[PATCH] PR middle-end/104885: Fix ICE with large stack frame on powerpc64.

[PATCH] PR middle-end/104885: Fix ICE with large stack frame on powerpc64.

[Roger Sayle]

[-- Attachment #1: Type: text/plain, Size: 1283 bytes --]


My recent testcase for PR c++/84964.C stress tests the middle-end by
attempting to pass a UINT_MAX sized structure on the stack.  Although
my fix to PR84964 avoids the ICE after sorry() on x86_64 and similar
targets, a related issue still exists on powerpc64 (and similar
ACCUMULATE_OUTGOING_ARGS/ARGS_GROW_DOWNWARD targets) which don't
issue a "sorry, unimplemented" message, but instead ICE elsewhere.

After attempting several alternate fixes, the simplest solution is
to just defensively check in mark_stack_region_used that the upper
bound of the region lies within the allocated stack_usage_map
array, which is of size highest_outgoing_arg_in_use.  When this isn't
the case, the code now follows the same path as for variable sized
regions, and uses stack_usage_watermark rather than a map.

This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
and make -k check to confirm there are no surprises, and with the cc1plus
of a cross-compiler to powerpc64-linux-gnu to confirm the new test
case no longer ICEs.  Ok for mainline?


2022-03-12  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
	PR middle-end/104885
	* calls.cc (mark_stack_region_used): Check that the region
	is within the allocated size of stack_usage_map.


Thanks in advance,
Roger
--


[-- Attachment #2: patchpp.txt --]
[-- Type: text/plain, Size: 583 bytes --]

diff --git a/gcc/calls.cc b/gcc/calls.cc
index 50fa7b8..1ca96e7 100644
--- a/gcc/calls.cc
+++ b/gcc/calls.cc
@@ -201,7 +201,8 @@ mark_stack_region_used (poly_uint64 lower_bound, poly_uint64 upper_bound)
 {
   unsigned HOST_WIDE_INT const_lower, const_upper;
   const_lower = constant_lower_bound (lower_bound);
-  if (upper_bound.is_constant (&const_upper))
+  if (upper_bound.is_constant (&const_upper)
+      && const_upper <= highest_outgoing_arg_in_use)
     for (unsigned HOST_WIDE_INT i = const_lower; i < const_upper; ++i)
       stack_usage_map[i] = 1;
   else

Re: [PATCH] PR middle-end/104885: Fix ICE with large stack frame on powerpc64.

[Jeff Law]

On 3/12/2022 12:40 PM, Roger Sayle wrote:
> My recent testcase for PR c++/84964.C stress tests the middle-end by
> attempting to pass a UINT_MAX sized structure on the stack.  Although
> my fix to PR84964 avoids the ICE after sorry() on x86_64 and similar
> targets, a related issue still exists on powerpc64 (and similar
> ACCUMULATE_OUTGOING_ARGS/ARGS_GROW_DOWNWARD targets) which don't
> issue a "sorry, unimplemented" message, but instead ICE elsewhere.
>
> After attempting several alternate fixes, the simplest solution is
> to just defensively check in mark_stack_region_used that the upper
> bound of the region lies within the allocated stack_usage_map
> array, which is of size highest_outgoing_arg_in_use.  When this isn't
> the case, the code now follows the same path as for variable sized
> regions, and uses stack_usage_watermark rather than a map.
>
> This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> and make -k check to confirm there are no surprises, and with the cc1plus
> of a cross-compiler to powerpc64-linux-gnu to confirm the new test
> case no longer ICEs.  Ok for mainline?
>
>
> 2022-03-12  Roger Sayle  <roger@nextmovesoftware.com>
>
> gcc/ChangeLog
> 	PR middle-end/104885
> 	* calls.cc (mark_stack_region_used): Check that the region
> 	is within the allocated size of stack_usage_map.
OK
jeff

Re: [PATCH] PR middle-end/104885: Fix ICE with large stack frame on powerpc64.

[Segher Boessenkool]

On Sat, Mar 12, 2022 at 07:40:08PM -0000, Roger Sayle wrote:
> My recent testcase for PR c++/84964.C stress tests the middle-end by
> attempting to pass a UINT_MAX sized structure on the stack.  Although
> my fix to PR84964 avoids the ICE after sorry() on x86_64 and similar
> targets, a related issue still exists on powerpc64 (and similar
> ACCUMULATE_OUTGOING_ARGS/ARGS_GROW_DOWNWARD targets) which don't
> issue a "sorry, unimplemented" message, but instead ICE elsewhere.
> 
> After attempting several alternate fixes, the simplest solution is
> to just defensively check in mark_stack_region_used that the upper
> bound of the region lies within the allocated stack_usage_map
> array, which is of size highest_outgoing_arg_in_use.  When this isn't
> the case, the code now follows the same path as for variable sized
> regions, and uses stack_usage_watermark rather than a map.

Thanks for the fix!


Segher