[dev.simd] cmd/compile: Generated code for AVX2 SIMD masked load/store

This adds to the change in the earlier dev.simd CL.
Generated by arch/internal/simdgen CL 689276 .

Also includes one test for "it at least works once".

Change-Id: I44a268cfc3bea06c5522ac2cfa04fe13a833e1dd
Reviewed-on: https://go-review.googlesource.com/c/go/+/689335
Reviewed-by: Junyang Shao <shaojunyang@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>

src/cmd/compile/internal/ssagen/simdintrinsics.go

@@ -2132,6 +2132,30 @@ func simdIntrinsics(addF func(pkg, fn string, b intrinsicBuilder, archFamilies .
 	addF(simdPackage, "Uint64x4.Store", simdStore(), sys.AMD64)
 	addF(simdPackage, "LoadUint64x8", simdLoad(), sys.AMD64)
 	addF(simdPackage, "Uint64x8.Store", simdStore(), sys.AMD64)
+	addF(simdPackage, "LoadMaskedFloat32x4", simdMaskedLoad(ssa.OpLoadMasked32), sys.AMD64)
+	addF(simdPackage, "Float32x4.StoreMasked", simdMaskedStore(ssa.OpStoreMasked32), sys.AMD64)
+	addF(simdPackage, "LoadMaskedFloat32x8", simdMaskedLoad(ssa.OpLoadMasked32), sys.AMD64)
+	addF(simdPackage, "Float32x8.StoreMasked", simdMaskedStore(ssa.OpStoreMasked32), sys.AMD64)
+	addF(simdPackage, "LoadMaskedFloat64x2", simdMaskedLoad(ssa.OpLoadMasked64), sys.AMD64)
+	addF(simdPackage, "Float64x2.StoreMasked", simdMaskedStore(ssa.OpStoreMasked64), sys.AMD64)
+	addF(simdPackage, "LoadMaskedFloat64x4", simdMaskedLoad(ssa.OpLoadMasked64), sys.AMD64)
+	addF(simdPackage, "Float64x4.StoreMasked", simdMaskedStore(ssa.OpStoreMasked64), sys.AMD64)
+	addF(simdPackage, "LoadMaskedInt32x4", simdMaskedLoad(ssa.OpLoadMasked32), sys.AMD64)
+	addF(simdPackage, "Int32x4.StoreMasked", simdMaskedStore(ssa.OpStoreMasked32), sys.AMD64)
+	addF(simdPackage, "LoadMaskedInt32x8", simdMaskedLoad(ssa.OpLoadMasked32), sys.AMD64)
+	addF(simdPackage, "Int32x8.StoreMasked", simdMaskedStore(ssa.OpStoreMasked32), sys.AMD64)
+	addF(simdPackage, "LoadMaskedInt64x2", simdMaskedLoad(ssa.OpLoadMasked64), sys.AMD64)
+	addF(simdPackage, "Int64x2.StoreMasked", simdMaskedStore(ssa.OpStoreMasked64), sys.AMD64)
+	addF(simdPackage, "LoadMaskedInt64x4", simdMaskedLoad(ssa.OpLoadMasked64), sys.AMD64)
+	addF(simdPackage, "Int64x4.StoreMasked", simdMaskedStore(ssa.OpStoreMasked64), sys.AMD64)
+	addF(simdPackage, "LoadMaskedUint32x4", simdMaskedLoad(ssa.OpLoadMasked32), sys.AMD64)
+	addF(simdPackage, "Uint32x4.StoreMasked", simdMaskedStore(ssa.OpStoreMasked32), sys.AMD64)
+	addF(simdPackage, "LoadMaskedUint32x8", simdMaskedLoad(ssa.OpLoadMasked32), sys.AMD64)
+	addF(simdPackage, "Uint32x8.StoreMasked", simdMaskedStore(ssa.OpStoreMasked32), sys.AMD64)
+	addF(simdPackage, "LoadMaskedUint64x2", simdMaskedLoad(ssa.OpLoadMasked64), sys.AMD64)
+	addF(simdPackage, "Uint64x2.StoreMasked", simdMaskedStore(ssa.OpStoreMasked64), sys.AMD64)
+	addF(simdPackage, "LoadMaskedUint64x4", simdMaskedLoad(ssa.OpLoadMasked64), sys.AMD64)
+	addF(simdPackage, "Uint64x4.StoreMasked", simdMaskedStore(ssa.OpStoreMasked64), sys.AMD64)
 	addF(simdPackage, "Mask8x16.AsInt8x16", func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value { return args[0] }, sys.AMD64)
 	addF(simdPackage, "Int8x16.AsMask8x16", func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value { return args[0] }, sys.AMD64)
 	addF(simdPackage, "Mask8x16.And", opLen2(ssa.OpAndInt32x4, types.TypeVec128), sys.AMD64)

src/simd/simd_test.go

@@ -206,16 +206,6 @@ func TestPairDotProdAccumulate(t *testing.T) {
 	}
 }
 
-// checkInt8Slices ensures that b and a are equal, to the end of b.
-// also serves to use the slices, to prevent accidental optimization.
-func checkInt8Slices(t *testing.T, a, b []int8) {
-	for i := range b {
-		if a[i] != b[i] {
-			t.Errorf("a and b differ at index %d, a=%d, b=%d", i, a[i], b[i])
-		}
-	}
-}
-
 func TestSlicesInt8(t *testing.T) {
 	a := []int8{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
 		17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32}

src/simd/slicepart_amd64.go

@@ -37,6 +37,10 @@ func int64atP32(p *int32) *int64 {
 	return (*int64)(unsafe.Pointer(p))
 }
 
+func int32atP64(p *int64) *int32 {
+	return (*int32)(unsafe.Pointer(p))
+}
+
 /* unsigned versions of integer slice part loads */
 
 // LoadUint8x16SlicePart loads a Uint8x16 from the slice s.
@@ -385,3 +389,70 @@ func (x Int16x8) StoreSlicePart(s []int16) {
 	}
 	return
 }
+
+var vecMask64 = [16]int64{
+	-1, -1, -1, -1,
+	-1, -1, -1, -1,
+	0, 0, 0, 0,
+	0, 0, 0, 0,
+}
+
+// paInt32x4 is an unchecked cast from a slice to an
+// pointer-to-array type, for used in a masked
+// load/store.  In practice, the slice will be too
+// short, so this has to be unsafe, and its only
+// use must be with an instruction with masked
+// load/store effect (including faults).
+func paInt32x4(s []int32) *[4]int32 {
+	return (*[4]int32)(unsafe.Pointer(&s[0]))
+}
+
+/* 32 and 64-bit slice-part loads for AVX2 (128 and 256 bit) */
+
+func LoadInt32x4SlicePart(s []int32) Int32x4 {
+	l := len(s)
+	if l >= 4 {
+		return LoadInt32x4Slice(s)
+	}
+	if l == 0 {
+		var x Int32x4
+		return x
+	}
+	p := int32atP64(&vecMask64[0])
+	mask := unsafe.Slice(p, 32)[16-l:]
+	return LoadMaskedInt32x4(paInt32x4(s), LoadInt32x4Slice(mask).AsMask32x4())
+}
+
+func (x Int32x4) StoreSlicePart(s []int32) {
+	l := len(s)
+	if l >= 4 {
+		x.StoreSlice(s)
+		return
+	}
+	if l == 0 {
+		return
+	}
+	p := int32atP64(&vecMask64[0])
+	mask := unsafe.Slice(p, 32)[16-l:]
+	x.StoreMasked(paInt32x4(s), LoadInt32x4Slice(mask).AsMask32x4())
+}
+
+// func LoadInt32x8SlicePart(s []int32) Int32x8 {
+// }
+
+// func LoadInt64x2SlicePart(s []int64) Int64x2 {
+// }
+
+// func LoadInt64x4SlicePart(s []int64) Int64x4 {
+// }
+
+// func (x Int32x8) StoreSlicePart(s []int32) {
+// }
+
+// func (x Int64x4) StoreSlicePart(s []int64) {
+// }
+
+// func (x Int64x8) StoreSlicePart(s []int64) {
+// }
+
+// Handle float32, float64, uint32, and uint64 with ugly casts.

src/simd/slicepart_test.go

@@ -177,3 +177,43 @@ func TestSlicesPartStoreUint8x32(t *testing.T) {
 		}
 	}
 }
+
+func TestSlicePartInt32(t *testing.T) {
+	L := 4
+	c := []int32{1, 2, 3, 4, 5, -1, -1, -1, -1}
+	a := c[:L+1]
+	for i := range a {
+		// Test the load first
+		// e is a partial slice.
+		e := a[i:]
+		v := simd.LoadInt32x4SlicePart(e)
+		// d contains what a ought to contain
+		d := make([]int32, L)
+		for j := 0; j < len(e) && j < len(d); j++ {
+			d[j] = e[j]
+		}
+
+		b := make([]int32, L)
+		v.StoreSlice(b)
+		// test the load
+		checkSlices(t, d, b)
+
+		// Test the store
+		f := make([]int32, L+1)
+		for i := range f {
+			f[i] = 99
+		}
+
+		v.StoreSlicePart(f[:len(e)])
+		if len(e) < len(b) {
+			checkSlices(t, f, b[:len(e)])
+		} else {
+			checkSlices(t, f, b)
+		}
+		for i := len(e); i < len(f); i++ {
+			if f[i] != 99 {
+				t.Errorf("StoreSlicePart altered f[%d], expected 99, saw %d", i, f[i])
+			}
+		}
+	}
+}

src/simd/types_amd64.go

@@ -28,6 +28,18 @@ func LoadFloat32x4(y *[4]float32) Float32x4
 //go:noescape
 func (x Float32x4) Store(y *[4]float32)
 
+// LoadMaskedFloat32x4 loads a Float32x4 from an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func LoadMaskedFloat32x4(y *[4]float32, mask Mask32x4) Float32x4
+
+// StoreMasked stores a Float32x4 to an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func (x Float32x4) StoreMasked(y *[4]float32, mask Mask32x4)
+
 // Float64x2 is a 128-bit SIMD vector of 2 float64
 type Float64x2 struct {
 	float64x2 v128
@@ -47,6 +59,18 @@ func LoadFloat64x2(y *[2]float64) Float64x2
 //go:noescape
 func (x Float64x2) Store(y *[2]float64)
 
+// LoadMaskedFloat64x2 loads a Float64x2 from an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func LoadMaskedFloat64x2(y *[2]float64, mask Mask64x2) Float64x2
+
+// StoreMasked stores a Float64x2 to an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func (x Float64x2) StoreMasked(y *[2]float64, mask Mask64x2)
+
 // Int8x16 is a 128-bit SIMD vector of 16 int8
 type Int8x16 struct {
 	int8x16 v128
@@ -104,6 +128,18 @@ func LoadInt32x4(y *[4]int32) Int32x4
 //go:noescape
 func (x Int32x4) Store(y *[4]int32)
 
+// LoadMaskedInt32x4 loads a Int32x4 from an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func LoadMaskedInt32x4(y *[4]int32, mask Mask32x4) Int32x4
+
+// StoreMasked stores a Int32x4 to an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func (x Int32x4) StoreMasked(y *[4]int32, mask Mask32x4)
+
 // Int64x2 is a 128-bit SIMD vector of 2 int64
 type Int64x2 struct {
 	int64x2 v128
@@ -123,6 +159,18 @@ func LoadInt64x2(y *[2]int64) Int64x2
 //go:noescape
 func (x Int64x2) Store(y *[2]int64)
 
+// LoadMaskedInt64x2 loads a Int64x2 from an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func LoadMaskedInt64x2(y *[2]int64, mask Mask64x2) Int64x2
+
+// StoreMasked stores a Int64x2 to an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func (x Int64x2) StoreMasked(y *[2]int64, mask Mask64x2)
+
 // Uint8x16 is a 128-bit SIMD vector of 16 uint8
 type Uint8x16 struct {
 	uint8x16 v128
@@ -180,6 +228,18 @@ func LoadUint32x4(y *[4]uint32) Uint32x4
 //go:noescape
 func (x Uint32x4) Store(y *[4]uint32)
 
+// LoadMaskedUint32x4 loads a Uint32x4 from an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func LoadMaskedUint32x4(y *[4]uint32, mask Mask32x4) Uint32x4
+
+// StoreMasked stores a Uint32x4 to an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func (x Uint32x4) StoreMasked(y *[4]uint32, mask Mask32x4)
+
 // Uint64x2 is a 128-bit SIMD vector of 2 uint64
 type Uint64x2 struct {
 	uint64x2 v128
@@ -199,6 +259,18 @@ func LoadUint64x2(y *[2]uint64) Uint64x2
 //go:noescape
 func (x Uint64x2) Store(y *[2]uint64)
 
+// LoadMaskedUint64x2 loads a Uint64x2 from an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func LoadMaskedUint64x2(y *[2]uint64, mask Mask64x2) Uint64x2
+
+// StoreMasked stores a Uint64x2 to an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func (x Uint64x2) StoreMasked(y *[2]uint64, mask Mask64x2)
+
 // Mask8x16 is a 128-bit SIMD vector of 16 int8
 type Mask8x16 struct {
 	int8x16 v128
@@ -311,6 +383,18 @@ func LoadFloat32x8(y *[8]float32) Float32x8
 //go:noescape
 func (x Float32x8) Store(y *[8]float32)
 
+// LoadMaskedFloat32x8 loads a Float32x8 from an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func LoadMaskedFloat32x8(y *[8]float32, mask Mask32x8) Float32x8
+
+// StoreMasked stores a Float32x8 to an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func (x Float32x8) StoreMasked(y *[8]float32, mask Mask32x8)
+
 // Float64x4 is a 256-bit SIMD vector of 4 float64
 type Float64x4 struct {
 	float64x4 v256
@@ -330,6 +414,18 @@ func LoadFloat64x4(y *[4]float64) Float64x4
 //go:noescape
 func (x Float64x4) Store(y *[4]float64)
 
+// LoadMaskedFloat64x4 loads a Float64x4 from an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func LoadMaskedFloat64x4(y *[4]float64, mask Mask64x4) Float64x4
+
+// StoreMasked stores a Float64x4 to an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func (x Float64x4) StoreMasked(y *[4]float64, mask Mask64x4)
+
 // Int8x32 is a 256-bit SIMD vector of 32 int8
 type Int8x32 struct {
 	int8x32 v256
@@ -387,6 +483,18 @@ func LoadInt32x8(y *[8]int32) Int32x8
 //go:noescape
 func (x Int32x8) Store(y *[8]int32)
 
+// LoadMaskedInt32x8 loads a Int32x8 from an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func LoadMaskedInt32x8(y *[8]int32, mask Mask32x8) Int32x8
+
+// StoreMasked stores a Int32x8 to an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func (x Int32x8) StoreMasked(y *[8]int32, mask Mask32x8)
+
 // Int64x4 is a 256-bit SIMD vector of 4 int64
 type Int64x4 struct {
 	int64x4 v256
@@ -406,6 +514,18 @@ func LoadInt64x4(y *[4]int64) Int64x4
 //go:noescape
 func (x Int64x4) Store(y *[4]int64)
 
+// LoadMaskedInt64x4 loads a Int64x4 from an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func LoadMaskedInt64x4(y *[4]int64, mask Mask64x4) Int64x4
+
+// StoreMasked stores a Int64x4 to an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func (x Int64x4) StoreMasked(y *[4]int64, mask Mask64x4)
+
 // Uint8x32 is a 256-bit SIMD vector of 32 uint8
 type Uint8x32 struct {
 	uint8x32 v256
@@ -463,6 +583,18 @@ func LoadUint32x8(y *[8]uint32) Uint32x8
 //go:noescape
 func (x Uint32x8) Store(y *[8]uint32)
 
+// LoadMaskedUint32x8 loads a Uint32x8 from an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func LoadMaskedUint32x8(y *[8]uint32, mask Mask32x8) Uint32x8
+
+// StoreMasked stores a Uint32x8 to an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func (x Uint32x8) StoreMasked(y *[8]uint32, mask Mask32x8)
+
 // Uint64x4 is a 256-bit SIMD vector of 4 uint64
 type Uint64x4 struct {
 	uint64x4 v256
@@ -482,6 +614,18 @@ func LoadUint64x4(y *[4]uint64) Uint64x4
 //go:noescape
 func (x Uint64x4) Store(y *[4]uint64)
 
+// LoadMaskedUint64x4 loads a Uint64x4 from an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func LoadMaskedUint64x4(y *[4]uint64, mask Mask64x4) Uint64x4
+
+// StoreMasked stores a Uint64x4 to an array,
+// at those elements enabled by mask
+//
+//go:noescape
+func (x Uint64x4) StoreMasked(y *[4]uint64, mask Mask64x4)
+
 // Mask8x32 is a 256-bit SIMD vector of 32 int8
 type Mask8x32 struct {
 	int8x32 v256