[PATCH] D37294: [X86][Skylake] Adding the scheduling information for the SkylakeClient target

[Gadi Haber via Phabricator via llvm-commits]

gadi.haber added inline comments.


================
Comment at: test/CodeGen/X86/avx2-intrinsics-x86.ll:1625
+; AVX512VL-NEXT:    vmovaps %ymm2, %ymm3 ## EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xda]
+; AVX512VL-NEXT:    vgatherdps %ymm3, (%ecx,%ymm1,4), %ymm0 ## encoding: [0xc4,0xe2,0x65,0x92,0x04,0x89]
 ; AVX512VL-NEXT:    vmovups %ymm2, (%eax) ## EVEX TO VEX Compression encoding: [0xc5,0xfc,0x11,0x10]
----------------
gadi.haber wrote:
> RKSimon wrote:
> > gadi.haber wrote:
> > > RKSimon wrote:
> > > > gadi.haber wrote:
> > > > > RKSimon wrote:
> > > > > > I feels tests like this shouldn't change with schedule model tweaks - it's supposed to be a codegen test not a target specific test.
> > > > > > 
> > > > > > Would it be possible to reduce the use of -mcpu from tests that aren't actually testing for scheduling? 
> > > > > Sure. I can modify the tests.  IS there a list of such tests that we can assume to tests only code gen?
> > > > Most of the non *-schedule.ll ones here for instance, the vector-shift-*.ll are another set I can remember kept changing.
> > > OK. In that case I suggest to abandon this review.  I will modify the tests and commit them as NFC and the re-upload it for review after the commit.
> > You don't need to do that - just fix the tests separately and then rebase this patch
> ok.
These are the attributes I though of using instead of -mcpu=skx:
-mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq

This is the attribute to use instead of -mcpu=knl:
-mattr=+avx512f

There are roughly 100 tests affected by this.

Any thoughts?



Repository:
  rL LLVM

https://reviews.llvm.org/D37294