On Sat, Feb 02, 2019 at 11:22:55AM +0100, Jakub Jelinek wrote:
> The only "regression" was gcc.target/i386/call-1.c with
> -fstack-protector-strong, but that is because the test is invalid:
> void set_eax(int val)
> {
> __asm__ __volatile__ ("mov %0, %%eax" : : "m" (val));
> }
> - missing "eax" clobber or "=a" (dummy) output and when set_eax is inlined
> that can break optimizations badly. Of course, in addition to fixing that,
> I'd expect if the tests wants to test what it originally wanted to test, it
> needs to disable inlining or perhaps all IPA opts, not sure if just for
> set_eax or also for foo/bar. Perhaps we can keep the testcase as is with
> the "eax" clobber and add another one with __attribute__((noipa)) on
> set_eax/foo/bar.
Regardless of the PR87485 decision, I think we should fix this testcase.
So here it is in patch form, regtested on x86_64-linux
(\{-m32,-m32/-fstack-protector-strong,-m64,-m64/-fstack-protector-strong\}),
ok for trunk?
2019-02-02 Jakub Jelinek <ja...@redhat.com>
PR rtl-optimization/11304
* gcc.target/i386/call-1.c (set_eax): Add "eax" clobber.
* gcc.target/i386/call-2.c: New test.
--- gcc/testsuite/gcc.target/i386/call-1.c.jj 2008-09-05 12:54:23.000000000
+0200
+++ gcc/testsuite/gcc.target/i386/call-1.c 2019-02-02 11:23:25.566902736
+0100
@@ -11,7 +11,7 @@ volatile int r;
void set_eax(int val)
{
- __asm__ __volatile__ ("mov %0, %%eax" : : "m" (val));
+ __asm__ __volatile__ ("mov %0, %%eax" : : "m" (val) : "eax");
}
void foo(int val)
--- gcc/testsuite/gcc.target/i386/call-2.c.jj 2019-02-02 11:23:31.922797178
+0100
+++ gcc/testsuite/gcc.target/i386/call-2.c 2019-02-02 11:23:47.757534186
+0100
@@ -0,0 +1,12 @@
+/* PR optimization/11304 */
+/* Originator: <manuel.serr...@sophia.inria.fr> */
+/* { dg-do run } */
+/* { dg-options "-O -fomit-frame-pointer" } */
+
+/* Verify that %eax is always restored after a call. */
+
+__attribute__((noipa)) void set_eax(int val);
+__attribute__((noipa)) void foo(int val);
+__attribute__((noipa)) int bar(int x);
+
+#include "call-1.c"
Jakub
