[PATCH] D71785: [InstCombine][AMDGPU] Trim components of s_buffer_load

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp:1027
+
+    // Start assuming the prefix of elements is demanded, but possibly clear some other bits if
+    // there are trailing zeros (unused components at front) and update offset.
----------------
I think this goes over the 80 column limit


================
Comment at: llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp:1048
+        break;
+      case Intrinsic::amdgcn_buffer_load:
+      case Intrinsic::amdgcn_struct_buffer_load:
----------------
Probably shouldn't bother with the legacy intrinsic, I have a patch almost ready to remove them


================
Comment at: llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp:1062
+        auto Offset = II->getArgOperand(OffsetIdx);
+        unsigned SingleComponentSizeInBits = getDataLayout().getTypeSizeInBits(II->getType()->getScalarType());
+        unsigned OffsetAdd = UnusedComponentsAtFront * SingleComponentSizeInBits / 8;
----------------
Line length


================
Comment at: llvm/test/Transforms/InstCombine/AMDGPU/amdgcn-demanded-vector-elts.ll:847
+}
+
+declare float @llvm.amdgcn.s.buffer.load.f32(<4 x i32>, i32, i32) #1
----------------
Need some 8 and 16 bit element tests at least


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D71785/new/

https://reviews.llvm.org/D71785