[PATCH] D99565: [X86] Support replacing aligned vector moves with unaligned moves when avx is enabled.

[LiuChen via Phabricator via llvm-commits]

LiuChen3 added inline comments.


================
Comment at: clang/include/clang/Driver/Options.td:1642
+    Group<f_Group>, Flags<[CoreOption, HelpHidden]>,
+    HelpText<"Enable transforming aligned vector move instruction to "
+             "unaligned vector move.">;
----------------
craig.topper wrote:
> As far the user is concerned this isn’t a transform. From their perspective it’s always use unaligned move instructions.
How about "Always emit unaligned move instructions." ? Do you have any suggestion here?


================
Comment at: llvm/lib/Target/X86/X86MCInstLower.cpp:2708
+  // enabled.
+  if (EnableX86UnalignedVecMove) {
+    for (const auto Pair : AlignedMovToUnalignedMovTable) {
----------------
Forgot to check the SSE level. I will add in next patch.


================
Comment at: llvm/test/CodeGen/X86/avx512vl-unaligned-load-store.ll:6
+define <8 x i32> @test_256_1(i8 * %addr) {
+; CHECK-LABEL: test_256_1:
+; CHECK:       # %bb.0:
----------------
craig.topper wrote:
> CHECK isn’t a valid prefix for this file
Thanks for reminding. I forgot to delete the these. I will remove these in next patch.


================
Comment at: llvm/test/CodeGen/X86/avx512vl-unaligned-load-store.ll:21
+  %vaddr = bitcast i8* %addr to <8 x i32>*
+  %res = load <8 x i32>, <8 x i32>* %vaddr, align 1
+  ret <8 x i32>%res
----------------
craig.topper wrote:
> What are the tests with align 1 intended to show?
This is to distinguish the unaligned-mov converted from aligned-move and the original unaligned-move. See the difference between line12 and line28.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D99565/new/

https://reviews.llvm.org/D99565