runtime: when dying from a signal use the previous signal handler

Before this CL, whenever the Go runtime wanted to kill its own
process with a signal dieFromSignal would reset the signal handler
to _SIG_DFL.

Unfortunately, if any signal handler were installed before the Go
runtime initialized, it wouldn't be invoked either.

Instead, use whatever signal handler was installed before
initialization.

The motivating use case is Crashlytics on Android. Before this CL,
Crashlytics would not consider a crash from a panic() since the
corresponding SIGABRT never reached its signal handler.

Updates #11382
Updates #20392 (perhaps even fixes it)
Fixes #19389

Change-Id: I0c8633329433b45cbb3b16571bea227e38e8be2e
Reviewed-on: https://go-review.googlesource.com/49590
Run-TryBot: Elias Naur <elias.naur@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>

src/runtime/signal_unix.go

@@ -394,8 +394,11 @@ func sigpanic() {
 //go:nosplit
 //go:nowritebarrierrec
 func dieFromSignal(sig uint32) {
-	setsig(sig, _SIG_DFL)
 	unblocksig(sig)
+	// First, try any signal handler installed before the runtime
+	// initialized.
+	fn := atomic.Loaduintptr(&fwdSig[sig])
+	setsig(sig, fn)
 	raise(sig)
 
 	// That should have killed us. On some systems, though, raise
@@ -407,6 +410,14 @@ func dieFromSignal(sig uint32) {
 	osyield()
 	osyield()
 
+	// If that didn't work, try _SIG_DFL.
+	setsig(sig, _SIG_DFL)
+	raise(sig)
+
+	osyield()
+	osyield()
+	osyield()
+
 	// If we are still somehow running, just exit with the wrong status.
 	exit(2)
 }