This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH, AArch64] Enable shrink wrap
- From: Jiong Wang <wong dot kwongyuan dot tools at gmail dot com>
- To: Sebastian Pop <sebpop at gmail dot com>
- Cc: Marcus Shawcroft <marcus dot shawcroft at gmail dot com>, Jiong Wang <jiong dot wang at arm dot com>, "gcc-patches at gcc dot gnu dot org" <gcc-patches at gcc dot gnu dot org>
- Date: Tue, 23 Sep 2014 22:51:49 +0100
- Subject: Re: [PATCH, AArch64] Enable shrink wrap
- Authentication-results: sourceware.org; auth=none
- References: <540873E1 dot 1090009 at arm dot com> <CAFqB+PwLMzHTE+Pib61jcYue1NoZk5kRnvqUpvmK3Ajcn3jUMA at mail dot gmail dot com> <20140923205909 dot GA13176 at f1 dot c dot bardezibar dot internal>
2014-09-23 21:59 GMT+01:00 Sebastian Pop <sebpop@gmail.com>:
> Marcus Shawcroft wrote:
>> On 4 September 2014 15:14, Jiong Wang <jiong.wang@arm.com> wrote:
>> > this patch enabled stack shrink-wrap support on AArch64.
>> >
>> > no regression on aarch64-none-elf bare-metal.
>> > aarch64 bootstrap OK.
>> >
>> > ok to install?
>> >
>> > 2014-09-04 Renlin Li<renlin.li@arm.com>
>> >
>> > gcc/
>> > * config/aarch64/aarch64.md (return): New expand.
>> > (simple_return): Likewise.
>> > * config/aarch64/aarch64.c (aarch64_use_return_insn_p): New function.
>> > * config/aarch64/aarch64-protos.h (aarch64_use_return_insn_p): New
>> > declaration.
>> >
>> > gcc/testsuite
>> > * gcc.dg/ira-shrinkwrap-prep-1.c: Enable aarch64.
>> > * gcc.dg/ira-shrinkwrap-prep-2.c: Likewise.
>> > * gcc.dg/pr10474.c: Likewise.
>>
>> OK, committed as 215508
>> /Marcus
>
> This patch causes a regression when compiling this testcase with -O2 and above:
>
> void mm(double** A, double** B, double** C, unsigned N, unsigned S)
> {
> unsigned i, j, k;
>
> for (i = 0; i < N; i += S) {
> for (j = 0; j < N; j += S) {
> for (k = 0; k < N; k += S) {
>
> unsigned ib = (N) < ((i + S)) ? N : (i + S);
> unsigned jb = (N) < ((j + S)) ? N : (j + S);
> unsigned kb = (N) < ((k + S)) ? N : (k + S);
> unsigned i0, j0, k0;
>
> for (i0 = i; i0 < ib; i0++) {
> for (j0 = j; j0 < jb; j0++) {
> double* a = A[i0];
> double* b = B[j0];
> double scratch = C[i0][j0];
> for (k0 = k; k0 < kb; k0++) {
> scratch += a[k0] * b[k0];
> }
> C[i0][j0] = scratch;
> }
> }
> }
> }
> }
> asm volatile ("bar:");
> }
>
> $ aarch64-gcc -O2 mm.c -S -o - | grep 'bar:' | wc -l
> 2
> $ aarch64-gcc -O1 mm.c -S -o - | grep 'bar:' | wc -l
> 1
Interesting.
I have done a quick investigation on x86/mips/arm32/aarch64 and found
* x86/mips couldn't shrink wrap this function.
* arm32/aarch64 could shrink wrap it.
but both generate the same code layout, with redundant exit basic block.
So, looks like it's caused by some hiding issue in generic code.
>
> Thanks,
> Sebastian