Add not_index_byte and not_index_byte2 functions to simdstring package

Fixes #9646

tools/simdstring/benchmarks_test.go

@@ -64,3 +64,45 @@ func BenchmarkIndexByte2(b *testing.B) {
 		t(pos, "scalar")
 	}
 }
+
+func BenchmarkNotIndexByte(b *testing.B) {
+	t := func(pos int, which string) {
+		// Fill with 'a' and place 'q' (a non-matching byte) at the target position
+		data := haystack('a', 'q', pos)
+		f := NotIndexByte
+		switch which {
+		case "scalar":
+			f = not_index_byte_scalar
+		}
+		b.Run(fmt.Sprintf("%s_sz=%d", which, pos), func(b *testing.B) {
+			for b.Loop() {
+				f(data, 'a')
+			}
+		})
+	}
+	for _, pos := range sizes {
+		t(pos, "simdstring")
+		t(pos, "scalar")
+	}
+}
+
+func BenchmarkNotIndexByte2(b *testing.B) {
+	t := func(pos int, which string) {
+		// Fill with 'a' and place 'q' (neither 'a' nor 'x') at the target position
+		data := haystack('a', 'q', pos)
+		f := NotIndexByte2
+		switch which {
+		case "scalar":
+			f = not_index_byte2_scalar
+		}
+		b.Run(fmt.Sprintf("%s_sz=%d", which, pos), func(b *testing.B) {
+			for b.Loop() {
+				f(data, 'a', 'x')
+			}
+		})
+	}
+	for _, pos := range sizes {
+		t(pos, "simdstring")
+		t(pos, "scalar")
+	}
+}

tools/simdstring/generate.go

@@ -404,6 +404,12 @@ func encode_cmgt16b(a, b, dest Register) (ans uint32) {
 	return 0x271<<21 | b.ARMId()<<16 | 0xd<<10 | a.ARMId()<<5 | dest.ARMId()
 }
 
+func encode_not16b(src, dest Register) uint32 {
+	// NOT Vd.16B, Vn.16B (alias of MVN)
+	// Encoding: 0 Q 1 01110 size 10000 00101 10 Rn Rd (Q=1, size=00 for .16B)
+	return 0x6E205800 | (src.ARMId() << 5) | dest.ARMId()
+}
+
 func (f *Function) MaskForCountDestructive(vec, ans Register) {
 	// vec is clobbered by this function
 	f.Comment("Count the number of bytes to the first 0xff byte and put the result in", ans)
@@ -688,6 +694,24 @@ func (f *Function) Or(a, b, dest Register) {
 	f.AddTrailingComment(dest, "=", a, "|", b, "(bitwise)")
 }
 
+func (f *Function) NotSelf(r Register) {
+	if f.ISA.Goarch == ARM64 {
+		f.Comment("Go assembler doesn't support the VMVN instruction, below we have: NOT", r.ARMFullWidth()+",", r.ARMFullWidth())
+		f.instr("WORD", fmt.Sprintf("$0x%x", encode_not16b(r, r)))
+		f.AddTrailingComment(r, "= ~", r, "(bitwise NOT)")
+		return
+	}
+	all_ones := f.Vec(r.Size)
+	defer f.ReleaseReg(all_ones)
+	f.AllOnesRegister(all_ones)
+	if r.Size == 128 {
+		f.instr("PXOR", all_ones, r)
+	} else {
+		f.instr("VPXOR", all_ones, r, r)
+	}
+	f.AddTrailingComment(r, "= ~", r, "(bitwise NOT)")
+}
+
 func (f *Function) And(a, b, dest Register) {
 	if f.ISA.Goarch == ARM64 {
 		f.instr("VAND", a.ARMFullWidth(), b.ARMFullWidth(), dest.ARMFullWidth())
@@ -1504,6 +1528,54 @@ func (s *State) indexc0() {
 
 }
 
+func (s *State) not_index_byte_body(f *Function) {
+	b := f.Vec()
+	f.Set1Epi8("b", b)
+	test_bytes := func(bytes_to_test, test_ans Register) {
+		f.CmpEqEpi8(bytes_to_test, b, test_ans)
+		f.NotSelf(test_ans)
+	}
+	s.index_func(f, test_bytes)
+}
+
+func (s *State) not_index_byte() {
+	f := s.NewFunction("not_index_byte_asm", "Find the index of the first byte that is not b", []FunctionParam{{"data", ByteSlice}, {"b", types.Byte}}, []FunctionParam{{"ans", types.Int}})
+	if s.ISA.HasSIMD {
+		s.not_index_byte_body(f)
+	}
+	f = s.NewFunction("not_index_byte_string_asm", "Find the index of the first byte that is not b", []FunctionParam{{"data", types.String}, {"b", types.Byte}}, []FunctionParam{{"ans", types.Int}})
+	if s.ISA.HasSIMD {
+		s.not_index_byte_body(f)
+	}
+
+}
+
+func (s *State) not_index_byte2_body(f *Function) {
+	b1 := f.Vec()
+	b2 := f.Vec()
+	f.Set1Epi8("b1", b1)
+	f.Set1Epi8("b2", b2)
+	test_bytes := func(bytes_to_test, test_ans Register) {
+		f.CmpEqEpi8(bytes_to_test, b1, test_ans)
+		f.CmpEqEpi8(bytes_to_test, b2, bytes_to_test)
+		f.Or(test_ans, bytes_to_test, test_ans)
+		f.NotSelf(test_ans)
+	}
+	s.index_func(f, test_bytes)
+}
+
+func (s *State) not_index_byte2() {
+	f := s.NewFunction("not_index_byte2_asm", "Find the index of the first byte that is neither b1 nor b2", []FunctionParam{{"data", ByteSlice}, {"b1", types.Byte}, {"b2", types.Byte}}, []FunctionParam{{"ans", types.Int}})
+	if s.ISA.HasSIMD {
+		s.not_index_byte2_body(f)
+	}
+	f = s.NewFunction("not_index_byte2_string_asm", "Find the index of the first byte that is neither b1 nor b2", []FunctionParam{{"data", types.String}, {"b1", types.Byte}, {"b2", types.Byte}}, []FunctionParam{{"ans", types.Int}})
+	if s.ISA.HasSIMD {
+		s.not_index_byte2_body(f)
+	}
+
+}
+
 func (s *State) Generate() {
 	s.test_load()
 	s.test_set1_epi8()
@@ -1516,6 +1588,8 @@ func (s *State) Generate() {
 	s.indexbyte2()
 	s.indexc0()
 	s.indexbyte()
+	s.not_index_byte()
+	s.not_index_byte2()
 
 	s.OutputFunction()
 }

tools/simdstring/intrinsics.go

@@ -33,6 +33,18 @@ var IndexC0 func(data []byte) int = index_c0_scalar
 // Return the index at which the first C0 byte is found or -1 when no such bytes are present.
 var IndexC0String func(data string) int = index_c0_string_scalar
 
+// Return the index of the first byte in data that is not equal to b. If all bytes equal b, -1 is returned.
+var NotIndexByte func(data []byte, b byte) int = not_index_byte_scalar
+
+// Return the index of the first byte in text that is not equal to b. If all bytes equal b, -1 is returned.
+var NotIndexByteString func(text string, b byte) int = not_index_byte_string_scalar
+
+// Return the index of the first byte in data that is neither a nor b. If all bytes are a or b, -1 is returned.
+var NotIndexByte2 func(data []byte, a, b byte) int = not_index_byte2_scalar
+
+// Return the index of the first byte in text that is neither a nor b. If all bytes are a or b, -1 is returned.
+var NotIndexByte2String func(text string, a, b byte) int = not_index_byte2_string_scalar
+
 func init() {
 	switch runtime.GOARCH {
 	case "amd64":
@@ -51,6 +63,10 @@ func init() {
 		IndexByte2String = index_byte2_string_asm_256
 		IndexC0 = index_c0_asm_256
 		IndexC0String = index_c0_string_asm_256
+		NotIndexByte = not_index_byte_asm_256
+		NotIndexByteString = not_index_byte_string_asm_256
+		NotIndexByte2 = not_index_byte2_asm_256
+		NotIndexByte2String = not_index_byte2_string_asm_256
 		VectorSize = 32
 	} else if Have128bit {
 		IndexByte = index_byte_asm_128
@@ -59,6 +75,10 @@ func init() {
 		IndexByte2String = index_byte2_string_asm_128
 		IndexC0 = index_c0_asm_128
 		IndexC0String = index_c0_string_asm_128
+		NotIndexByte = not_index_byte_asm_128
+		NotIndexByteString = not_index_byte_string_asm_128
+		NotIndexByte2 = not_index_byte2_asm_128
+		NotIndexByte2String = not_index_byte2_string_asm_128
 		VectorSize = 16
 	}
 }

tools/simdstring/intrinsics_test.go

@@ -244,6 +244,65 @@ func TestSIMDStringOps(t *testing.T) {
 	index_test([]byte("abc"), 'x')
 	index_test([]byte("abc"), 'b')
 
+	not_index_test := func(haystack []byte, needle byte) {
+		var actual int
+		expected := not_index_byte_scalar(haystack, needle)
+
+		for _, sz := range sizes {
+			switch sz {
+			case 16:
+				actual = not_index_byte_asm_128(haystack, needle)
+			case 32:
+				actual = not_index_byte_asm_256(haystack, needle)
+			}
+			if actual != expected {
+				t.Fatalf("not_index failed in: %#v (%d != %d) at size: %d with needle: %#v", string(haystack), expected, actual, sz, needle)
+			}
+		}
+	}
+	not_index_test(nil, 'a')
+	not_index_test([]byte{}, 'a')
+	not_index_test([]byte("aaa"), 'a')
+	not_index_test([]byte("aaab"), 'a')
+	not_index_test([]byte("baaa"), 'a')
+	not_index_test([]byte("abc"), 'a')
+	for _, sz := range []int{0, 16, 32, 64, 79} {
+		q := strings.Repeat("a", sz) + "b"
+		not_index_test([]byte(q), 'a')
+		not_index_test([]byte(q), 'b')
+		not_index_test([]byte(strings.Repeat("a", sz)), 'a')
+	}
+
+	not_index2_test := func(haystack []byte, a, b byte) {
+		var actual int
+		expected := not_index_byte2_scalar(haystack, a, b)
+
+		for _, sz := range sizes {
+			switch sz {
+			case 16:
+				actual = not_index_byte2_asm_128(haystack, a, b)
+			case 32:
+				actual = not_index_byte2_asm_256(haystack, a, b)
+			}
+			if actual != expected {
+				t.Fatalf("not_index2 failed in: %#v (%d != %d) at size: %d with needles: %#v %#v", string(haystack), expected, actual, sz, a, b)
+			}
+		}
+	}
+	not_index2_test(nil, 'a', 'b')
+	not_index2_test([]byte{}, 'a', 'b')
+	not_index2_test([]byte("aabb"), 'a', 'b')
+	not_index2_test([]byte("aabbc"), 'a', 'b')
+	not_index2_test([]byte("caabb"), 'a', 'b')
+	for _, sz := range []int{0, 16, 32, 64, 79} {
+		q := strings.Repeat("ab", sz) + "c"
+		not_index2_test([]byte(q), 'a', 'b')
+		not_index2_test([]byte(strings.Repeat("ab", sz)), 'a', 'b')
+		for align := range 32 {
+			not_index2_test([]byte(strings.Repeat(" ", align)+q), 'a', 'b')
+		}
+	}
+
 }
 
 func TestIntrinsics(t *testing.T) {

tools/simdstring/scalar.go

@@ -57,3 +57,39 @@ func index_c0_string_scalar(data string) int {
 	}
 	return -1
 }
+
+func not_index_byte_scalar(data []byte, b byte) int {
+	for i, ch := range data {
+		if ch != b {
+			return i
+		}
+	}
+	return -1
+}
+
+func not_index_byte_string_scalar(data string, b byte) int {
+	for i := 0; i < len(data); i++ {
+		if data[i] != b {
+			return i
+		}
+	}
+	return -1
+}
+
+func not_index_byte2_scalar(data []byte, a, b byte) int {
+	for i, ch := range data {
+		if ch != a && ch != b {
+			return i
+		}
+	}
+	return -1
+}
+
+func not_index_byte2_string_scalar(data string, a, b byte) int {
+	for i := 0; i < len(data); i++ {
+		if data[i] != a && data[i] != b {
+			return i
+		}
+	}
+	return -1
+}