test(verify): prove the arm64 lse atomics through the link-parity kernel

Assisted-by: GLM 5.3 Flash

verify/testdata/linkatomic_arm64.s

@@ -0,0 +1,59 @@
+// 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