runtime/cgo: fix darwin/amd64 signal handling setup

Was not allocating space for the frame above sigpanic,
nor was it pushing the LR into the right place.
Because traceback past sigpanic only needs the
LR for faulting leaves, this was not noticed too much.
But it did break the sync/atomic nil deref tests.

Change-Id: Icba53fffa193423aab744c37f21ee893ce2ee3ac
Reviewed-on: https://go-review.googlesource.com/12926Reviewed-by: David Crawshaw <crawshaw@golang.org>

src/runtime/cgo/signal_darwin_arm64.s

@@ -18,8 +18,9 @@ TEXT ·panicmem(SB),NOSPLIT,$-8
 
 	// On a foreign thread.
 	// TODO(crawshaw): call badsignal
+	MOVD.W $0, -16(RSP)
 	MOVW $139, R1
-	MOVW R1, (RSP)
+	MOVW R1, 8(RSP)
 	B    runtime·exit(SB)
 
 ongothread:
@@ -33,10 +34,23 @@ ongothread:
 	// To do this we call into runtime·setsigsegv, which sets the
 	// appropriate state inside the g object. We give it the faulting
 	// PC on the stack, then put it in the LR before calling sigpanic.
-	STP.W (R1, R2), -16(RSP)
-	BL runtime·setsigsegv(SB)
-	LDP.P 16(RSP), (R1, R2)
 
+	// Build a 32-byte stack frame for us for this call.
+	// Saved LR (none available) is at the bottom,
+	// then the PC argument for setsigsegv, 
+	// then a copy of the LR for us to restore.
+	MOVD.W $0, -32(RSP)
 	MOVD R1, 8(RSP)
-	MOVD R2, R30 // link register
+	MOVD R2, 16(RSP)
+	BL runtime·setsigsegv(SB)
+	MOVD 8(RSP), R1
+	MOVD 16(RSP), R2
+
+	// Build a 16-byte stack frame for the simulated
+	// call to sigpanic, by taking 16 bytes away from the
+	// 32-byte stack frame above.
+	// The saved LR in this frame is the LR at time of fault,
+	// and the LR on entry to sigpanic is the PC at time of fault.
+	MOVD.W R1, 16(RSP)
+	MOVD R2, R30
 	B runtime·sigpanic(SB)