Hello,
recently we got a bootstrap failure in libfortran's transfer.c:sset
function. This is caused by optimizing alloca into prologue. This
can cause for x64 with SEH to too large prologue-frame-size for SEH
information.
This patch simply assumes that iff the function is using an alloca
call, then we should set frame's hard_frame_pointer_offset to its
current stack_pointer_offset (minus 128 delta for smaller common
stack-address ranges).
ChangeLog
2011-08-04 Kai Tietz <kti...@redhat.com>
* config/i386/i386.c (ix86_compute_frame_layout): Adjust
hard_frame_pointer_offset
if function uses alloca.
Regression test in progress. Ok for apply after successful test?
Regards,
Kai
Index: i386.c
===================================================================
--- i386.c (revision 177412)
+++ i386.c (working copy)
@@ -8860,7 +8860,7 @@
/* If we can leave the frame pointer where it is, do so. */
diff = frame->stack_pointer_offset - frame->hard_frame_pointer_offset;
- if (diff > 240 || (diff & 15) != 0)
+ if (diff > 240 || (diff & 15) != 0 || cfun->calls_alloca != 0)
{
/* Ideally we'd determine what portion of the local stack frame
(within the constraint of the lowest 240) is most heavily used.
