// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
// SPDX-License-Identifier: BSD-3-Clause

#include "textflag.h"

// The link-regression kernel for the amd64 SSE families: fold XOR-folds
// every 16-byte block of a buffer into one octa with MOVOU and PXOR,
// folds any remaining bytes with a scalar loop, combines the two octa
// lanes with PEXTRQ, passes the total to mixq through the ABI0 stack slot
// and XORs the seed into the returned mix.  mixq multiplies by the odd
// golden-ratio constant, the same arithmetic the Go side of the regression
// mirrors bit for bit.  The run claim is native: the host is amd64.

// func mixq(x uint64) uint64
TEXT ·mixq(SB), NOSPLIT, $0-16
	MOVQ  x+0(FP), AX
	MOVQ  $0x9e3779b97f4a7c15, CX
	IMULQ CX, AX
	MOVQ  AX, ret+8(FP)
	RET

// func fold(buf []byte, seed uint64) uint64
TEXT ·fold(SB), NOSPLIT, $16-40
	MOVQ    buf+0(FP), SI
	MOVQ    buf_len+8(FP), CX
	MOVQ    seed+24(FP), BX
	PXOR    X1, X1
	XORQ    AX, AX

blocks:
	CMPQ    CX, $16
	JB      tail
	MOVOU   (SI), X2
	PXOR    X2, X1
	ADDQ    $16, SI
	SUBQ    $16, CX
	JMP     blocks

tail:
	TESTQ   CX, CX
	JE      gathered
	MOVBQZX (SI), DX
	ADDQ    DX, AX
	ADDQ    $1, SI
	SUBQ    $1, CX
	JMP     tail

gathered:
	MOVQ    X1, DX
	PEXTRQ  $1, X1, CX
	ADDQ    CX, DX
	ADDQ    AX, DX
	MOVQ    DX, 0(SP)
	CALL    ·mixq(SB)
	MOVQ    8(SP), DX
	XORQ    BX, DX
	MOVQ    DX, ret+32(FP)
	RET