Hi!
IMNSHO we want .cfi_restore rbp on leave insn even when that insn isn't
setting cfa (on the shrink-wrapped testcase below already the previous
load of %r10 sets cfa, but there is no .cfi_restore rbp and we ICE when
processing the merge label to which pre-prologue bb jumps (which obviously
doesn't have %rbp saved or restored). When not emitting leave, but instead
doing ix86_emit_restore_reg_using_pop, we emit the restore note too.
Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
trunk?
2011-12-06 Jakub Jelinek <ja...@redhat.com>
PR target/51354
* config/i386/i386.c (ix86_emit_leave): Call ix86_add_cfa_restore_note
unconditionally.
* gcc.c-torture/compile/pr51354.c: New test.
--- gcc/config/i386/i386.c.jj 2011-12-02 01:52:26.000000000 +0100
+++ gcc/config/i386/i386.c 2011-12-06 10:38:40.370565431 +0100
@@ -10545,9 +10545,9 @@ ix86_emit_leave (void)
add_reg_note (insn, REG_CFA_DEF_CFA,
plus_constant (stack_pointer_rtx, m->fs.sp_offset));
RTX_FRAME_RELATED_P (insn) = 1;
- ix86_add_cfa_restore_note (insn, hard_frame_pointer_rtx,
- m->fs.fp_offset);
}
+ ix86_add_cfa_restore_note (insn, hard_frame_pointer_rtx,
+ m->fs.fp_offset);
}
/* Emit code to restore saved registers using MOV insns.
--- gcc/testsuite/gcc.c-torture/compile/pr51354.c.jj 2011-12-06
10:41:52.069421458 +0100
+++ gcc/testsuite/gcc.c-torture/compile/pr51354.c 2011-12-06
10:41:01.000000000 +0100
@@ -0,0 +1,17 @@
+/* PR target/51354 */
+
+extern void abort (void);
+
+typedef int __attribute__ ((aligned (32))) ai;
+
+void foo (int *x, ai * y);
+
+int
+bar (int x)
+{
+ if (x == 12346)
+ return 24;
+ ai i;
+ foo (__builtin_alloca (x), &i);
+ return 128;
+}
Jakub
