cmd/compile: intrinsics for math/bits.TrailingZerosX

Implement math/bits.TrailingZerosX using intrinsics.

Generally reorganize the intrinsic spec a bit.
The instrinsics data structure is now built at init time.
This will make doing the other functions in math/bits easier.

Update sys.CtzX to return int instead of uint{64,32} so it
matches math/bits.TrailingZerosX.

Improve the intrinsics a bit for amd64.  We don't need the CMOV
for <64 bit versions.

Update #18616

Change-Id: Ic1c5339c943f961d830ae56f12674d7b29d4ff39
Reviewed-on: https://go-review.googlesource.com/38155
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>

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

@@ -162,7 +162,7 @@ var allAsmTests = []*asmTests{
 	{
 		arch:    "amd64",
 		os:      "linux",
-		imports: []string{"encoding/binary"},
+		imports: []string{"encoding/binary", "math/bits"},
 		tests:   linuxAMD64Tests,
 	},
 	{
@@ -174,7 +174,7 @@ var allAsmTests = []*asmTests{
 	{
 		arch:    "s390x",
 		os:      "linux",
-		imports: []string{"encoding/binary"},
+		imports: []string{"encoding/binary", "math/bits"},
 		tests:   linuxS390XTests,
 	},
 	{
@@ -543,6 +543,39 @@ var linuxAMD64Tests = []*asmTest{
 		`,
 		[]string{"\tBTQ\t\\$60"},
 	},
+	// Intrinsic tests for math/bits
+	{
+		`
+		func f41(a uint64) int {
+			return bits.TrailingZeros64(a)
+		}
+		`,
+		[]string{"\tBSFQ\t", "\tMOVQ\t\\$64,", "\tCMOVQEQ\t"},
+	},
+	{
+		`
+		func f42(a uint32) int {
+			return bits.TrailingZeros32(a)
+		}
+		`,
+		[]string{"\tBSFQ\t", "\tORQ\t[^$]", "\tMOVQ\t\\$4294967296,"},
+	},
+	{
+		`
+		func f43(a uint16) int {
+			return bits.TrailingZeros16(a)
+		}
+		`,
+		[]string{"\tBSFQ\t", "\tORQ\t\\$65536,"},
+	},
+	{
+		`
+		func f44(a uint8) int {
+			return bits.TrailingZeros8(a)
+		}
+		`,
+		[]string{"\tBSFQ\t", "\tORQ\t\\$256,"},
+	},
 }
 
 var linux386Tests = []*asmTest{
@@ -710,6 +743,39 @@ var linuxS390XTests = []*asmTest{
 		`,
 		[]string{"\tFMSUBS\t"},
 	},
+	// Intrinsic tests for math/bits
+	{
+		`
+		func f18(a uint64) int {
+			return bits.TrailingZeros64(a)
+		}
+		`,
+		[]string{"\tFLOGR\t"},
+	},
+	{
+		`
+		func f19(a uint32) int {
+			return bits.TrailingZeros32(a)
+		}
+		`,
+		[]string{"\tFLOGR\t", "\tMOVWZ\t"},
+	},
+	{
+		`
+		func f20(a uint16) int {
+			return bits.TrailingZeros16(a)
+		}
+		`,
+		[]string{"\tFLOGR\t", "\tOR\t\\$65536,"},
+	},
+	{
+		`
+		func f21(a uint8) int {
+			return bits.TrailingZeros8(a)
+		}
+		`,
+		[]string{"\tFLOGR\t", "\tOR\t\\$256,"},
+	},
 }
 
 var linuxARMTests = []*asmTest{

src/cmd/compile/internal/ssa/gen/AMD64.rules

@@ -98,7 +98,7 @@
 
 // Lowering other arithmetic
 (Ctz64 <t> x) -> (CMOVQEQ (Select0 <t> (BSFQ x)) (MOVQconst <t> [64]) (Select1 <TypeFlags> (BSFQ x)))
-(Ctz32 <t> x) -> (CMOVLEQ (Select0 <t> (BSFL x)) (MOVLconst <t> [32]) (Select1 <TypeFlags> (BSFL x)))
+(Ctz32 x) -> (Select0 (BSFQ (ORQ <config.Frontend().TypeUInt64()> (MOVQconst [1<<32]) x)))
 
 (Bswap64 x) -> (BSWAPQ x)
 (Bswap32 x) -> (BSWAPL x)
@@ -2083,3 +2083,9 @@
 	(CMPXCHGQlock [off1+off2] {sym} ptr old new_ mem)
 (CMPXCHGLlock [off1] {sym} (ADDQconst [off2] ptr) old new_ mem) && is32Bit(off1+off2) ->
 	(CMPXCHGLlock [off1+off2] {sym} ptr old new_ mem)
+
+// We don't need the conditional move if we know the arg of BSF is not zero.
+(CMOVQEQ x _ (Select1 (BSFQ (ORQconst [c] _)))) && c != 0 -> x
+// Extension is unnecessary for trailing zeros.
+(BSFQ (ORQconst <t> [1<<8] (MOVBQZX x))) -> (BSFQ (ORQconst <t> [1<<8] x))
+(BSFQ (ORQconst <t> [1<<16] (MOVWQZX x))) -> (BSFQ (ORQconst <t> [1<<16] x))

src/cmd/compile/internal/ssa/gen/dec64.rules

@@ -108,13 +108,11 @@
 		(Com32 <config.fe.TypeUInt32()> (Int64Lo x)))
 
 (Ctz64 x) ->
-	(Int64Make
-		(Const32 <config.fe.TypeUInt32()> [0])
 	(Add32 <config.fe.TypeUInt32()>
 		(Ctz32 <config.fe.TypeUInt32()> (Int64Lo x))
 		(And32 <config.fe.TypeUInt32()>
 			(Com32 <config.fe.TypeUInt32()> (Zeromask (Int64Lo x)))
-				(Ctz32 <config.fe.TypeUInt32()> (Int64Hi x)))))
+			(Ctz32 <config.fe.TypeUInt32()> (Int64Hi x))))
 
 (Bswap64 x) ->
 	(Int64Make

src/cmd/compile/internal/ssa/rewriteAMD64.go

@@ -28,8 +28,12 @@ func rewriteValueAMD64(v *Value, config *Config) bool {
 		return rewriteValueAMD64_OpAMD64ANDQ(v, config)
 	case OpAMD64ANDQconst:
 		return rewriteValueAMD64_OpAMD64ANDQconst(v, config)
+	case OpAMD64BSFQ:
+		return rewriteValueAMD64_OpAMD64BSFQ(v, config)
 	case OpAMD64BTQconst:
 		return rewriteValueAMD64_OpAMD64BTQconst(v, config)
+	case OpAMD64CMOVQEQ:
+		return rewriteValueAMD64_OpAMD64CMOVQEQ(v, config)
 	case OpAMD64CMPB:
 		return rewriteValueAMD64_OpAMD64CMPB(v, config)
 	case OpAMD64CMPBconst:
@@ -2158,6 +2162,59 @@ func rewriteValueAMD64_OpAMD64ANDQconst(v *Value, config *Config) bool {
 	}
 	return false
 }
+func rewriteValueAMD64_OpAMD64BSFQ(v *Value, config *Config) bool {
+	b := v.Block
+	_ = b
+	// match: (BSFQ (ORQconst <t> [1<<8] (MOVBQZX x)))
+	// cond:
+	// result: (BSFQ (ORQconst <t> [1<<8] x))
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpAMD64ORQconst {
+			break
+		}
+		t := v_0.Type
+		if v_0.AuxInt != 1<<8 {
+			break
+		}
+		v_0_0 := v_0.Args[0]
+		if v_0_0.Op != OpAMD64MOVBQZX {
+			break
+		}
+		x := v_0_0.Args[0]
+		v.reset(OpAMD64BSFQ)
+		v0 := b.NewValue0(v.Pos, OpAMD64ORQconst, t)
+		v0.AuxInt = 1 << 8
+		v0.AddArg(x)
+		v.AddArg(v0)
+		return true
+	}
+	// match: (BSFQ (ORQconst <t> [1<<16] (MOVWQZX x)))
+	// cond:
+	// result: (BSFQ (ORQconst <t> [1<<16] x))
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpAMD64ORQconst {
+			break
+		}
+		t := v_0.Type
+		if v_0.AuxInt != 1<<16 {
+			break
+		}
+		v_0_0 := v_0.Args[0]
+		if v_0_0.Op != OpAMD64MOVWQZX {
+			break
+		}
+		x := v_0_0.Args[0]
+		v.reset(OpAMD64BSFQ)
+		v0 := b.NewValue0(v.Pos, OpAMD64ORQconst, t)
+		v0.AuxInt = 1 << 16
+		v0.AddArg(x)
+		v.AddArg(v0)
+		return true
+	}
+	return false
+}
 func rewriteValueAMD64_OpAMD64BTQconst(v *Value, config *Config) bool {
 	b := v.Block
 	_ = b
@@ -2177,6 +2234,37 @@ func rewriteValueAMD64_OpAMD64BTQconst(v *Value, config *Config) bool {
 	}
 	return false
 }
+func rewriteValueAMD64_OpAMD64CMOVQEQ(v *Value, config *Config) bool {
+	b := v.Block
+	_ = b
+	// match: (CMOVQEQ x _ (Select1 (BSFQ (ORQconst [c] _))))
+	// cond: c != 0
+	// result: x
+	for {
+		x := v.Args[0]
+		v_2 := v.Args[2]
+		if v_2.Op != OpSelect1 {
+			break
+		}
+		v_2_0 := v_2.Args[0]
+		if v_2_0.Op != OpAMD64BSFQ {
+			break
+		}
+		v_2_0_0 := v_2_0.Args[0]
+		if v_2_0_0.Op != OpAMD64ORQconst {
+			break
+		}
+		c := v_2_0_0.AuxInt
+		if !(c != 0) {
+			break
+		}
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
+	return false
+}
 func rewriteValueAMD64_OpAMD64CMPB(v *Value, config *Config) bool {
 	b := v.Block
 	_ = b
@@ -17902,26 +17990,20 @@ func rewriteValueAMD64_OpConvert(v *Value, config *Config) bool {
 func rewriteValueAMD64_OpCtz32(v *Value, config *Config) bool {
 	b := v.Block
 	_ = b
-	// match: (Ctz32 <t> x)
+	// match: (Ctz32 x)
 	// cond:
-	// result: (CMOVLEQ (Select0 <t> (BSFL x)) (MOVLconst <t> [32]) (Select1 <TypeFlags> (BSFL x)))
+	// result: (Select0 (BSFQ (ORQ <config.Frontend().TypeUInt64()> (MOVQconst [1<<32]) x)))
 	for {
-		t := v.Type
 		x := v.Args[0]
-		v.reset(OpAMD64CMOVLEQ)
-		v0 := b.NewValue0(v.Pos, OpSelect0, t)
-		v1 := b.NewValue0(v.Pos, OpAMD64BSFL, MakeTuple(config.fe.TypeUInt32(), TypeFlags))
+		v.reset(OpSelect0)
+		v0 := b.NewValue0(v.Pos, OpAMD64BSFQ, MakeTuple(config.fe.TypeUInt64(), TypeFlags))
+		v1 := b.NewValue0(v.Pos, OpAMD64ORQ, config.Frontend().TypeUInt64())
+		v2 := b.NewValue0(v.Pos, OpAMD64MOVQconst, config.fe.TypeUInt64())
+		v2.AuxInt = 1 << 32
+		v1.AddArg(v2)
 		v1.AddArg(x)
 		v0.AddArg(v1)
 		v.AddArg(v0)
-		v2 := b.NewValue0(v.Pos, OpAMD64MOVLconst, t)
-		v2.AuxInt = 32
-		v.AddArg(v2)
-		v3 := b.NewValue0(v.Pos, OpSelect1, TypeFlags)
-		v4 := b.NewValue0(v.Pos, OpAMD64BSFL, MakeTuple(config.fe.TypeUInt32(), TypeFlags))
-		v4.AddArg(x)
-		v3.AddArg(v4)
-		v.AddArg(v3)
 		return true
 	}
 }

src/cmd/compile/internal/ssa/rewritedec64.go

@@ -368,34 +368,30 @@ func rewriteValuedec64_OpCtz64(v *Value, config *Config) bool {
 	_ = b
 	// match: (Ctz64 x)
 	// cond:
-	// result: (Int64Make 		(Const32 <config.fe.TypeUInt32()> [0]) 		(Add32 <config.fe.TypeUInt32()> 			(Ctz32 <config.fe.TypeUInt32()> (Int64Lo x)) 			(And32 <config.fe.TypeUInt32()> 				(Com32 <config.fe.TypeUInt32()> (Zeromask (Int64Lo x))) 				(Ctz32 <config.fe.TypeUInt32()> (Int64Hi x)))))
+	// result: (Add32 <config.fe.TypeUInt32()> 		(Ctz32 <config.fe.TypeUInt32()> (Int64Lo x)) 		(And32 <config.fe.TypeUInt32()> 			(Com32 <config.fe.TypeUInt32()> (Zeromask (Int64Lo x))) 			(Ctz32 <config.fe.TypeUInt32()> (Int64Hi x))))
 	for {
 		x := v.Args[0]
-		v.reset(OpInt64Make)
-		v0 := b.NewValue0(v.Pos, OpConst32, config.fe.TypeUInt32())
-		v0.AuxInt = 0
+		v.reset(OpAdd32)
+		v.Type = config.fe.TypeUInt32()
+		v0 := b.NewValue0(v.Pos, OpCtz32, config.fe.TypeUInt32())
+		v1 := b.NewValue0(v.Pos, OpInt64Lo, config.fe.TypeUInt32())
+		v1.AddArg(x)
+		v0.AddArg(v1)
 		v.AddArg(v0)
-		v1 := b.NewValue0(v.Pos, OpAdd32, config.fe.TypeUInt32())
-		v2 := b.NewValue0(v.Pos, OpCtz32, config.fe.TypeUInt32())
-		v3 := b.NewValue0(v.Pos, OpInt64Lo, config.fe.TypeUInt32())
-		v3.AddArg(x)
+		v2 := b.NewValue0(v.Pos, OpAnd32, config.fe.TypeUInt32())
+		v3 := b.NewValue0(v.Pos, OpCom32, config.fe.TypeUInt32())
+		v4 := b.NewValue0(v.Pos, OpZeromask, config.fe.TypeUInt32())
+		v5 := b.NewValue0(v.Pos, OpInt64Lo, config.fe.TypeUInt32())
+		v5.AddArg(x)
+		v4.AddArg(v5)
+		v3.AddArg(v4)
 		v2.AddArg(v3)
-		v1.AddArg(v2)
-		v4 := b.NewValue0(v.Pos, OpAnd32, config.fe.TypeUInt32())
-		v5 := b.NewValue0(v.Pos, OpCom32, config.fe.TypeUInt32())
-		v6 := b.NewValue0(v.Pos, OpZeromask, config.fe.TypeUInt32())
-		v7 := b.NewValue0(v.Pos, OpInt64Lo, config.fe.TypeUInt32())
+		v6 := b.NewValue0(v.Pos, OpCtz32, config.fe.TypeUInt32())
+		v7 := b.NewValue0(v.Pos, OpInt64Hi, config.fe.TypeUInt32())
 		v7.AddArg(x)
 		v6.AddArg(v7)
-		v5.AddArg(v6)
-		v4.AddArg(v5)
-		v8 := b.NewValue0(v.Pos, OpCtz32, config.fe.TypeUInt32())
-		v9 := b.NewValue0(v.Pos, OpInt64Hi, config.fe.TypeUInt32())
-		v9.AddArg(x)
-		v8.AddArg(v9)
-		v4.AddArg(v8)
-		v1.AddArg(v4)
-		v.AddArg(v1)
+		v2.AddArg(v6)
+		v.AddArg(v2)
 		return true
 	}
 }

src/runtime/internal/sys/intrinsics.go

@@ -32,22 +32,22 @@ var deBruijnIdx32 = [32]byte{
 
 // Ctz64 counts trailing (low-order) zeroes,
 // and if all are zero, then 64.
-func Ctz64(x uint64) uint64 {
+func Ctz64(x uint64) int {
 	x &= -x                      // isolate low-order bit
 	y := x * deBruijn64 >> 58    // extract part of deBruijn sequence
-	y = uint64(deBruijnIdx64[y]) // convert to bit index
-	z := (x - 1) >> 57 & 64      // adjustment if zero
-	return y + z
+	i := int(deBruijnIdx64[y])   // convert to bit index
+	z := int((x - 1) >> 57 & 64) // adjustment if zero
+	return i + z
 }
 
 // Ctz32 counts trailing (low-order) zeroes,
 // and if all are zero, then 32.
-func Ctz32(x uint32) uint32 {
+func Ctz32(x uint32) int {
 	x &= -x                      // isolate low-order bit
 	y := x * deBruijn32 >> 27    // extract part of deBruijn sequence
-	y = uint32(deBruijnIdx32[y]) // convert to bit index
-	z := (x - 1) >> 26 & 32      // adjustment if zero
-	return y + z
+	i := int(deBruijnIdx32[y])   // convert to bit index
+	z := int((x - 1) >> 26 & 32) // adjustment if zero
+	return i + z
 }
 
 // Bswap64 returns its input with byte order reversed

src/runtime/internal/sys/intrinsics_386.s

@@ -4,14 +4,12 @@
 
 #include "textflag.h"
 
-TEXT runtime∕internal∕sys·Ctz64(SB), NOSPLIT, $0-16
-	MOVL	$0, ret_hi+12(FP)
-
+TEXT runtime∕internal∕sys·Ctz64(SB), NOSPLIT, $0-12
 	// Try low 32 bits.
 	MOVL	x_lo+0(FP), AX
 	BSFL	AX, AX
 	JZ	tryhigh
-	MOVL	AX, ret_lo+8(FP)
+	MOVL	AX, ret+8(FP)
 	RET
 
 tryhigh:
@@ -20,12 +18,12 @@ tryhigh:
 	BSFL	AX, AX
 	JZ	none
 	ADDL	$32, AX
-	MOVL	AX, ret_lo+8(FP)
+	MOVL	AX, ret+8(FP)
 	RET
 
 none:
 	// No bits are set.
-	MOVL	$64, ret_lo+8(FP)
+	MOVL	$64, ret+8(FP)
 	RET
 
 TEXT runtime∕internal∕sys·Ctz32(SB), NOSPLIT, $0-8

src/runtime/internal/sys/intrinsics_stubs.go

@@ -6,7 +6,7 @@
 
 package sys
 
-func Ctz64(x uint64) uint64
-func Ctz32(x uint32) uint32
+func Ctz64(x uint64) int
+func Ctz32(x uint32) int
 func Bswap64(x uint64) uint64
 func Bswap32(x uint32) uint32

src/runtime/internal/sys/intrinsics_test.go

@@ -6,17 +6,17 @@ import (
 )
 
 func TestCtz64(t *testing.T) {
-	for i := uint(0); i <= 64; i++ {
-		x := uint64(5) << i
-		if got := sys.Ctz64(x); got != uint64(i) {
+	for i := 0; i <= 64; i++ {
+		x := uint64(5) << uint(i)
+		if got := sys.Ctz64(x); got != i {
 			t.Errorf("Ctz64(%d)=%d, want %d", x, got, i)
 		}
 	}
 }
 func TestCtz32(t *testing.T) {
-	for i := uint(0); i <= 32; i++ {
-		x := uint32(5) << i
-		if got := sys.Ctz32(x); got != uint32(i) {
+	for i := 0; i <= 32; i++ {
+		x := uint32(5) << uint(i)
+		if got := sys.Ctz32(x); got != i {
 			t.Errorf("Ctz32(%d)=%d, want %d", x, got, i)
 		}
 	}

src/runtime/malloc.go

@@ -491,7 +491,7 @@ func nextFreeFast(s *mspan) gclinkptr {
 			if freeidx%64 == 0 && freeidx != s.nelems {
 				return 0
 			}
-			s.allocCache >>= (theBit + 1)
+			s.allocCache >>= uint(theBit + 1)
 			s.freeindex = freeidx
 			v := gclinkptr(result*s.elemsize + s.base())
 			s.allocCount++

src/runtime/mbitmap.go

@@ -248,7 +248,7 @@ func (s *mspan) nextFreeIndex() uintptr {
 		return snelems
 	}
 
-	s.allocCache >>= (bitIndex + 1)
+	s.allocCache >>= uint(bitIndex + 1)
 	sfreeindex = result + 1
 
 	if sfreeindex%64 == 0 && sfreeindex != snelems {

test/intrinsic.dir/main.go

@@ -22,7 +22,7 @@ func logf(f string, args ...interface{}) {
 	}
 }
 
-func test(i, x uint64) {
+func test(i int, x uint64) {
 	t := T.Ctz64(x) // ERROR "intrinsic substitution for Ctz64"
 	if i != t {
 		logf("Ctz64(0x%x) expected %d but got %d\n", x, i, t)
@@ -36,12 +36,12 @@ func test(i, x uint64) {
 	if i <= 32 {
 		x32 := uint32(x)
 		t32 := T.Ctz32(x32) // ERROR "intrinsic substitution for Ctz32"
-		if uint32(i) != t32 {
+		if i != t32 {
 			logf("Ctz32(0x%x) expected %d but got %d\n", x32, i, t32)
 		}
 		x32 = -x32
 		t32 = T.Ctz32(x32) // ERROR "intrinsic substitution for Ctz32"
-		if uint32(i) != t32 {
+		if i != t32 {
 			logf("Ctz32(0x%x) expected %d but got %d\n", x32, i, t32)
 		}
 	}
@@ -83,10 +83,10 @@ func main() {
 		logf("ctz64(0) != 64")
 	}
 
-	for i := uint64(0); i <= 64; i++ {
+	for i := 0; i <= 64; i++ {
 		for j := uint64(1); j <= 255; j += 2 {
 			for k := uint64(1); k <= 65537; k += 128 {
-				x := (j * k) << i
+				x := (j * k) << uint(i)
 				test(i, x)
 			}
 		}