64F x y) -> (FGreaterEqual (FCMPU x y))
 
 (Geq8U x y)  -> (GreaterEqual (CMPU (ZeroExt8to32 x) (ZeroExt8to32 y)))
 (Geq16U x y) -> (GreaterEqual (CMPU (ZeroExt16to32 x) (ZeroExt16to32 y)))
@@ -250,6 +268,10 @@
 (If (LessEqual cc) yes no) -> (LE cc yes no)
 (If (GreaterThan cc) yes no) -> (GT cc yes no)
 (If (GreaterEqual cc) yes no) -> (GE cc yes no)
+(If (FLessThan cc) yes no) -> (FLT cc yes no)
+(If (FLessEqual cc) yes no) -> (FLE cc yes no)
+(If (FGreaterThan cc) yes no) -> (FGT cc yes no)
+(If (FGreaterEqual cc) yes no) -> (FGE cc yes no)
 
 (If cond yes no) -> (NE (CMPWconst [0] cond) yes no)
 
@@ -260,6 +282,10 @@
 (NE (CMPWconst [0] (LessEqual cc)) yes no) -> (LE cc yes no)
 (NE (CMPWconst [0] (GreaterThan cc)) yes no) -> (GT cc yes no)
 (NE (CMPWconst [0] (GreaterEqual cc)) yes no) -> (GE cc yes no)
+// (NE (CMPWconst [0] (FLessThan cc)) yes no) -> (FLT cc yes no)
+// (NE (CMPWconst [0] (FLessEqual cc)) yes no) -> (FLE cc yes no)
+// (NE (CMPWconst [0] (FGreaterThan cc)) yes no) -> (FGT cc yes no)
+// (NE (CMPWconst [0] (FGreaterEqual cc)) yes no) -> (FGE cc yes no)
 
 // absorb flag constants into branches
 (EQ (FlagEQ) yes no) -> (First nil yes no)
@@ -294,6 +320,11 @@
 (EQ (InvertFlags cmp) yes no) -> (EQ cmp yes no)
 (NE (InvertFlags cmp) yes no) -> (NE cmp yes no)
 
+// (FLT (InvertFlags cmp) yes no) -> (FGT cmp yes no)
+// (FGT (InvertFlags cmp) yes no) -> (FLT cmp yes no)
+// (FLE (InvertFlags cmp) yes no) -> (FGE cmp yes no)
+// (FGE (InvertFlags cmp) yes no) -> (FLE cmp yes no)
+
 // constant comparisons
 (CMPWconst (MOVWconst [x]) [y]) && int32(x)==int32(y) -> (FlagEQ)
 (CMPWconst (MOVWconst [x]) [y]) && int32(x)<int32(y)  -> (FlagLT)
@@ -349,6 +380,10 @@
 (GreaterThan (InvertFlags x)) -> (LessThan x)
 (LessEqual (InvertFlags x)) -> (GreaterEqual x)
 (GreaterEqual (InvertFlags x)) -> (LessEqual x)
+(FLessThan (InvertFlags x)) -> (FGreaterThan x)
+(FGreaterThan (InvertFlags x)) -> (FLessThan x)
+(FLessEqual (InvertFlags x)) -> (FGreaterEqual x)
+(FGreaterEqual (InvertFlags x)) -> (FLessEqual x)
 
 
 // Lowering loads
@@ -364,6 +399,7 @@
 (Load <t> ptr mem) && is64BitFloat(t) -> (FMOVDload ptr mem)
 
 (Store [8] ptr val mem) && is64BitFloat(val.Type) -> (FMOVDstore ptr val mem)
+(Store [8] ptr val mem) && is32BitFloat(val.Type) -> (FMOVDstore ptr val mem) // glitch from (Cvt32Fto64F x) -> x -- type is wrong
 (Store [4] ptr val mem) && is32BitFloat(val.Type) -> (FMOVSstore ptr val mem)
 (Store [8] ptr val mem) && (is64BitInt(val.Type) || isPtr(val.Type)) -> (MOVDstore ptr val mem)
 (Store [4] ptr val mem) && is32BitInt(val.Type) -> (MOVWstore ptr val mem)

src/cmd/compile/internal/ssa/gen/PPC64Ops.go

@@ -77,11 +77,11 @@ var regNamesPPC64 = []string{
 	"F24",
 	"F25",
 	"F26",
-	"F27",
-	"F28",
-	"F29",
-	"F30",
-	"F31",
+	// "F27", // reserved for "floating conversion constant"
+	// "F28", // 0.0
+	// "F29", // 0.5
+	// "F30", // 1.0
+	// "F31", // 2.0
 
 	// "CR0",
 	// "CR1",
@@ -121,16 +121,16 @@ func init() {
 
 	var (
 		gp = buildReg("R3 R4 R5 R6 R7 R8 R9 R10 R11 R12 R14 R15 R16 R17 R18 R19 R20 R21 R22 R23 R24 R25 R26 R27 R28 R29")
-		fp = buildReg("F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31")
+		fp = buildReg("F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26")
 		sp = buildReg("SP")
 		sb = buildReg("SB")
-		// gr = buildReg("g")
-		//cr = buildReg("CR")
-		//ctr  = buildReg("CTR")
-		//lr   = buildReg("LR")
+		// gr  = buildReg("g")
+		// cr  = buildReg("CR")
+		// ctr = buildReg("CTR")
+		// lr  = buildReg("LR")
 		tmp  = buildReg("R31")
 		ctxt = buildReg("R11")
-		//		tls	= buildReg("R13")
+		// tls = buildReg("R13")
 		gp01        = regInfo{inputs: nil, outputs: []regMask{gp}}
 		gp11        = regInfo{inputs: []regMask{gp | sp | sb}, outputs: []regMask{gp}}
 		gp21        = regInfo{inputs: []regMask{gp | sp | sb, gp | sp | sb}, outputs: []regMask{gp}}
@@ -142,6 +142,8 @@ func init() {
 		gpstorezero = regInfo{inputs: []regMask{gp | sp | sb}} // ppc64.REGZERO is reserved zero value
 		fp01        = regInfo{inputs: nil, outputs: []regMask{fp}}
 		fp11        = regInfo{inputs: []regMask{fp}, outputs: []regMask{fp}}
+		fpgp        = regInfo{inputs: []regMask{fp}, outputs: []regMask{gp}}
+		gpfp        = regInfo{inputs: []regMask{gp}, outputs: []regMask{fp}}
 		fp21        = regInfo{inputs: []regMask{fp, fp}, outputs: []regMask{fp}}
 		fp2cr       = regInfo{inputs: []regMask{fp, fp}}
 		fpload      = regInfo{inputs: []regMask{gp | sp | sb}, outputs: []regMask{fp}}
@@ -195,6 +197,21 @@ func init() {
 
 		// MOD is implemented as rem := arg0 - (arg0/arg1) * arg1
 
+		// Conversions are all float-to-float register operations.  "Integer" refers to encoding in the FP register.
+		{name: "FCTIDZ", argLength: 1, reg: fp11, asm: "FCTIDZ", typ: "Float64"}, // convert float to 64-bit int round towards zero
+		{name: "FCTIWZ", argLength: 1, reg: fp11, asm: "FCTIWZ", typ: "Float64"}, // convert float to 32-bit int round towards zero
+		{name: "FCFID", argLength: 1, reg: fp11, asm: "FCFID", typ: "Float64"},   // convert 64-bit integer to float
+		{name: "FRSP", argLength: 1, reg: fp11, asm: "FRSP", typ: "Float64"},     // round float to 32-bit value
+
+		// Movement between float and integer registers with no change in bits; accomplished with stores+loads on PPC.
+		// Because the 32-bit load-literal-bits instructions have impoverished addressability, always widen the
+		// data instead and use FMOVDload and FMOVDstore instead (this will also dodge endianess issues).
+		// There are optimizations that should apply -- (Xi2f64 (MOVWload (not-ADD-ptr+offset) ) ) could use
+		// the word-load instructions.  (Xi2f64 (MOVDload ptr )) can be (FMOVDload ptr)
+
+		{name: "Xf2i64", argLength: 1, reg: fpgp, typ: "Int64"},   // move 64 bits of F register into G register
+		{name: "Xi2f64", argLength: 1, reg: gpfp, typ: "Float64"}, // move 64 bits of G register into F register
+
 		{name: "AND", argLength: 2, reg: gp21, asm: "AND", commutative: true},               // arg0&arg1
 		{name: "ANDN", argLength: 2, reg: gp21, asm: "ANDN"},                                // arg0&^arg1
 		{name: "OR", argLength: 2, reg: gp21, asm: "OR", commutative: true},                 // arg0|arg1
@@ -203,6 +220,8 @@ func init() {
 		{name: "EQV", argLength: 2, reg: gp21, asm: "EQV", typ: "Int64", commutative: true}, // arg0^^arg1
 		{name: "NEG", argLength: 1, reg: gp11, asm: "NEG"},                                  // -arg0 (integer)
 		{name: "FNEG", argLength: 1, reg: fp11, asm: "FNEG"},                                // -arg0 (floating point)
+		{name: "FSQRT", argLength: 1, reg: fp11, asm: "FSQRT"},                              // sqrt(arg0) (floating point)
+		{name: "FSQRTS", argLength: 1, reg: fp11, asm: "FSQRTS"},                            // sqrt(arg0) (floating point, single precision)
 
 		{name: "ORconst", argLength: 1, reg: gp11, asm: "OR", aux: "Int64"},                                                                                     // arg0|aux
 		{name: "XORconst", argLength: 1, reg: gp11, asm: "XOR", aux: "Int64"},                                                                                   // arg0^aux
@@ -254,12 +273,16 @@ func init() {
 		{name: "CMPWUconst", argLength: 1, reg: gp1cr, asm: "CMPWU", aux: "Int32", typ: "Flags"},
 
 		// pseudo-ops
-		{name: "Equal", argLength: 1, reg: crgp},        // bool, true flags encode x==y false otherwise.
-		{name: "NotEqual", argLength: 1, reg: crgp},     // bool, true flags encode x!=y false otherwise.
-		{name: "LessThan", argLength: 1, reg: crgp},     // bool, true flags encode  x<y false otherwise.
-		{name: "LessEqual", argLength: 1, reg: crgp},    // bool, true flags encode  x<=y false otherwise.
-		{name: "GreaterThan", argLength: 1, reg: crgp},  // bool, true flags encode  x>y false otherwise.
-		{name: "GreaterEqual", argLength: 1, reg: crgp}, // bool, true flags encode  x>=y false otherwise.
+		{name: "Equal", argLength: 1, reg: crgp},         // bool, true flags encode x==y false otherwise.
+		{name: "NotEqual", argLength: 1, reg: crgp},      // bool, true flags encode x!=y false otherwise.
+		{name: "LessThan", argLength: 1, reg: crgp},      // bool, true flags encode  x<y false otherwise.
+		{name: "FLessThan", argLength: 1, reg: crgp},     // bool, true flags encode  x<y false otherwise.
+		{name: "LessEqual", argLength: 1, reg: crgp},     // bool, true flags encode  x<=y false otherwise.
+		{name: "FLessEqual", argLength: 1, reg: crgp},    // bool, true flags encode  x<=y false otherwise; PPC <= === !> which is wrong for NaN
+		{name: "GreaterThan", argLength: 1, reg: crgp},   // bool, true flags encode  x>y false otherwise.
+		{name: "FGreaterThan", argLength: 1, reg: crgp},  // bool, true flags encode  x>y false otherwise.
+		{name: "GreaterEqual", argLength: 1, reg: crgp},  // bool, true flags encode  x>=y false otherwise.
+		{name: "FGreaterEqual", argLength: 1, reg: crgp}, // bool, true flags encode  x>=y false otherwise.; PPC >= === !< which is wrong for NaN
 
 		// Scheduler ensures LoweredGetClosurePtr occurs only in entry block,
 		// and sorts it to the very beginning of the block to prevent other
@@ -352,10 +375,10 @@ func init() {
 		{name: "LE"},
 		{name: "GT"},
 		{name: "GE"},
-		{name: "ULT"},
-		{name: "ULE"},
-		{name: "UGT"},
-		{name: "UGE"},
+		{name: "FLT"},
+		{name: "FLE"},
+		{name: "FGT"},
+		{name: "FGE"},
 	}
 
 	archs = append(archs, arch{

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

@@ -29,7 +29,7 @@ func checkLower(f *Func) {
 					continue // ok not to lower
 				}
 			}
-			s := "not lowered: " + v.Op.String() + " " + v.Type.SimpleString()
+			s := "not lowered: " + v.String() + ", " + v.Op.String() + " " + v.Type.SimpleString()
 			for _, a := range v.Args {
 				s += " " + a.Type.SimpleString()
 			}

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

@@ -380,7 +380,7 @@ func flush387(s *gc.SSAGenState) {
 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.Node = s.ScratchFpMem
+	a.Sym = gc.Linksym(s.ScratchFpMem.Sym)
 	a.Reg = x86.REG_SP
 }

src/cmd/internal/obj/ppc64/a.out.go

@@ -219,8 +219,8 @@ const (
 	C_LAUTO
 	C_SEXT
 	C_LEXT
-	C_ZOREG
-	C_SOREG
+	C_ZOREG // conjecture: either (1) register + zeroed offset, or (2) "R0" implies zero or C_REG
+	C_SOREG // register + signed offset
 	C_LOREG
 	C_FPSCR
 	C_MSR
@@ -324,6 +324,8 @@ const (
 	AFMOVDU
 	AFMOVS
 	AFMOVSU
+	AFMOVSX
+	AFMOVSZ
 	AFMSUB
 	AFMSUBCC
 	AFMSUBS

src/cmd/internal/obj/ppc64/anames.go

@@ -91,6 +91,8 @@ var Anames = []string{
 	"FMOVDU",
 	"FMOVS",
 	"FMOVSU",
+	"FMOVSX",
+	"FMOVSZ",
 	"FMSUB",
 	"FMSUBCC",
 	"FMSUBS",

src/cmd/internal/obj/ppc64/asm9.go

@@ -53,7 +53,7 @@ type Optab struct {
 	a2    uint8
 	a3    uint8
 	a4    uint8
-	type_ int8
+	type_ int8 // cases in asmout below. E.g., 44 = st r,(ra+rb); 45 = ld (ra+rb), r
 	size  int8
 	param int16
 }
@@ -310,6 +310,12 @@ var optab = []Optab{
 	{AFMOVD, C_FREG, C_NONE, C_NONE, C_LAUTO, 35, 8, REGSP},
 	{AFMOVD, C_FREG, C_NONE, C_NONE, C_LOREG, 35, 8, REGZERO},
 	{AFMOVD, C_FREG, C_NONE, C_NONE, C_ADDR, 74, 8, 0},
+	{AFMOVSX, C_ZOREG, C_REG, C_NONE, C_FREG, 45, 4, 0},
+	{AFMOVSX, C_ZOREG, C_NONE, C_NONE, C_FREG, 45, 4, 0},
+	{AFMOVSX, C_FREG, C_REG, C_NONE, C_ZOREG, 44, 4, 0},
+	{AFMOVSX, C_FREG, C_NONE, C_NONE, C_ZOREG, 44, 4, 0},
+	{AFMOVSZ, C_ZOREG, C_REG, C_NONE, C_FREG, 45, 4, 0},
+	{AFMOVSZ, C_ZOREG, C_NONE, C_NONE, C_FREG, 45, 4, 0},
 	{ASYNC, C_NONE, C_NONE, C_NONE, C_NONE, 46, 4, 0},
 	{AWORD, C_LCON, C_NONE, C_NONE, C_NONE, 40, 4, 0},
 	{ADWORD, C_LCON, C_NONE, C_NONE, C_NONE, 31, 8, 0},
@@ -920,7 +926,7 @@ func buildop(ctxt *obj.Link) {
 		switch r {
 		default:
 			ctxt.Diag("unknown op in build: %v", obj.Aconv(r))
-			log.Fatalf("bad code")
+			log.Fatalf("instruction missing from switch in asm9.go:buildop: %v", obj.Aconv(r))
 
 		case ADCBF: /* unary indexed: op (b+a); op (b) */
 			opset(ADCBI, r0)
@@ -1265,6 +1271,8 @@ func buildop(ctxt *obj.Link) {
 
 		case AADD,
 			AANDCC, /* and. Rb,Rs,Ra; andi. $uimm,Rs,Ra; andis. $uimm,Rs,Ra */
+			AFMOVSX,
+			AFMOVSZ,
 			ALSW,
 			AMOVW,
 			/* load/store/move word with sign extension; special 32-bit move; move 32-bit literals */
@@ -3238,6 +3246,10 @@ func oploadx(ctxt *obj.Link, a obj.As) uint32 {
 		return OPVCC(31, 535, 0, 0) /* lfsx */
 	case AFMOVSU:
 		return OPVCC(31, 567, 0, 0) /* lfsux */
+	case AFMOVSX:
+		return OPVCC(31, 855, 0, 0) /* lfiwax - power6, isa 2.05 */
+	case AFMOVSZ:
+		return OPVCC(31, 887, 0, 0) /* lfiwzx - power7, isa 2.06 */
 	case AMOVH:
 		return OPVCC(31, 343, 0, 0) /* lhax */
 	case AMOVHU:
@@ -3332,6 +3344,8 @@ func opstorex(ctxt *obj.Link, a obj.As) uint32 {
 		return OPVCC(31, 663, 0, 0) /* stfsx */
 	case AFMOVSU:
 		return OPVCC(31, 695, 0, 0) /* stfsux */
+	case AFMOVSX:
+		return OPVCC(31, 983, 0, 0) /* stfiwx */
 
 	case AMOVHZ, AMOVH:
 		return OPVCC(31, 407, 0, 0) /* sthx */