[PATCH] D144534: [DAGCombine] Fold SETCC_OEQ(x, Inf) into IS_FPCLASS

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp:4969
+
+    if (isOperationLegalOrCustom(ISD::IS_FPCLASS, N0.getValueType())) {
+      bool IsFabs = N0.getOpcode() == ISD::FABS;
----------------
Is the advantage that you avoid materializing the inf constant? Can you check if the constantfp is legal?


================
Comment at: llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp:4972
+      SDValue Op = IsFabs ? N0.getOperand(0) : N0;
+      if (Cond == ISD::SETOEQ && CFP->isInfinity()) {
+        FPClassTest Flag = CFP->isNegative() ? (IsFabs ? fcNone : fcNegInf)
----------------
Could also handle the unordered case 


================
Comment at: llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp:4976
+        return DAG.getNode(ISD::IS_FPCLASS, dl, VT, Op,
+                           DAG.getConstant(Flag, dl, MVT::i32, true));
+      }
----------------
getTargetConstant 


================
Comment at: llvm/test/CodeGen/AMDGPU/fp-classify.ll:616
+; SI-NEXT:    v_cvt_f32_f16_e64 v1, |s0|
+; SI-NEXT:    v_cmp_class_f32_e32 vcc, v1, v0
 ; SI-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc
----------------
The result is the same but I’d somewhat prefer to keep the original isa. The infinity constant is more likely reusable with other uses in the program 


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D144534/new/

https://reviews.llvm.org/D144534