[PATCH] D153034: [PowerPC][Future] Enable __builtin_mma_xxm[t|f]acc

Kamau Bridgeman via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Fri Jun 30 14:29:51 PDT 2023

kamaub added inline comments.

Comment at: llvm/test/CodeGen/PowerPC/mmaplus-intrinsics.ll:152
+; CHECK-O0-NEXT:    lxv vs0, 32(r3)
+; CHECK-O0-NEXT:    # implicit-def: $vsrp17
+; CHECK-O0-NEXT:    xxlor v3, vs0, vs0
lei wrote:
> I realized these -O0 run line was not added as part of this patch, but I am wondering why we need to be explicitly running `-O0` for these intrinsic tests.  Are they generated differently for `-O0` vs opt?
I think the reason for adding O0 test cases was to make sure that FastISEL worked in these tested situations. For this patch specifically the `xxm[t|f]acc` intrinsics are not removed differently at O0, but having the check to make sure the the quad vector type is lowered to the expected dmr moves as a result of the intrinsic removal makes sense to me.

Even then, the amount of new O0 checks just for this change could be considered overkill. 

Here I am adding a few new test cases but also moving some from some from `mma-instrinsics.ll` so i could move those new test to their own file where I do not need to test fastISEL, but I did not see a strong reasoning to do so at the time.

  rG LLVM Github Monorepo



More information about the llvm-commits mailing list