Add SIMD indexByteTwo/lastIndexByteTwo for faster prefiltering

Use SIMD to search for two bytes simultaneously, replacing the
two-pass bytes.IndexByte approach in trySkip and the scalar backward
loop in asciiFuzzyIndex. AVX2+SSE2 on amd64, NEON on arm64, with
scalar fallback for other architectures.

=== query: 'l' ===
  [all]   baseline:   100.61ms  current:    98.88ms  (+1.7%)  matches: 5069891 (94.78%)
  [1T]    baseline:   889.28ms  current:   852.71ms  (+4.1%)  matches: 5069891 (94.78%)

=== query: 'lin' ===
  [all]   baseline:   281.31ms  current:   269.35ms  (+4.3%)  matches: 3516507 (65.74%)
  [1T]    baseline:  2266.51ms  current:  2238.24ms  (+1.2%)  matches: 3516507 (65.74%)

=== query: 'linux' ===
  [all]   baseline:    69.94ms  current:    68.33ms  (+2.3%)  matches: 307229 (5.74%)
  [1T]    baseline:   642.66ms  current:   589.10ms  (+8.3%)  matches: 307229 (5.74%)

=== query: 'linuxlinux' ===
  [all]   baseline:    39.56ms  current:    35.48ms  (+10.3%)  matches: 12230 (0.23%)
  [1T]    baseline:   367.88ms  current:   333.49ms  (+9.3%)  matches: 12230 (0.23%)

=== query: 'linuxlinuxlinux' ===
  [all]   baseline:    36.22ms  current:    31.59ms  (+12.8%)  matches: 865 (0.02%)
  [1T]    baseline:   339.48ms  current:   293.02ms  (+13.7%)  matches: 865 (0.02%)

src/algo/algo.go

@@ -321,22 +321,15 @@ type Algo func(caseSensitive bool, normalize bool, forward bool, input *util.Cha
 
 func trySkip(input *util.Chars, caseSensitive bool, b byte, from int) int {
 	byteArray := input.Bytes()[from:]
-	idx := bytes.IndexByte(byteArray, b)
+	// For case-insensitive search of a letter, search for both cases in one pass
-	if idx == 0 {
-		// Can't skip any further
-		return from
-	}
-	// We may need to search for the uppercase letter again. We don't have to
-	// consider normalization as we can be sure that this is an ASCII string.
 	if !caseSensitive && b >= 'a' && b <= 'z' {
-		if idx > 0 {
+		idx := indexByteTwo(byteArray, b, b-32)
-			byteArray = byteArray[:idx]
+		if idx < 0 {
-		}
+			return -1
-		uidx := bytes.IndexByte(byteArray, b-32)
-		if uidx >= 0 {
-			idx = uidx
 		}
+		return from + idx
 	}
+	idx := bytes.IndexByte(byteArray, b)
 	if idx < 0 {
 		return -1
 	}
@@ -380,14 +373,17 @@ func asciiFuzzyIndex(input *util.Chars, pattern []rune, caseSensitive bool) (int
 	}
 
 	// Find the last appearance of the last character of the pattern to limit the search scope
-	bu := b
-	if !caseSensitive && b >= 'a' && b <= 'z' {
-		bu = b - 32
-	}
 	scope := input.Bytes()[lastIdx:]
-	for offset := len(scope) - 1; offset > 0; offset-- {
+	if len(scope) > 1 {
-		if scope[offset] == b || scope[offset] == bu {
+		tail := scope[1:]
-			return firstIdx, lastIdx + offset + 1
+		var end int
+		if !caseSensitive && b >= 'a' && b <= 'z' {
+			end = lastIndexByteTwo(tail, b, b-32)
+		} else {
+			end = bytes.LastIndexByte(tail, b)
+		}
+		if end >= 0 {
+			return firstIdx, lastIdx + 1 + end + 1
 		}
 	}
 	return firstIdx, lastIdx + 1

src/algo/indexbyte2_amd64.go

@@ -0,0 +1,24 @@
+//go:build amd64
+
+package algo
+
+var _useAVX2 bool
+
+func init() {
+	_useAVX2 = cpuHasAVX2()
+}
+
+//go:noescape
+func cpuHasAVX2() bool
+
+// indexByteTwo returns the index of the first occurrence of b1 or b2 in s,
+// or -1 if neither is present. Uses AVX2 when available, SSE2 otherwise.
+//
+//go:noescape
+func indexByteTwo(s []byte, b1, b2 byte) int
+
+// lastIndexByteTwo returns the index of the last occurrence of b1 or b2 in s,
+// or -1 if neither is present. Uses AVX2 when available, SSE2 otherwise.
+//
+//go:noescape
+func lastIndexByteTwo(s []byte, b1, b2 byte) int

src/algo/indexbyte2_amd64.s

@@ -0,0 +1,377 @@
+#include "textflag.h"
+
+// func cpuHasAVX2() bool
+//
+// Checks CPUID and XGETBV for AVX2 + OS YMM support.
+TEXT ·cpuHasAVX2(SB),NOSPLIT,$0-1
+	MOVQ	BX, R8          // save BX (callee-saved, clobbered by CPUID)
+
+	// Check max CPUID leaf >= 7
+	MOVL	$0, AX
+	CPUID
+	CMPL	AX, $7
+	JL	cpuid_no
+
+	// Check OSXSAVE (CPUID.1:ECX bit 27)
+	MOVL	$1, AX
+	CPUID
+	TESTL	$(1<<27), CX
+	JZ	cpuid_no
+
+	// Check AVX2 (CPUID.7.0:EBX bit 5)
+	MOVL	$7, AX
+	MOVL	$0, CX
+	CPUID
+	TESTL	$(1<<5), BX
+	JZ	cpuid_no
+
+	// Check OS YMM state support via XGETBV
+	MOVL	$0, CX
+	BYTE	$0x0F; BYTE $0x01; BYTE $0xD0  // XGETBV → EDX:EAX
+	ANDL	$6, AX          // bits 1 (XMM) and 2 (YMM)
+	CMPL	AX, $6
+	JNE	cpuid_no
+
+	MOVQ	R8, BX          // restore BX
+	MOVB	$1, ret+0(FP)
+	RET
+
+cpuid_no:
+	MOVQ	R8, BX
+	MOVB	$0, ret+0(FP)
+	RET
+
+// func indexByteTwo(s []byte, b1, b2 byte) int
+//
+// Returns the index of the first occurrence of b1 or b2 in s, or -1.
+// Uses AVX2 (32 bytes/iter) when available, SSE2 (16 bytes/iter) otherwise.
+TEXT ·indexByteTwo(SB),NOSPLIT,$0-40
+	MOVQ	s_base+0(FP), SI
+	MOVQ	s_len+8(FP), BX
+	MOVBLZX	b1+24(FP), AX
+	MOVBLZX	b2+25(FP), CX
+	LEAQ	ret+32(FP), R8
+
+	TESTQ	BX, BX
+	JEQ	fwd_failure
+
+	// Try AVX2 for inputs >= 32 bytes
+	CMPQ	BX, $32
+	JLT	fwd_sse2
+	CMPB	·_useAVX2(SB), $1
+	JNE	fwd_sse2
+
+	// ====== AVX2 forward search ======
+	MOVD	AX, X0
+	VPBROADCASTB	X0, Y0       // Y0 = splat(b1)
+	MOVD	CX, X1
+	VPBROADCASTB	X1, Y1       // Y1 = splat(b2)
+
+	MOVQ	SI, DI
+	LEAQ	-32(SI)(BX*1), AX    // AX = last valid 32-byte chunk
+	JMP	fwd_avx2_entry
+
+fwd_avx2_loop:
+	VMOVDQU	(DI), Y2
+	VPCMPEQB	Y0, Y2, Y3
+	VPCMPEQB	Y1, Y2, Y4
+	VPOR	Y3, Y4, Y3
+	VPMOVMSKB	Y3, DX
+	BSFL	DX, DX
+	JNZ	fwd_avx2_success
+	ADDQ	$32, DI
+
+fwd_avx2_entry:
+	CMPQ	DI, AX
+	JB	fwd_avx2_loop
+
+	// Last 32-byte chunk (may overlap with previous)
+	MOVQ	AX, DI
+	VMOVDQU	(AX), Y2
+	VPCMPEQB	Y0, Y2, Y3
+	VPCMPEQB	Y1, Y2, Y4
+	VPOR	Y3, Y4, Y3
+	VPMOVMSKB	Y3, DX
+	BSFL	DX, DX
+	JNZ	fwd_avx2_success
+
+	MOVQ	$-1, (R8)
+	VZEROUPPER
+	RET
+
+fwd_avx2_success:
+	SUBQ	SI, DI
+	ADDQ	DX, DI
+	MOVQ	DI, (R8)
+	VZEROUPPER
+	RET
+
+	// ====== SSE2 forward search (< 32 bytes or no AVX2) ======
+
+fwd_sse2:
+	// Broadcast b1 into X0
+	MOVD	AX, X0
+	PUNPCKLBW	X0, X0
+	PUNPCKLBW	X0, X0
+	PSHUFL	$0, X0, X0
+
+	// Broadcast b2 into X4
+	MOVD	CX, X4
+	PUNPCKLBW	X4, X4
+	PUNPCKLBW	X4, X4
+	PSHUFL	$0, X4, X4
+
+	CMPQ	BX, $16
+	JLT	fwd_small
+
+	MOVQ	SI, DI
+	LEAQ	-16(SI)(BX*1), AX
+	JMP	fwd_sseloopentry
+
+fwd_sseloop:
+	MOVOU	(DI), X1
+	MOVOU	X1, X2
+	PCMPEQB	X0, X1
+	PCMPEQB	X4, X2
+	POR	X2, X1
+	PMOVMSKB	X1, DX
+	BSFL	DX, DX
+	JNZ	fwd_ssesuccess
+	ADDQ	$16, DI
+
+fwd_sseloopentry:
+	CMPQ	DI, AX
+	JB	fwd_sseloop
+
+	// Search the last 16-byte chunk (may overlap)
+	MOVQ	AX, DI
+	MOVOU	(AX), X1
+	MOVOU	X1, X2
+	PCMPEQB	X0, X1
+	PCMPEQB	X4, X2
+	POR	X2, X1
+	PMOVMSKB	X1, DX
+	BSFL	DX, DX
+	JNZ	fwd_ssesuccess
+
+fwd_failure:
+	MOVQ	$-1, (R8)
+	RET
+
+fwd_ssesuccess:
+	SUBQ	SI, DI
+	ADDQ	DX, DI
+	MOVQ	DI, (R8)
+	RET
+
+fwd_small:
+	// Check if loading 16 bytes from SI would cross a page boundary
+	LEAQ	16(SI), AX
+	TESTW	$0xff0, AX
+	JEQ	fwd_endofpage
+
+	MOVOU	(SI), X1
+	MOVOU	X1, X2
+	PCMPEQB	X0, X1
+	PCMPEQB	X4, X2
+	POR	X2, X1
+	PMOVMSKB	X1, DX
+	BSFL	DX, DX
+	JZ	fwd_failure
+	CMPL	DX, BX
+	JAE	fwd_failure
+	MOVQ	DX, (R8)
+	RET
+
+fwd_endofpage:
+	MOVOU	-16(SI)(BX*1), X1
+	MOVOU	X1, X2
+	PCMPEQB	X0, X1
+	PCMPEQB	X4, X2
+	POR	X2, X1
+	PMOVMSKB	X1, DX
+	MOVL	BX, CX
+	SHLL	CX, DX
+	SHRL	$16, DX
+	BSFL	DX, DX
+	JZ	fwd_failure
+	MOVQ	DX, (R8)
+	RET
+
+// func lastIndexByteTwo(s []byte, b1, b2 byte) int
+//
+// Returns the index of the last occurrence of b1 or b2 in s, or -1.
+// Uses AVX2 (32 bytes/iter) when available, SSE2 (16 bytes/iter) otherwise.
+TEXT ·lastIndexByteTwo(SB),NOSPLIT,$0-40
+	MOVQ	s_base+0(FP), SI
+	MOVQ	s_len+8(FP), BX
+	MOVBLZX	b1+24(FP), AX
+	MOVBLZX	b2+25(FP), CX
+	LEAQ	ret+32(FP), R8
+
+	TESTQ	BX, BX
+	JEQ	back_failure
+
+	// Try AVX2 for inputs >= 32 bytes
+	CMPQ	BX, $32
+	JLT	back_sse2
+	CMPB	·_useAVX2(SB), $1
+	JNE	back_sse2
+
+	// ====== AVX2 backward search ======
+	MOVD	AX, X0
+	VPBROADCASTB	X0, Y0
+	MOVD	CX, X1
+	VPBROADCASTB	X1, Y1
+
+	// DI = start of last 32-byte chunk
+	LEAQ	-32(SI)(BX*1), DI
+
+back_avx2_loop:
+	CMPQ	DI, SI
+	JBE	back_avx2_first
+
+	VMOVDQU	(DI), Y2
+	VPCMPEQB	Y0, Y2, Y3
+	VPCMPEQB	Y1, Y2, Y4
+	VPOR	Y3, Y4, Y3
+	VPMOVMSKB	Y3, DX
+	BSRL	DX, DX
+	JNZ	back_avx2_success
+	SUBQ	$32, DI
+	JMP	back_avx2_loop
+
+back_avx2_first:
+	// First 32 bytes (DI <= SI, load from SI)
+	VMOVDQU	(SI), Y2
+	VPCMPEQB	Y0, Y2, Y3
+	VPCMPEQB	Y1, Y2, Y4
+	VPOR	Y3, Y4, Y3
+	VPMOVMSKB	Y3, DX
+	BSRL	DX, DX
+	JNZ	back_avx2_firstsuccess
+
+	MOVQ	$-1, (R8)
+	VZEROUPPER
+	RET
+
+back_avx2_success:
+	SUBQ	SI, DI
+	ADDQ	DX, DI
+	MOVQ	DI, (R8)
+	VZEROUPPER
+	RET
+
+back_avx2_firstsuccess:
+	MOVQ	DX, (R8)
+	VZEROUPPER
+	RET
+
+	// ====== SSE2 backward search (< 32 bytes or no AVX2) ======
+
+back_sse2:
+	// Broadcast b1 into X0
+	MOVD	AX, X0
+	PUNPCKLBW	X0, X0
+	PUNPCKLBW	X0, X0
+	PSHUFL	$0, X0, X0
+
+	// Broadcast b2 into X4
+	MOVD	CX, X4
+	PUNPCKLBW	X4, X4
+	PUNPCKLBW	X4, X4
+	PSHUFL	$0, X4, X4
+
+	CMPQ	BX, $16
+	JLT	back_small
+
+	// DI = start of last 16-byte chunk
+	LEAQ	-16(SI)(BX*1), DI
+
+back_sseloop:
+	CMPQ	DI, SI
+	JBE	back_ssefirst
+
+	MOVOU	(DI), X1
+	MOVOU	X1, X2
+	PCMPEQB	X0, X1
+	PCMPEQB	X4, X2
+	POR	X2, X1
+	PMOVMSKB	X1, DX
+	BSRL	DX, DX
+	JNZ	back_ssesuccess
+	SUBQ	$16, DI
+	JMP	back_sseloop
+
+back_ssefirst:
+	// First 16 bytes (DI <= SI, load from SI)
+	MOVOU	(SI), X1
+	MOVOU	X1, X2
+	PCMPEQB	X0, X1
+	PCMPEQB	X4, X2
+	POR	X2, X1
+	PMOVMSKB	X1, DX
+	BSRL	DX, DX
+	JNZ	back_ssefirstsuccess
+
+back_failure:
+	MOVQ	$-1, (R8)
+	RET
+
+back_ssesuccess:
+	SUBQ	SI, DI
+	ADDQ	DX, DI
+	MOVQ	DI, (R8)
+	RET
+
+back_ssefirstsuccess:
+	// DX = byte offset from base
+	MOVQ	DX, (R8)
+	RET
+
+back_small:
+	// Check page boundary
+	LEAQ	16(SI), AX
+	TESTW	$0xff0, AX
+	JEQ	back_endofpage
+
+	MOVOU	(SI), X1
+	MOVOU	X1, X2
+	PCMPEQB	X0, X1
+	PCMPEQB	X4, X2
+	POR	X2, X1
+	PMOVMSKB	X1, DX
+	// Mask to first BX bytes: keep bits 0..BX-1
+	MOVL	$1, AX
+	MOVL	BX, CX
+	SHLL	CX, AX
+	DECL	AX
+	ANDL	AX, DX
+	BSRL	DX, DX
+	JZ	back_failure
+	MOVQ	DX, (R8)
+	RET
+
+back_endofpage:
+	// Load 16 bytes ending at base+n
+	MOVOU	-16(SI)(BX*1), X1
+	MOVOU	X1, X2
+	PCMPEQB	X0, X1
+	PCMPEQB	X4, X2
+	POR	X2, X1
+	PMOVMSKB	X1, DX
+	// Bits correspond to bytes [base+n-16, base+n).
+	// We want original bytes [0, n), which are bits [16-n, 16).
+	// Mask: keep bits (16-n) through 15.
+	MOVL	$16, CX
+	SUBL	BX, CX
+	SHRL	CX, DX
+	SHLL	CX, DX
+	BSRL	DX, DX
+	JZ	back_failure
+	// DX is the bit position in the loaded chunk.
+	// Original byte index = DX - (16 - n) = DX + n - 16
+	ADDL	BX, DX
+	SUBL	$16, DX
+	MOVQ	DX, (R8)
+	RET

src/algo/indexbyte2_arm64.go

@@ -0,0 +1,17 @@
+//go:build arm64
+
+package algo
+
+// indexByteTwo returns the index of the first occurrence of b1 or b2 in s,
+// or -1 if neither is present. Implemented in assembly using ARM64 NEON
+// to search for both bytes in a single pass.
+//
+//go:noescape
+func indexByteTwo(s []byte, b1, b2 byte) int
+
+// lastIndexByteTwo returns the index of the last occurrence of b1 or b2 in s,
+// or -1 if neither is present. Implemented in assembly using ARM64 NEON,
+// scanning backward.
+//
+//go:noescape
+func lastIndexByteTwo(s []byte, b1, b2 byte) int

src/algo/indexbyte2_arm64.s

@@ -0,0 +1,249 @@
+#include "textflag.h"
+
+// func indexByteTwo(s []byte, b1, b2 byte) int
+//
+// Returns the index of the first occurrence of b1 or b2 in s, or -1.
+// Uses ARM64 NEON to search for both bytes in a single pass over the data.
+// Adapted from Go's internal/bytealg/indexbyte_arm64.s (single-byte version).
+TEXT ·indexByteTwo(SB),NOSPLIT,$0-40
+	MOVD	s_base+0(FP), R0
+	MOVD	s_len+8(FP), R2
+	MOVBU	b1+24(FP), R1
+	MOVBU	b2+25(FP), R7
+	MOVD	$ret+32(FP), R8
+
+	// Core algorithm:
+	// For each 32-byte chunk we calculate a 64-bit syndrome value,
+	// with two bits per byte. We compare against both b1 and b2,
+	// OR the results, then use the same syndrome extraction as
+	// Go's IndexByte.
+
+	CBZ	R2, fail
+	MOVD	R0, R11
+	// Magic constant 0x40100401 allows us to identify which lane matches.
+	// Each byte in the group of 4 gets a distinct bit: 1, 4, 16, 64.
+	MOVD	$0x40100401, R5
+	VMOV	R1, V0.B16    // V0 = splat(b1)
+	VMOV	R7, V7.B16    // V7 = splat(b2)
+	// Work with aligned 32-byte chunks
+	BIC	$0x1f, R0, R3
+	VMOV	R5, V5.S4
+	ANDS	$0x1f, R0, R9
+	AND	$0x1f, R2, R10
+	BEQ	loop
+
+	// Input string is not 32-byte aligned. Process the first
+	// aligned 32-byte block and mask off bytes before our start.
+	VLD1.P	(R3), [V1.B16, V2.B16]
+	SUB	$0x20, R9, R4
+	ADDS	R4, R2, R2
+	// Compare against both needles
+	VCMEQ	V0.B16, V1.B16, V3.B16  // b1 vs first 16 bytes
+	VCMEQ	V7.B16, V1.B16, V8.B16  // b2 vs first 16 bytes
+	VORR	V8.B16, V3.B16, V3.B16  // combine
+	VCMEQ	V0.B16, V2.B16, V4.B16  // b1 vs second 16 bytes
+	VCMEQ	V7.B16, V2.B16, V9.B16  // b2 vs second 16 bytes
+	VORR	V9.B16, V4.B16, V4.B16  // combine
+	// Build syndrome
+	VAND	V5.B16, V3.B16, V3.B16
+	VAND	V5.B16, V4.B16, V4.B16
+	VADDP	V4.B16, V3.B16, V6.B16
+	VADDP	V6.B16, V6.B16, V6.B16
+	VMOV	V6.D[0], R6
+	// Clear the irrelevant lower bits
+	LSL	$1, R9, R4
+	LSR	R4, R6, R6
+	LSL	R4, R6, R6
+	// The first block can also be the last
+	BLS	masklast
+	// Have we found something already?
+	CBNZ	R6, tail
+
+loop:
+	VLD1.P	(R3), [V1.B16, V2.B16]
+	SUBS	$0x20, R2, R2
+	// Compare against both needles, OR results
+	VCMEQ	V0.B16, V1.B16, V3.B16
+	VCMEQ	V7.B16, V1.B16, V8.B16
+	VORR	V8.B16, V3.B16, V3.B16
+	VCMEQ	V0.B16, V2.B16, V4.B16
+	VCMEQ	V7.B16, V2.B16, V9.B16
+	VORR	V9.B16, V4.B16, V4.B16
+	// If we're out of data we finish regardless of the result
+	BLS	end
+	// Fast check: OR both halves and check for any match
+	VORR	V4.B16, V3.B16, V6.B16
+	VADDP	V6.D2, V6.D2, V6.D2
+	VMOV	V6.D[0], R6
+	CBZ	R6, loop
+
+end:
+	// Found something or out of data — build full syndrome
+	VAND	V5.B16, V3.B16, V3.B16
+	VAND	V5.B16, V4.B16, V4.B16
+	VADDP	V4.B16, V3.B16, V6.B16
+	VADDP	V6.B16, V6.B16, V6.B16
+	VMOV	V6.D[0], R6
+	// Only mask for the last block
+	BHS	tail
+
+masklast:
+	// Clear irrelevant upper bits
+	ADD	R9, R10, R4
+	AND	$0x1f, R4, R4
+	SUB	$0x20, R4, R4
+	NEG	R4<<1, R4
+	LSL	R4, R6, R6
+	LSR	R4, R6, R6
+
+tail:
+	CBZ	R6, fail
+	RBIT	R6, R6
+	SUB	$0x20, R3, R3
+	CLZ	R6, R6
+	ADD	R6>>1, R3, R0
+	SUB	R11, R0, R0
+	MOVD	R0, (R8)
+	RET
+
+fail:
+	MOVD	$-1, R0
+	MOVD	R0, (R8)
+	RET
+
+// func lastIndexByteTwo(s []byte, b1, b2 byte) int
+//
+// Returns the index of the last occurrence of b1 or b2 in s, or -1.
+// Scans backward using ARM64 NEON.
+TEXT ·lastIndexByteTwo(SB),NOSPLIT,$0-40
+	MOVD	s_base+0(FP), R0
+	MOVD	s_len+8(FP), R2
+	MOVBU	b1+24(FP), R1
+	MOVBU	b2+25(FP), R7
+	MOVD	$ret+32(FP), R8
+
+	CBZ	R2, lfail
+	MOVD	R0, R11          // save base
+	ADD	R0, R2, R12      // R12 = end = base + len
+	MOVD	$0x40100401, R5
+	VMOV	R1, V0.B16       // V0 = splat(b1)
+	VMOV	R7, V7.B16       // V7 = splat(b2)
+	VMOV	R5, V5.S4
+
+	// Align: find the aligned block containing the last byte
+	SUB	$1, R12, R3
+	BIC	$0x1f, R3, R3    // R3 = start of aligned block containing last byte
+
+	// --- Process tail block ---
+	VLD1	(R3), [V1.B16, V2.B16]
+	VCMEQ	V0.B16, V1.B16, V3.B16
+	VCMEQ	V7.B16, V1.B16, V8.B16
+	VORR	V8.B16, V3.B16, V3.B16
+	VCMEQ	V0.B16, V2.B16, V4.B16
+	VCMEQ	V7.B16, V2.B16, V9.B16
+	VORR	V9.B16, V4.B16, V4.B16
+	VAND	V5.B16, V3.B16, V3.B16
+	VAND	V5.B16, V4.B16, V4.B16
+	VADDP	V4.B16, V3.B16, V6.B16
+	VADDP	V6.B16, V6.B16, V6.B16
+	VMOV	V6.D[0], R6
+
+	// Mask upper bits (bytes past end of slice)
+	// tail_bytes = end - R3 (1..32)
+	SUB	R3, R12, R10     // R10 = tail_bytes
+	MOVD	$64, R4
+	SUB	R10<<1, R4, R4   // R4 = 64 - 2*tail_bytes
+	LSL	R4, R6, R6
+	LSR	R4, R6, R6
+
+	// Is this also the head block?
+	CMP	R11, R3          // R3 - R11
+	BLO	lmaskfirst       // R3 < base: head+tail in same block
+	BEQ	ltailonly         // R3 == base: single aligned block
+
+	// R3 > base: more blocks before this one
+	CBNZ	R6, llast
+	B	lbacksetup
+
+ltailonly:
+	// Single block, already masked upper bits
+	CBNZ	R6, llast
+	B	lfail
+
+lmaskfirst:
+	// Mask lower bits (bytes before start of slice)
+	SUB	R3, R11, R4      // R4 = base - R3
+	LSL	$1, R4, R4
+	LSR	R4, R6, R6
+	LSL	R4, R6, R6
+	CBNZ	R6, llast
+	B	lfail
+
+lbacksetup:
+	SUB	$0x20, R3
+
+lbackloop:
+	VLD1	(R3), [V1.B16, V2.B16]
+	VCMEQ	V0.B16, V1.B16, V3.B16
+	VCMEQ	V7.B16, V1.B16, V8.B16
+	VORR	V8.B16, V3.B16, V3.B16
+	VCMEQ	V0.B16, V2.B16, V4.B16
+	VCMEQ	V7.B16, V2.B16, V9.B16
+	VORR	V9.B16, V4.B16, V4.B16
+	// Quick check: any match in this block?
+	VORR	V4.B16, V3.B16, V6.B16
+	VADDP	V6.D2, V6.D2, V6.D2
+	VMOV	V6.D[0], R6
+
+	// Is this a head block? (R3 < base)
+	CMP	R11, R3
+	BLO	lheadblock
+
+	// Full block (R3 >= base)
+	CBNZ	R6, lbackfound
+	// More blocks?
+	BEQ	lfail            // R3 == base, no more
+	SUB	$0x20, R3
+	B	lbackloop
+
+lbackfound:
+	// Build full syndrome
+	VAND	V5.B16, V3.B16, V3.B16
+	VAND	V5.B16, V4.B16, V4.B16
+	VADDP	V4.B16, V3.B16, V6.B16
+	VADDP	V6.B16, V6.B16, V6.B16
+	VMOV	V6.D[0], R6
+	B	llast
+
+lheadblock:
+	// R3 < base. Build full syndrome if quick check had a match.
+	CBZ	R6, lfail
+	VAND	V5.B16, V3.B16, V3.B16
+	VAND	V5.B16, V4.B16, V4.B16
+	VADDP	V4.B16, V3.B16, V6.B16
+	VADDP	V6.B16, V6.B16, V6.B16
+	VMOV	V6.D[0], R6
+	// Mask lower bits
+	SUB	R3, R11, R4      // R4 = base - R3
+	LSL	$1, R4, R4
+	LSR	R4, R6, R6
+	LSL	R4, R6, R6
+	CBZ	R6, lfail
+
+llast:
+	// Find last match: highest set bit in syndrome
+	// Syndrome has bit 2i set for matching byte i.
+	// CLZ gives leading zeros; byte_offset = (63 - CLZ) / 2.
+	CLZ	R6, R6
+	MOVD	$63, R4
+	SUB	R6, R4, R6       // R6 = 63 - CLZ = bit position
+	LSR	$1, R6            // R6 = byte offset within block
+	ADD	R3, R6, R0        // R0 = absolute address
+	SUB	R11, R0, R0       // R0 = slice index
+	MOVD	R0, (R8)
+	RET
+
+lfail:
+	MOVD	$-1, R0
+	MOVD	R0, (R8)
+	RET

src/algo/indexbyte2_other.go

@@ -0,0 +1,33 @@
+//go:build !arm64 && !amd64
+
+package algo
+
+import "bytes"
+
+// indexByteTwo returns the index of the first occurrence of b1 or b2 in s,
+// or -1 if neither is present.
+func indexByteTwo(s []byte, b1, b2 byte) int {
+	i1 := bytes.IndexByte(s, b1)
+	if i1 == 0 {
+		return 0
+	}
+	scope := s
+	if i1 > 0 {
+		scope = s[:i1]
+	}
+	if i2 := bytes.IndexByte(scope, b2); i2 >= 0 {
+		return i2
+	}
+	return i1
+}
+
+// lastIndexByteTwo returns the index of the last occurrence of b1 or b2 in s,
+// or -1 if neither is present.
+func lastIndexByteTwo(s []byte, b1, b2 byte) int {
+	for i := len(s) - 1; i >= 0; i-- {
+		if s[i] == b1 || s[i] == b2 {
+			return i
+		}
+	}
+	return -1
+}

src/algo/indexbyte2_test.go

@@ -0,0 +1,259 @@
+package algo
+
+import (
+	"bytes"
+	"testing"
+)
+
+func TestIndexByteTwo(t *testing.T) {
+	tests := []struct {
+		name string
+		s    string
+		b1   byte
+		b2   byte
+		want int
+	}{
+		{"empty", "", 'a', 'b', -1},
+		{"single_b1", "a", 'a', 'b', 0},
+		{"single_b2", "b", 'a', 'b', 0},
+		{"single_none", "c", 'a', 'b', -1},
+		{"b1_first", "xaxb", 'a', 'b', 1},
+		{"b2_first", "xbxa", 'a', 'b', 1},
+		{"same_byte", "xxa", 'a', 'a', 2},
+		{"at_end", "xxxxa", 'a', 'b', 4},
+		{"not_found", "xxxxxxxx", 'a', 'b', -1},
+		{"long_b1_at_3000", string(make([]byte, 3000)) + "a" + string(make([]byte, 1000)), 'a', 'b', 3000},
+		{"long_b2_at_3000", string(make([]byte, 3000)) + "b" + string(make([]byte, 1000)), 'a', 'b', 3000},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := indexByteTwo([]byte(tt.s), tt.b1, tt.b2)
+			if got != tt.want {
+				t.Errorf("indexByteTwo(%q, %c, %c) = %d, want %d", tt.s[:min(len(tt.s), 40)], tt.b1, tt.b2, got, tt.want)
+			}
+		})
+	}
+
+	// Exhaustive test: compare against loop reference for various lengths,
+	// including sizes around SIMD block boundaries (16, 32, 64).
+	for n := 0; n <= 256; n++ {
+		data := make([]byte, n)
+		for i := range data {
+			data[i] = byte('c' + (i % 20))
+		}
+		// Test with match at every position
+		for pos := 0; pos < n; pos++ {
+			for _, b := range []byte{'A', 'B'} {
+				data[pos] = b
+				got := indexByteTwo(data, 'A', 'B')
+				want := loopIndexByteTwo(data, 'A', 'B')
+				if got != want {
+					t.Fatalf("indexByteTwo(len=%d, match=%c@%d) = %d, want %d", n, b, pos, got, want)
+				}
+				data[pos] = byte('c' + (pos % 20))
+			}
+		}
+		// Test with no match
+		got := indexByteTwo(data, 'A', 'B')
+		if got != -1 {
+			t.Fatalf("indexByteTwo(len=%d, no match) = %d, want -1", n, got)
+		}
+		// Test with both bytes present
+		if n >= 2 {
+			data[n/3] = 'A'
+			data[n*2/3] = 'B'
+			got := indexByteTwo(data, 'A', 'B')
+			want := loopIndexByteTwo(data, 'A', 'B')
+			if got != want {
+				t.Fatalf("indexByteTwo(len=%d, both@%d,%d) = %d, want %d", n, n/3, n*2/3, got, want)
+			}
+			data[n/3] = byte('c' + ((n / 3) % 20))
+			data[n*2/3] = byte('c' + ((n * 2 / 3) % 20))
+		}
+	}
+}
+
+func TestLastIndexByteTwo(t *testing.T) {
+	tests := []struct {
+		name string
+		s    string
+		b1   byte
+		b2   byte
+		want int
+	}{
+		{"empty", "", 'a', 'b', -1},
+		{"single_b1", "a", 'a', 'b', 0},
+		{"single_b2", "b", 'a', 'b', 0},
+		{"single_none", "c", 'a', 'b', -1},
+		{"b1_last", "xbxa", 'a', 'b', 3},
+		{"b2_last", "xaxb", 'a', 'b', 3},
+		{"same_byte", "axx", 'a', 'a', 0},
+		{"at_start", "axxxx", 'a', 'b', 0},
+		{"both_present", "axbx", 'a', 'b', 2},
+		{"not_found", "xxxxxxxx", 'a', 'b', -1},
+		{"long_b1_at_3000", string(make([]byte, 3000)) + "a" + string(make([]byte, 1000)), 'a', 'b', 3000},
+		{"long_b2_at_end", string(make([]byte, 4000)) + "b", 'a', 'b', 4000},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := lastIndexByteTwo([]byte(tt.s), tt.b1, tt.b2)
+			if got != tt.want {
+				t.Errorf("lastIndexByteTwo(%q, %c, %c) = %d, want %d", tt.s[:min(len(tt.s), 40)], tt.b1, tt.b2, got, tt.want)
+			}
+		})
+	}
+
+	// Exhaustive test against loop reference
+	for n := 0; n <= 256; n++ {
+		data := make([]byte, n)
+		for i := range data {
+			data[i] = byte('c' + (i % 20))
+		}
+		for pos := 0; pos < n; pos++ {
+			for _, b := range []byte{'A', 'B'} {
+				data[pos] = b
+				got := lastIndexByteTwo(data, 'A', 'B')
+				want := refLastIndexByteTwo(data, 'A', 'B')
+				if got != want {
+					t.Fatalf("lastIndexByteTwo(len=%d, match=%c@%d) = %d, want %d", n, b, pos, got, want)
+				}
+				data[pos] = byte('c' + (pos % 20))
+			}
+		}
+		// No match
+		got := lastIndexByteTwo(data, 'A', 'B')
+		if got != -1 {
+			t.Fatalf("lastIndexByteTwo(len=%d, no match) = %d, want -1", n, got)
+		}
+		// Both bytes present
+		if n >= 2 {
+			data[n/3] = 'A'
+			data[n*2/3] = 'B'
+			got := lastIndexByteTwo(data, 'A', 'B')
+			want := refLastIndexByteTwo(data, 'A', 'B')
+			if got != want {
+				t.Fatalf("lastIndexByteTwo(len=%d, both@%d,%d) = %d, want %d", n, n/3, n*2/3, got, want)
+			}
+			data[n/3] = byte('c' + ((n / 3) % 20))
+			data[n*2/3] = byte('c' + ((n * 2 / 3) % 20))
+		}
+	}
+}
+
+func FuzzIndexByteTwo(f *testing.F) {
+	f.Add([]byte("hello world"), byte('o'), byte('l'))
+	f.Add([]byte(""), byte('a'), byte('b'))
+	f.Add([]byte("aaa"), byte('a'), byte('a'))
+	f.Fuzz(func(t *testing.T, data []byte, b1, b2 byte) {
+		got := indexByteTwo(data, b1, b2)
+		want := loopIndexByteTwo(data, b1, b2)
+		if got != want {
+			t.Errorf("indexByteTwo(len=%d, b1=%d, b2=%d) = %d, want %d", len(data), b1, b2, got, want)
+		}
+	})
+}
+
+func FuzzLastIndexByteTwo(f *testing.F) {
+	f.Add([]byte("hello world"), byte('o'), byte('l'))
+	f.Add([]byte(""), byte('a'), byte('b'))
+	f.Add([]byte("aaa"), byte('a'), byte('a'))
+	f.Fuzz(func(t *testing.T, data []byte, b1, b2 byte) {
+		got := lastIndexByteTwo(data, b1, b2)
+		want := refLastIndexByteTwo(data, b1, b2)
+		if got != want {
+			t.Errorf("lastIndexByteTwo(len=%d, b1=%d, b2=%d) = %d, want %d", len(data), b1, b2, got, want)
+		}
+	})
+}
+
+// Reference implementations for correctness checking
+func refIndexByteTwo(s []byte, b1, b2 byte) int {
+	i1 := bytes.IndexByte(s, b1)
+	if i1 == 0 {
+		return 0
+	}
+	scope := s
+	if i1 > 0 {
+		scope = s[:i1]
+	}
+	if i2 := bytes.IndexByte(scope, b2); i2 >= 0 {
+		return i2
+	}
+	return i1
+}
+
+func loopIndexByteTwo(s []byte, b1, b2 byte) int {
+	for i, b := range s {
+		if b == b1 || b == b2 {
+			return i
+		}
+	}
+	return -1
+}
+
+func refLastIndexByteTwo(s []byte, b1, b2 byte) int {
+	for i := len(s) - 1; i >= 0; i-- {
+		if s[i] == b1 || s[i] == b2 {
+			return i
+		}
+	}
+	return -1
+}
+
+func benchIndexByteTwo(b *testing.B, size int, pos int) {
+	data := make([]byte, size)
+	for i := range data {
+		data[i] = byte('a' + (i % 20))
+	}
+	data[pos] = 'Z'
+
+	type impl struct {
+		name string
+		fn   func([]byte, byte, byte) int
+	}
+	impls := []impl{
+		{"asm", indexByteTwo},
+		{"2xIndexByte", refIndexByteTwo},
+		{"loop", loopIndexByteTwo},
+	}
+	for _, im := range impls {
+		b.Run(im.name, func(b *testing.B) {
+			for i := 0; i < b.N; i++ {
+				im.fn(data, 'Z', 'z')
+			}
+		})
+	}
+}
+
+func benchLastIndexByteTwo(b *testing.B, size int, pos int) {
+	data := make([]byte, size)
+	for i := range data {
+		data[i] = byte('a' + (i % 20))
+	}
+	data[pos] = 'Z'
+
+	type impl struct {
+		name string
+		fn   func([]byte, byte, byte) int
+	}
+	impls := []impl{
+		{"asm", lastIndexByteTwo},
+		{"loop", refLastIndexByteTwo},
+	}
+	for _, im := range impls {
+		b.Run(im.name, func(b *testing.B) {
+			for i := 0; i < b.N; i++ {
+				im.fn(data, 'Z', 'z')
+			}
+		})
+	}
+}
+
+func BenchmarkIndexByteTwo_10(b *testing.B)       { benchIndexByteTwo(b, 10, 8) }
+func BenchmarkIndexByteTwo_100(b *testing.B)      { benchIndexByteTwo(b, 100, 80) }
+func BenchmarkIndexByteTwo_1000(b *testing.B)     { benchIndexByteTwo(b, 1000, 800) }
+func BenchmarkLastIndexByteTwo_10(b *testing.B)   { benchLastIndexByteTwo(b, 10, 2) }
+func BenchmarkLastIndexByteTwo_100(b *testing.B)  { benchLastIndexByteTwo(b, 100, 20) }
+func BenchmarkLastIndexByteTwo_1000(b *testing.B) { benchLastIndexByteTwo(b, 1000, 200) }