cmd/compile: update comment about x86 nop instruction generator

The comment about losing the high bits is incorrect.  We now use these
nops in places where they really need to be a nop.  (Before inline
marks, we used them just before deferreturn calls, so they could
clobber any caller-saved values.)

Change-Id: I433d1ec455aa37dab8fef6eb7d407f3737dbb97f
Reviewed-on: https://go-review.googlesource.com/c/158057
Reviewed-by: Ilya Tocar <ilya.tocar@intel.com>

src/cmd/compile/internal/amd64/ggen.go

@@ -142,9 +142,12 @@ func zeroAuto(pp *gc.Progs, n *gc.Node) {
 }
 
 func ginsnop(pp *gc.Progs) *obj.Prog {
-	// This is actually not the x86 NOP anymore,
-	// but at the point where it gets used, AX is dead
-	// so it's okay if we lose the high bits.
+	// This is a hardware nop (1-byte 0x90) instruction,
+	// even though we describe it as an explicit XCHGL here.
+	// Particularly, this does not zero the high 32 bits
+	// like typical *L opcodes.
+	// (gas assembles "xchg %eax,%eax" to 0x87 0xc0, which
+	// does zero the high 32 bits.)
 	p := pp.Prog(x86.AXCHGL)
 	p.From.Type = obj.TYPE_REG
 	p.From.Reg = x86.REG_AX

src/cmd/compile/internal/x86/ggen.go

@@ -54,6 +54,7 @@ func zeroAuto(pp *gc.Progs, n *gc.Node) {
 }
 
 func ginsnop(pp *gc.Progs) *obj.Prog {
+	// See comment in ../amd64/ggen.go.
 	p := pp.Prog(x86.AXCHGL)
 	p.From.Type = obj.TYPE_REG
 	p.From.Reg = x86.REG_AX