DSE patch RFA: Don't delete trapping insn

DSE patch RFA: Don't delete trapping insn

[Ian Lance Taylor]

[-- Attachment #1: Type: text/plain, Size: 427 bytes --]

The DSE pass can delete a dead store even if the instruction can trap.
That is incorrect when using -fnon-call-exceptions
-fno-delete-dead-exceptions.  This led to a bug report against gccgo:
https://go.dev/issue/53012.  However, the bug is not specific to Go.

This patch fixes the problem in a simple way, and includes a C++
testcase.  Bootstrapped and ran C, C++, and Go tests on
x86_64-pc-linux-gnu.

OK for mainline?

Ian

[-- Attachment #2: patch.txt --]
[-- Type: text/plain, Size: 2210 bytes --]

From bae426745756896ec0ea27c9e12469c53b88d538 Mon Sep 17 00:00:00 2001
From: Ian Lance Taylor <iant@golang.org>
Date: Fri, 1 Jul 2022 14:51:45 -0700
Subject: [PATCH] tree-optimization: only DSE trapping insn if
 -fdelete-dead-exceptions

gcc/ChangeLog:

	* tree-ssa-dse.cc (dse_optimize_stmt): Only delete a trapping
	statement if -fdelete-dead-exceptions.

gcc/testsuite/ChangeLog:

	* g++.dg/torture/except-1.C: New test.
---
 gcc/testsuite/g++.dg/torture/except-1.C | 44 +++++++++++++++++++++++++
 gcc/tree-ssa-dse.cc                     |  3 +-
 2 files changed, 46 insertions(+), 1 deletion(-)
 create mode 100644 gcc/testsuite/g++.dg/torture/except-1.C

diff --git a/gcc/testsuite/g++.dg/torture/except-1.C b/gcc/testsuite/g++.dg/torture/except-1.C
new file mode 100644
index 00000000000..7050a33cc27
--- /dev/null
+++ b/gcc/testsuite/g++.dg/torture/except-1.C
@@ -0,0 +1,44 @@
+// { dg-do run { target { i?86-*-linux* i?86-*-gnu* x86_64-*-linux* } } }
+// { dg-additional-options "-fexceptions -fnon-call-exceptions -fno-delete-dead-exceptions" }
+
+#include <signal.h>
+#include <stdlib.h>
+#include <string.h>
+
+static void
+sighandler (int signo, siginfo_t* si, void* uc)
+{
+  throw (5);
+}
+
+struct S { void *p1, *p2; };
+
+struct S v;
+
+__attribute__ ((noinline))
+int
+dosegv ()
+{
+  struct S *p = 0;
+  struct S s __attribute__((unused)) = *p;
+  return 0;
+}
+
+int main ()
+{
+  struct sigaction sa;
+
+  memset (&sa, 0, sizeof sa);
+  sa.sa_sigaction = sighandler;
+  sigaction (SIGSEGV, &sa, NULL);
+  sigaction (SIGBUS, &sa, NULL);
+
+  try {
+    dosegv ();
+  }
+  catch (int x) {
+    return (x != 5);
+  }
+
+  return 1;
+}
diff --git a/gcc/tree-ssa-dse.cc b/gcc/tree-ssa-dse.cc
index 62efafe384d..8d1739a4510 100644
--- a/gcc/tree-ssa-dse.cc
+++ b/gcc/tree-ssa-dse.cc
@@ -1463,7 +1463,8 @@ dse_optimize_stmt (function *fun, gimple_stmt_iterator *gsi, sbitmap live_bytes)
       gimple_call_set_lhs (stmt, NULL_TREE);
       update_stmt (stmt);
     }
-  else
+  else if (!stmt_could_throw_p (fun, stmt)
+	   || fun->can_delete_dead_exceptions)
     delete_dead_or_redundant_assignment (gsi, "dead", need_eh_cleanup,
 					 need_ab_cleanup);
 }
-- 
2.37.0.rc0.161.g10f37bed90-goog

Re: DSE patch RFA: Don't delete trapping insn

[Jeff Law]

On 7/1/2022 4:04 PM, Ian Lance Taylor via Gcc-patches wrote:
> The DSE pass can delete a dead store even if the instruction can trap.
> That is incorrect when using -fnon-call-exceptions
> -fno-delete-dead-exceptions.  This led to a bug report against gccgo:
> https://go.dev/issue/53012.  However, the bug is not specific to Go.
>
> This patch fixes the problem in a simple way, and includes a C++
> testcase.  Bootstrapped and ran C, C++, and Go tests on
> x86_64-pc-linux-gnu.
>
> OK for mainline?
OK
jeff