[PATCH] Fix PR77916

[Christophe Lyon]

On 18 October 2016 at 05:18, Markus Trippelsdorf <markus@trippelsdorf.de> wrote:
> On 2016.10.18 at 05:13 +0200, Markus Trippelsdorf wrote:
>> On 2016.10.17 at 17:23 -0500, Bill Schmidt wrote:
>> > Hi,
>> >
>> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77916 identifies a situation
>> > where SLSR will ICE when exposed to a cast from integer to pointer.  This
>> > is because we try to convert a PLUS_EXPR with an addend of -1 * S into a
>> > MINUS_EXPR with a subtrahend of S, but the base operand is unexpectedly
>> > of pointer type.  This patch recognizes when pointer arithmetic is taking
>> > place and ensures that we use a POINTER_PLUS_EXPR at all such times.  In
>> > the case of the PR, this occurs in the logic where the stride S is a known
>> > constant value, but the same problem could occur when it is an SSA_NAME
>> > without known value.  Both possibilities are handled here.
>> >
>> > Fixing the code to ensure that the unknown stride case always uses an
>> > initializer for a negative increment allows us to remove the stopgap fix
>> > added for PR77937 as well.
>> >
>> > Bootstrapped and tested on powerpc64le-unknown-linux-gnu with no
>> > regressions, committed.
>>
>> Perhaps you should consider building ffmpeg with -O3 -march=amdfam10 on
>> X86_64 before committing these patches, because you broke it for the
>> third time in the last couple of days.
>>
>> markus@x4 ffmpeg % cat h264dsp.i
>> extern int fn2(int);
>> extern int fn3(int);
>> int a, b, c;
>> void fn1(long p1) {
>>   char *d;
>>   for (;; d += p1) {
>>     d[0] = fn2(1 >> c >> 1);
>>     fn2(c >> a);
>>     d[1] = fn3(d[1]) >> 1;
>>     d[6] = fn3(d[6] * b + 1) >> 1;
>>     d[7] = fn3(d[7] * b + 1) >> 1;
>>     d[8] = fn3(d[8] * b + 1) >> 1;
>>   }
>> }
>>
>> markus@x4 ffmpeg % gcc -O3 -march=amdfam10 -c h264dsp.i
>> h264dsp.i: In function ‘fn1’:
>> h264dsp.i:4:6: internal compiler error: in replace_one_candidate, at gimple-ssa-strength-reduction.c:3375
>>  void fn1(long p1) {
>>       ^~~
>> 0x12773a9 replace_one_candidate
>>         ../../gcc/gcc/gimple-ssa-strength-reduction.c:3375
>> 0x127af77 replace_profitable_candidates
>>         ../../gcc/gcc/gimple-ssa-strength-reduction.c:3486
>> 0x127aeeb replace_profitable_candidates
>>         ../../gcc/gcc/gimple-ssa-strength-reduction.c:3495
>> 0x127f3ee analyze_candidates_and_replace
>>         ../../gcc/gcc/gimple-ssa-strength-reduction.c:3574
>> 0x127f3ee execute
>>         ../../gcc/gcc/gimple-ssa-strength-reduction.c:3648
>
> Just figured out that this testcase is identical to:
> gcc/testsuite/gcc.dg/torture/pr77937-2.c
>
> So please run the testsuite on X86_64 in the future.
>


I'm not sure whether Markus means that pr77937-2 fails since this commit?

I'm seeing ICEs on pr77937-2 on some arm targets:
http://people.linaro.org/~christophe.lyon/cross-validation/gcc/trunk/241285/report-build-info.html

but I'm not 100% sure it was caused by this patch (the regression
happened between
r241273 and r241285), I haven't looked in details yet.

Christophe

> --
> Markus