cmd/compile: ensure we don't reuse temporary register

Before this CL, we could use the same register for both a temporary
register and for moving a value in the output register out of the way.

Fixes #71857

Change-Id: Iefbfd9d4139136174570d8aadf8a0fb391791ea9
Reviewed-on: https://go-review.googlesource.com/c/go/+/651221
Reviewed-by: David Chase <drchase@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Keith Randall <khr@google.com>

src/cmd/compile/internal/ssa/regalloc.go

@@ -1677,6 +1677,7 @@ func (s *regAllocState) regalloc(f *Func) {
 				}
 				tmpReg = s.allocReg(m, &tmpVal)
 				s.nospill |= regMask(1) << tmpReg
+				s.tmpused |= regMask(1) << tmpReg
 			}
 
 			// Now that all args are in regs, we're ready to issue the value itself.

test/fixedbugs/issue71857.go

@@ -0,0 +1,29 @@
+// run
+
+// Copyright 2025 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package main
+
+import "sync/atomic"
+
+//go:noinline
+func f(p0, p1, p2, p3, p4, p5, p6, p7 *uint64, a *atomic.Uint64) {
+	old := a.Or(0xaaa)
+	*p0 = old
+	*p1 = old
+	*p2 = old
+	*p3 = old
+	*p4 = old
+	*p5 = old
+	*p6 = old
+	*p7 = old
+}
+
+func main() {
+	a := new(atomic.Uint64)
+	p := new(uint64)
+	f(p, p, p, p, p, p, p, p, a)
+
+}