cmd/compile: fix scheduling of memory-producing tuple ops

Intrinsified atomic op produces <value,memory>. Make sure this
memory is considered in the store chain calculation.

Fixes #16948.

Change-Id: I029f164b123a7e830214297f8373f06ea0bf1e26
Reviewed-on: https://go-review.googlesource.com/28350
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>

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

@@ -128,9 +128,13 @@ func schedule(f *Func) {
 		// the calculated store chain is good only for this block.
 		for _, v := range b.Values {
 			if v.Op != OpPhi && v.Type.IsMemory() {
+				mem := v
+				if v.Op == OpSelect1 {
+					v = v.Args[0]
+				}
 				for _, w := range v.Args {
 					if w.Type.IsMemory() {
-						nextMem[w.ID] = v
+						nextMem[w.ID] = mem
 					}
 				}
 			}

test/fixedbugs/issue16948.go

+// run
+
+// Copyright 2016 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.
+
+// issue 16948: make sure intrinsified atomic ops won't
+// confuse the scheduler.
+
+package main
+
+import "sync/atomic"
+
+func main() {
+	f()
+}
+
+var x int32
+
+type T [10]int
+var sink *T
+
+func f() (t T) {
+	atomic.AddInt32(&x, 1)
+	g(42, 42, 42, 42, 42, &t) // use int values that is invalid pointer to smash the stack slot of return value of runtime.newobject
+	return
+}
+
+//go:noinline
+func g(a, b, c, d, e int, p *T) {
+	var t [10000]int // a large stack frame to trigger stack growing
+	_ = t
+	sink = p // force p (in caller) heap allocated
+}