[dev.ssa] cmd/compile: add floating point optimizations in SSA for ARM

Add some simplification rules for floating point ops.

cmd/internal/obj/arm supports instructions that compare FP register
to 0, but runtime softfloat simulator does not. This CL adds these
instructions to softfloat simulator as well.

Updates #15365.

Change-Id: I29405b2bfcb4c8cf106cb7a1a811409fec91b170
Reviewed-on: https://go-review.googlesource.com/24790
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>

8 files changed, 439 insertions, 0 deletions

diff --git a/src/cmd/compile/internal/arm/ssa.go b/src/cmd/compile/internal/arm/ssa.go
index 5dcd8ca1ac..bfd13464e5 100644
--- a/src/cmd/compile/internal/arm/ssa.go
+++ b/src/cmd/compile/internal/arm/ssa.go
@@ -523,6 +523,11 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 		p.From.Type = obj.TYPE_CONST
 		p.From.Offset = v.AuxInt
 		p.Reg = gc.SSARegNum(v.Args[0])
+	case ssa.OpARMCMPF0,
+		ssa.OpARMCMPD0:
+		p := gc.Prog(v.Op.Asm())
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = gc.SSARegNum(v.Args[0])
 	case ssa.OpARMCMPshiftLL:
 		genshift(v.Op.Asm(), gc.SSARegNum(v.Args[0]), gc.SSARegNum(v.Args[1]), 0, arm.SHIFT_LL, v.AuxInt)
 	case ssa.OpARMCMPshiftRL:
diff --git a/src/cmd/compile/internal/ssa/gen/ARM.rules b/src/cmd/compile/internal/ssa/gen/ARM.rules
index 7ec0e502ec..0002e8ea07 100644
--- a/src/cmd/compile/internal/ssa/gen/ARM.rules
+++ b/src/cmd/compile/internal/ssa/gen/ARM.rules
@@ -1203,3 +1203,7 @@
 (AND x (MVNshiftLL y [c])) -> (BICshiftLL x y [c])
 (AND x (MVNshiftRL y [c])) -> (BICshiftRL x y [c])
 (AND x (MVNshiftRA y [c])) -> (BICshiftRA x y [c])
+
+// floating point optimizations
+(CMPF x (MOVFconst [0])) -> (CMPF0 x)
+(CMPD x (MOVDconst [0])) -> (CMPD0 x)
diff --git a/src/cmd/compile/internal/ssa/gen/ARMOps.go b/src/cmd/compile/internal/ssa/gen/ARMOps.go
index 89576daf0e..789fef5819 100644
--- a/src/cmd/compile/internal/ssa/gen/ARMOps.go
+++ b/src/cmd/compile/internal/ssa/gen/ARMOps.go
@@ -119,6 +119,7 @@ func init() {
 		gp2store  = regInfo{inputs: []regMask{gpspsbg, gpg, gpg}, outputs: []regMask{}}
 		fp01      = regInfo{inputs: []regMask{}, outputs: []regMask{fp}}
 		fp11      = regInfo{inputs: []regMask{fp}, outputs: []regMask{fp}}
+		fp1flags  = regInfo{inputs: []regMask{fp}, outputs: []regMask{flags}}
 		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}}
@@ -299,6 +300,9 @@ func init() {
 		{name: "CMPshiftRLreg", argLength: 3, reg: gp3flags, asm: "CMP", typ: "Flags"}, // arg0 compare to arg1>>arg2, unsigned shift
 		{name: "CMPshiftRAreg", argLength: 3, reg: gp3flags, asm: "CMP", typ: "Flags"}, // arg0 compare to arg1>>arg2, signed shift
 
+		{name: "CMPF0", argLength: 1, reg: fp1flags, asm: "CMPF", typ: "Flags"}, // arg0 compare to 0, float32
+		{name: "CMPD0", argLength: 1, reg: fp1flags, asm: "CMPD", typ: "Flags"}, // arg0 compare to 0, float64
+
 		// moves
 		{name: "MOVWconst", argLength: 0, reg: gp01, aux: "Int32", asm: "MOVW", typ: "UInt32", rematerializeable: true},    // 32 low bits of auxint
 		{name: "MOVFconst", argLength: 0, reg: fp01, aux: "Float64", asm: "MOVF", typ: "Float32", rematerializeable: true}, // auxint as 64-bit float, convert to 32-bit float
diff --git a/src/cmd/compile/internal/ssa/gen/generic.rules b/src/cmd/compile/internal/ssa/gen/generic.rules
index d010403d9d..c0fe802aaa 100644
--- a/src/cmd/compile/internal/ssa/gen/generic.rules
+++ b/src/cmd/compile/internal/ssa/gen/generic.rules
@@ -840,3 +840,23 @@
   -> (Sub64 x (Mul64 <t> (Div64  <t> x (Const64 <t> [c])) (Const64 <t> [c])))
 (Mod64u <t> x (Const64 [c])) && x.Op != OpConst64 && umagic64ok(c)
   -> (Sub64 x (Mul64 <t> (Div64u <t> x (Const64 <t> [c])) (Const64 <t> [c])))
+
+// floating point optimizations
+(Add32F x (Const32F [0])) -> x
+(Add32F (Const32F [0]) x) -> x
+(Add64F x (Const64F [0])) -> x
+(Add64F (Const64F [0]) x) -> x
+(Sub32F x (Const32F [0])) -> x
+(Sub64F x (Const64F [0])) -> x
+(Mul32F x (Const32F [f2i(1)])) -> x
+(Mul32F (Const32F [f2i(1)]) x) -> x
+(Mul64F x (Const64F [f2i(1)])) -> x
+(Mul64F (Const64F [f2i(1)]) x) -> x
+(Mul32F x (Const32F [f2i(-1)])) -> (Neg32F x)
+(Mul32F (Const32F [f2i(-1)]) x) -> (Neg32F x)
+(Mul64F x (Const64F [f2i(-1)])) -> (Neg64F x)
+(Mul64F (Const64F [f2i(-1)]) x) -> (Neg64F x)
+(Div32F x (Const32F [f2i(1)])) -> x
+(Div64F x (Const64F [f2i(1)])) -> x
+(Div32F x (Const32F [f2i(-1)])) -> (Neg32F x)
+(Div64F x (Const64F [f2i(-1)])) -> (Neg32F x)
diff --git a/src/cmd/compile/internal/ssa/opGen.go b/src/cmd/compile/internal/ssa/opGen.go
index 159e1b26b4..01576c1217 100644
--- a/src/cmd/compile/internal/ssa/opGen.go
+++ b/src/cmd/compile/internal/ssa/opGen.go
@@ -696,6 +696,8 @@ const (
 	OpARMCMPshiftLLreg
 	OpARMCMPshiftRLreg
 	OpARMCMPshiftRAreg
+	OpARMCMPF0
+	OpARMCMPD0
 	OpARMMOVWconst
 	OpARMMOVFconst
 	OpARMMOVDconst
@@ -8867,6 +8869,32 @@ var opcodeTable = [...]opInfo{
 		},
 	},
 	{
+		name:   "CMPF0",
+		argLen: 1,
+		asm:    arm.ACMPF,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 4294901760}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15
+			},
+			outputs: []regMask{
+				4294967296, // FLAGS
+			},
+		},
+	},
+	{
+		name:   "CMPD0",
+		argLen: 1,
+		asm:    arm.ACMPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 4294901760}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15
+			},
+			outputs: []regMask{
+				4294967296, // FLAGS
+			},
+		},
+	},
+	{
 		name:              "MOVWconst",
 		auxType:           auxInt32,
 		argLen:            0,
diff --git a/src/cmd/compile/internal/ssa/rewriteARM.go b/src/cmd/compile/internal/ssa/rewriteARM.go
index ceac5839ef..87eaea265f 100644
--- a/src/cmd/compile/internal/ssa/rewriteARM.go
+++ b/src/cmd/compile/internal/ssa/rewriteARM.go
@@ -118,6 +118,10 @@ func rewriteValueARM(v *Value, config *Config) bool {
 		return rewriteValueARM_OpARMCMOVWLSconst(v, config)
 	case OpARMCMP:
 		return rewriteValueARM_OpARMCMP(v, config)
+	case OpARMCMPD:
+		return rewriteValueARM_OpARMCMPD(v, config)
+	case OpARMCMPF:
+		return rewriteValueARM_OpARMCMPF(v, config)
 	case OpARMCMPconst:
 		return rewriteValueARM_OpARMCMPconst(v, config)
 	case OpARMCMPshiftLL:
@@ -4117,6 +4121,48 @@ func rewriteValueARM_OpARMCMP(v *Value, config *Config) bool {
 	}
 	return false
 }
+func rewriteValueARM_OpARMCMPD(v *Value, config *Config) bool {
+	b := v.Block
+	_ = b
+	// match: (CMPD x (MOVDconst [0]))
+	// cond:
+	// result: (CMPD0 x)
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpARMMOVDconst {
+			break
+		}
+		if v_1.AuxInt != 0 {
+			break
+		}
+		v.reset(OpARMCMPD0)
+		v.AddArg(x)
+		return true
+	}
+	return false
+}
+func rewriteValueARM_OpARMCMPF(v *Value, config *Config) bool {
+	b := v.Block
+	_ = b
+	// match: (CMPF x (MOVFconst [0]))
+	// cond:
+	// result: (CMPF0 x)
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpARMMOVFconst {
+			break
+		}
+		if v_1.AuxInt != 0 {
+			break
+		}
+		v.reset(OpARMCMPF0)
+		v.AddArg(x)
+		return true
+	}
+	return false
+}
 func rewriteValueARM_OpARMCMPconst(v *Value, config *Config) bool {
 	b := v.Block
 	_ = b
diff --git a/src/cmd/compile/internal/ssa/rewritegeneric.go b/src/cmd/compile/internal/ssa/rewritegeneric.go
index 3f2325d003..00bb24a67b 100644
--- a/src/cmd/compile/internal/ssa/rewritegeneric.go
+++ b/src/cmd/compile/internal/ssa/rewritegeneric.go
@@ -54,8 +54,12 @@ func rewriteValuegeneric(v *Value, config *Config) bool {
 		return rewriteValuegeneric_OpCvt32Fto64F(v, config)
 	case OpCvt64Fto32F:
 		return rewriteValuegeneric_OpCvt64Fto32F(v, config)
+	case OpDiv32F:
+		return rewriteValuegeneric_OpDiv32F(v, config)
 	case OpDiv64:
 		return rewriteValuegeneric_OpDiv64(v, config)
+	case OpDiv64F:
+		return rewriteValuegeneric_OpDiv64F(v, config)
 	case OpDiv64u:
 		return rewriteValuegeneric_OpDiv64u(v, config)
 	case OpEq16:
@@ -498,6 +502,40 @@ func rewriteValuegeneric_OpAdd32F(v *Value, config *Config) bool {
 		v.AuxInt = f2i(float64(i2f32(c) + i2f32(d)))
 		return true
 	}
+	// match: (Add32F x (Const32F [0]))
+	// cond:
+	// result: x
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst32F {
+			break
+		}
+		if v_1.AuxInt != 0 {
+			break
+		}
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
+	// match: (Add32F (Const32F [0]) x)
+	// cond:
+	// result: x
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpConst32F {
+			break
+		}
+		if v_0.AuxInt != 0 {
+			break
+		}
+		x := v.Args[1]
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
 	return false
 }
 func rewriteValuegeneric_OpAdd64(v *Value, config *Config) bool {
@@ -582,6 +620,40 @@ func rewriteValuegeneric_OpAdd64F(v *Value, config *Config) bool {
 		v.AuxInt = f2i(i2f(c) + i2f(d))
 		return true
 	}
+	// match: (Add64F x (Const64F [0]))
+	// cond:
+	// result: x
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst64F {
+			break
+		}
+		if v_1.AuxInt != 0 {
+			break
+		}
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
+	// match: (Add64F (Const64F [0]) x)
+	// cond:
+	// result: x
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpConst64F {
+			break
+		}
+		if v_0.AuxInt != 0 {
+			break
+		}
+		x := v.Args[1]
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
 	return false
 }
 func rewriteValuegeneric_OpAdd8(v *Value, config *Config) bool {
@@ -1842,6 +1914,44 @@ func rewriteValuegeneric_OpCvt64Fto32F(v *Value, config *Config) bool {
 	}
 	return false
 }
+func rewriteValuegeneric_OpDiv32F(v *Value, config *Config) bool {
+	b := v.Block
+	_ = b
+	// match: (Div32F x (Const32F [f2i(1)]))
+	// cond:
+	// result: x
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst32F {
+			break
+		}
+		if v_1.AuxInt != f2i(1) {
+			break
+		}
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
+	// match: (Div32F x (Const32F [f2i(-1)]))
+	// cond:
+	// result: (Neg32F x)
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst32F {
+			break
+		}
+		if v_1.AuxInt != f2i(-1) {
+			break
+		}
+		v.reset(OpNeg32F)
+		v.AddArg(x)
+		return true
+	}
+	return false
+}
 func rewriteValuegeneric_OpDiv64(v *Value, config *Config) bool {
 	b := v.Block
 	_ = b
@@ -1997,6 +2107,44 @@ func rewriteValuegeneric_OpDiv64(v *Value, config *Config) bool {
 	}
 	return false
 }
+func rewriteValuegeneric_OpDiv64F(v *Value, config *Config) bool {
+	b := v.Block
+	_ = b
+	// match: (Div64F x (Const64F [f2i(1)]))
+	// cond:
+	// result: x
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst64F {
+			break
+		}
+		if v_1.AuxInt != f2i(1) {
+			break
+		}
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
+	// match: (Div64F x (Const64F [f2i(-1)]))
+	// cond:
+	// result: (Neg32F x)
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst64F {
+			break
+		}
+		if v_1.AuxInt != f2i(-1) {
+			break
+		}
+		v.reset(OpNeg32F)
+		v.AddArg(x)
+		return true
+	}
+	return false
+}
 func rewriteValuegeneric_OpDiv64u(v *Value, config *Config) bool {
 	b := v.Block
 	_ = b
@@ -5310,6 +5458,72 @@ func rewriteValuegeneric_OpMul32F(v *Value, config *Config) bool {
 		v.AuxInt = f2i(float64(i2f32(c) * i2f32(d)))
 		return true
 	}
+	// match: (Mul32F x (Const32F [f2i(1)]))
+	// cond:
+	// result: x
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst32F {
+			break
+		}
+		if v_1.AuxInt != f2i(1) {
+			break
+		}
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
+	// match: (Mul32F (Const32F [f2i(1)]) x)
+	// cond:
+	// result: x
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpConst32F {
+			break
+		}
+		if v_0.AuxInt != f2i(1) {
+			break
+		}
+		x := v.Args[1]
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
+	// match: (Mul32F x (Const32F [f2i(-1)]))
+	// cond:
+	// result: (Neg32F x)
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst32F {
+			break
+		}
+		if v_1.AuxInt != f2i(-1) {
+			break
+		}
+		v.reset(OpNeg32F)
+		v.AddArg(x)
+		return true
+	}
+	// match: (Mul32F (Const32F [f2i(-1)]) x)
+	// cond:
+	// result: (Neg32F x)
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpConst32F {
+			break
+		}
+		if v_0.AuxInt != f2i(-1) {
+			break
+		}
+		x := v.Args[1]
+		v.reset(OpNeg32F)
+		v.AddArg(x)
+		return true
+	}
 	return false
 }
 func rewriteValuegeneric_OpMul64(v *Value, config *Config) bool {
@@ -5446,6 +5660,72 @@ func rewriteValuegeneric_OpMul64F(v *Value, config *Config) bool {
 		v.AuxInt = f2i(i2f(c) * i2f(d))
 		return true
 	}
+	// match: (Mul64F x (Const64F [f2i(1)]))
+	// cond:
+	// result: x
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst64F {
+			break
+		}
+		if v_1.AuxInt != f2i(1) {
+			break
+		}
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
+	// match: (Mul64F (Const64F [f2i(1)]) x)
+	// cond:
+	// result: x
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpConst64F {
+			break
+		}
+		if v_0.AuxInt != f2i(1) {
+			break
+		}
+		x := v.Args[1]
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
+	// match: (Mul64F x (Const64F [f2i(-1)]))
+	// cond:
+	// result: (Neg64F x)
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst64F {
+			break
+		}
+		if v_1.AuxInt != f2i(-1) {
+			break
+		}
+		v.reset(OpNeg64F)
+		v.AddArg(x)
+		return true
+	}
+	// match: (Mul64F (Const64F [f2i(-1)]) x)
+	// cond:
+	// result: (Neg64F x)
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpConst64F {
+			break
+		}
+		if v_0.AuxInt != f2i(-1) {
+			break
+		}
+		x := v.Args[1]
+		v.reset(OpNeg64F)
+		v.AddArg(x)
+		return true
+	}
 	return false
 }
 func rewriteValuegeneric_OpMul8(v *Value, config *Config) bool {
@@ -9412,6 +9692,23 @@ func rewriteValuegeneric_OpSub32F(v *Value, config *Config) bool {
 		v.AuxInt = f2i(float64(i2f32(c) - i2f32(d)))
 		return true
 	}
+	// match: (Sub32F x (Const32F [0]))
+	// cond:
+	// result: x
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst32F {
+			break
+		}
+		if v_1.AuxInt != 0 {
+			break
+		}
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
 	return false
 }
 func rewriteValuegeneric_OpSub64(v *Value, config *Config) bool {
@@ -9527,6 +9824,23 @@ func rewriteValuegeneric_OpSub64F(v *Value, config *Config) bool {
 		v.AuxInt = f2i(i2f(c) - i2f(d))
 		return true
 	}
+	// match: (Sub64F x (Const64F [0]))
+	// cond:
+	// result: x
+	for {
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst64F {
+			break
+		}
+		if v_1.AuxInt != 0 {
+			break
+		}
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
 	return false
 }
 func rewriteValuegeneric_OpSub8(v *Value, config *Config) bool {
diff --git a/src/runtime/softfloat_arm.go b/src/runtime/softfloat_arm.go
index 5f609c80d3..802a151fbf 100644
--- a/src/runtime/softfloat_arm.go
+++ b/src/runtime/softfloat_arm.go
@@ -464,6 +464,24 @@ execute:
 		}
 		return 1
 
+	case 0xeeb50bc0: // D[regd] :: 0 (CMPD)
+		cmp, nan := fcmp64(fgetd(regd), 0)
+		m.fflag = fstatus(nan, cmp)
+
+		if fptrace > 0 {
+			print("*** cmp D[", regd, "]::0 ", hex(m.fflag), "\n")
+		}
+		return 1
+
+	case 0xeeb50ac0: // F[regd] :: 0 (CMPF)
+		cmp, nan := fcmp64(f32to64(m.freglo[regd]), 0)
+		m.fflag = fstatus(nan, cmp)
+
+		if fptrace > 0 {
+			print("*** cmp F[", regd, "]::0 ", hex(m.fflag), "\n")
+		}
+		return 1
+
 	case 0xeeb70ac0: // D[regd] = F[regm] (MOVFD)
 		fputd(regd, f32to64(m.freglo[regm]))