// 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 arm64 LSE atomics: atomicsum adds a
// value to a counter with LDADDALD in a loop, accumulating every old value
// the atomic returns, and finishes with the counter's final contents, so
// the result is addv times times times times plus one under any schedule.
// caller passes the pointer and the two scalars to atomicsum and mixes the
// total through the intra-file mixa call over the ABI0 stack slots.  mixa
// multiplies by the odd golden-ratio constant, the same arithmetic the Go
// side of the regression mirrors bit for bit.  The run claim follows the
// documented degradation: a qemu-user without LSE fails the baseline and
// the claim degrades to link-only.

// func atomicsum(p *int64, addv int64, times int64) int64
TEXT ·atomicsum(SB), NOSPLIT, $0-32
	MOVD     p+0(FP), R3
	MOVD     addv+8(FP), R4
	MOVD     times+16(FP), R5
	MOVD     $0, R6

loop:
	CBZ      R5, done
	LDADDALD R4, (R3), R7
	ADD      R7, R6, R6
	SUB      $1, R5, R5
	JMP      loop

done:
	MOVD     (R3), R8
	ADD      R8, R6, R6
	MOVD     R6, ret+24(FP)
	RET

// func mixa(x int64) int64
TEXT ·mixa(SB), NOSPLIT, $0-16
	MOVD x+0(FP), R0
	MOVD $0x9e3779b97f4a7c15, R1
	MUL  R1, R0, R0
	MOVD R0, ret+8(FP)
	RET

// func caller(p *int64, addv int64, times int64) int64
TEXT ·caller(SB), NOSPLIT, $32-32
	MOVD p+0(FP), R3
	MOVD addv+8(FP), R4
	MOVD times+16(FP), R5
	MOVD R3, 8(RSP)
	MOVD R4, 16(RSP)
	MOVD R5, 24(RSP)
	BL   ·atomicsum(SB)
	MOVD 32(RSP), R0
	MOVD R0, 8(RSP)
	BL   ·mixa(SB)
	MOVD 16(RSP), R0
	MOVD R0, ret+24(FP)
	RET