[PATCH] D64966: [AMDGPU/GlobalISel] Add llvm.amdgcn.fdiv.fast legalization.

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added a comment.

Can you also remove the case for it in AMDGPURegisterBankInfo? It should never make it to RegBankSelected



================
Comment at: llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp:1288-1293
+  const APFloat C0Val(BitsToFloat(0x6f800000));
+  const APFloat C1Val(BitsToFloat(0x2f800000));
+
+  auto C0 = B.buildFConstant(S32, C0Val);
+  auto C1 = B.buildFConstant(S32, C1Val);
+  auto C2 = B.buildFConstant(S32, 1.0);
----------------
We don't really need to use F_CONSTANT. You can just use G_CONSTANT and the constant values directly without going through APFloat


================
Comment at: llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp:1298
+
+  auto Mul0 = B.buildFMul(S32, RHS, Sel.getReg(0));
+
----------------
Should propagate the fast math flags


================
Comment at: llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp:1305-1308
+  B.buildInstr(TargetOpcode::G_FMUL)
+    .addDef(Res)
+    .addUse(Sel.getReg(0))
+    .addUse(Mul1.getReg(0));
----------------
buildFMul


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D64966/new/

https://reviews.llvm.org/D64966