[PATCH] D39802: Sched model improving on btver2: JFPU01 resource, vtestp* for xmm.

[Andrea Di Biagio via Phabricator via llvm-commits]

andreadb accepted this revision.
andreadb added a comment.
This revision is now accepted and ready to land.

In https://reviews.llvm.org/D39802#920299, @avt77 wrote:

> I made updates required by andreadb except the question about JFPU01: if we need perf experiments please help me with perf test(s). If we agree to use the current implementation then I could commit the patch if you get me LGTM.


I don't think it is worthy to run experiments just for the blendv case.
I just wanted to point out that there may be an alternative solution for that particular case. Having a resource cycle of 4 for the blendv is a bit odd (at least to me)..
But, as I wrote said, I really don't have a strong opinion on what numbers should go there (neither solution is ideal).
If you fix the remaining NumMicroOps issue I pointed out, then the patch looks good to me.



================
Comment at: lib/Target/X86/X86ScheduleBtVer2.td:656-659
 def WriteVTESTY: SchedWriteRes<[JFPU01, JFPU0]> {
   let Latency = 4;
-  let ResourceCycles = [4, 2];
+  let ResourceCycles = [2, 2];
 }
----------------
let NumMicroOps = 3;




================
Comment at: lib/Target/X86/X86ScheduleBtVer2.td:663-666
 def WriteVTESTYLd: SchedWriteRes<[JLAGU, JFPU01, JFPU0]> {
   let Latency = 9;
-  let ResourceCycles = [1, 4, 2];
+  let ResourceCycles = [1, 2, 2];
 }
----------------
Same.


https://reviews.llvm.org/D39802