This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [patch] Fix PR48183, NEON ICE in emit-rtl.c:immed_double_const() under -g


On Tue, Mar 29, 2011 at 1:52 PM, Richard Sandiford
<richard.sandiford@linaro.org> wrote:
> Richard Guenther <richard.guenther@gmail.com> writes:
>> On Thu, Mar 24, 2011 at 11:57 AM, Richard Sandiford
>> <richard.sandiford@linaro.org> wrote:
>>> Chung-Lin Tang <cltang@codesourcery.com> writes:
>>>> PR48183 is a case where ARM NEON instrinsics, under -O -g, produce debug
>>>> insns that tries to expand OImode (32-byte integer) zero constants, much
>>>> too large to represent as two HOST_WIDE_INTs; as the internals manual
>>>> indicates, such large constants are not supported in general, and ICEs
>>>> on the GET_MODE_BITSIZE(mode) == 2*HOST_BITS_PER_WIDE_INT assertion.
>>>>
>>>> This patch allows the cases where the large integer constant is still
>>>> representable using a single CONST_INT, such as zero(0). Bootstrapped
>>>> and tested on i686 and x86_64, cross-tested on ARM, all without
>>>> regressions. Okay for trunk?
>>>>
>>>> Thanks,
>>>> Chung-Lin
>>>>
>>>> 2011-03-20 ?Chung-Lin Tang ?<cltang@codesourcery.com>
>>>>
>>>> ? ? ? * emit-rtl.c (immed_double_const): Allow wider than
>>>> ? ? ? 2*HOST_BITS_PER_WIDE_INT mode constants when they are
>>>> ? ? ? representable as a single const_int RTX.
>>>
>>> I realise this might be seen as a good expedient fix, but it makes
>>> me a bit uneasy. ?Not a very constructive rationale, sorry.
>>>
>>> For this particular case, the problem is that vst2q_s32 and the
>>> like initialise a union directly:
>>>
>>> ?union { int32x4x2_t __i; __builtin_neon_oi __o; } __bu = { __b; };
>>>
>>> and this gets translated into a zeroing of the whole union followed
>>> by an assignment to __i:
>>>
>>> ?__bu = {};
>>> ?__bu.__i = __b;
>>
>> Btw, this looks like a missed optimization in gimplification. ?Worth
>> a bugreport (or even a fix). ?Might be a target but as well, dependent
>> on how __builtin_neon_oi looks like. ?Do you have a complete testcase
>> that reproduces the above with a cross?
>
> Yeah, build cc1 for arm-linux-gnueabi and compile the attached
> testcase (from Chung-Lin) using:
>
> ?-O2 -g -mfpu=neon -mfloat-abi=softfp

It seems that count_type_elements is confused by unions and thus clearing
is always performed.  I fail to see why count_type_elements could not
simply return 1 for all unions (non-initialized parts have undefined rather
than zero content).

categorize_ctor_elements also counts 8 elements for some reason.

The following fixes it for me:

Index: gcc/expr.c
===================================================================
--- gcc/expr.c  (revision 171606)
+++ gcc/expr.c  (working copy)
@@ -5059,7 +5059,7 @@ count_type_elements (const_tree type, bo

     case UNION_TYPE:
     case QUAL_UNION_TYPE:
-      return -1;
+      return 1;

     case COMPLEX_TYPE:
       return 2;

disclaimer: completely untested, might confuse the hell out of
output_init_constructor and friends.

Richard.

> Rchard
>
>


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]