[SPEC2017] Use -fno-strict-aliasing when compiling 502.gcc_r, 505.mcf_r

[labrinea]

Summary

As suggested on SPEC's documentation page (Known Portability Issues) [1][2] the spec_qsort.c routine does not strictly obey the ANSI aliasing rules. In the absence of -fno-strict-aliasing the test suite fails when Function Specialization is enabled with both LTO and PGO (see https://reviews.llvm.org/D140210).

[1] https://www.spec.org/cpu2017/Docs/benchmarks/502.gcc_r.html
[2] https://www.spec.org/cpu2017/Docs/benchmarks/505.mcf_r.html

[labrinea]

Hmm, I am wondering whether I should add the flag to 505.mcf_r too, despite it's not failing.

[xbolva00]

In D141474#4042886, @labrinea wrote:

Hmm, I am wondering whether I should add the flag to 505.mcf_r too, despite it's not failing.

Should be added later, when needed.

[SjoerdMeijer]

Why not adding it now for MCF as well? It's clear that this is a potential problem, so adding it now may save someone some time looking into this when it does start failing.

[fhahn]

Is this change perf-neutral?

[SjoerdMeijer]

It's not entirely perf neutral last time I actually played with this a couple of weeks ago. I can't remember exactly, but about 0.5% difference or so for MCF.
So nothing shocking I think, and I think we have a good reason to add it.

[xbolva00]

In D141474#4043044, @SjoerdMeijer wrote:

Why not adding it now for MCF as well? It's clear that this is a potential problem, so adding it now may save someone some time looking into this when it does start failing.

The questions is why there is no such failure for mcf currently... no func specialization for this benchmark in spec_qsort? But from the description it should happen

Measured +9.121% score increase for 505.mcf_r from SPEC Int 2017
(Tested on Neoverse N1 with -O3 + LTO)

So in this case yes, you should add this flag for mcf too.

[labrinea]

In D141474#4043064, @fhahn wrote:

Is this change perf-neutral?

I am not seeing much impact on performance. The score differences seem noisy (average of first three runs). On AArch64 -O3 + LTO (no PGO) I am seeing:

-fstrict-aliasing --> -fno-strict-aliasing

testname	FuncSpec=Off	FuncSpec=On
505.mcf_r	+ 0.101 %	+ 0.092 %
502.gcc_r	+ 0.085 %	- 0.189 %

That said I will be adding the flag to 505.mcf_r too.

[labrinea]

Added the flag to 505.mcf_r as well.

[SjoerdMeijer]

Thanks for checking the perf numbers. It matches with my experiences. I think I saw slightly bigger numbers, but yeah, like I said, nothing shocking.

I think adding this to MCF makes sense too, so LGTM.

External/SPEC/CINT2017rate/502.gcc_r/CMakeLists.txt

 # https://www.spec.org/cpu2017/Docs/benchmarks/502.gcc_r.html
 include(../../SpecCPU2017.cmake)
 if (NOT TEST_SUITE_SPEC2017_ROOT)
   return ()
 endif ()
 
 speccpu2017_benchmark(RATE)
 
 add_definitions(-DSPEC_502 -DIN_GCC -DHAVE_CONFIG_H)
 speccpu2017_add_include_dirs(. include spec_qsort)
-add_compile_options(-fgnu89-inline)
+add_compile_options(-fgnu89-inline -fno-strict-aliasing)
 
 ## test ########################################################################
 
 speccpu2017_run_test(
   "${RUN_test_DIR_REL}/t1.c" -O3 -finline-limit=50000 -o t1.opts-O3_-finline-limit_50000.s
   RUN_TYPE test
 )