commit 69192e0193e60c169c7776f444362dffba31eb7d
author Manish Kausik H <46352931+Nirhar@users.noreply.github.com> Mon Jul 08 18:31:32 2024 +0530
committer GitHub <noreply@github.com> Mon Jul 08 14:01:32 2024 +0100
tree dcccfa95f356e41810f0ff7bd994ade595479574
parent cb72aecea6e7da49c7a7fbeb1bde4e401324b9ba

[LegalizeDAG] Optimize CodeGen for `ISD::CTLZ_ZERO_UNDEF` (#83039)

Previously we had the same instructions being generated for `ISD::CTLZ` and `ISD::CTLZ_ZERO_UNDEF` which did not take advantage of the fact that zero is an invalid input for `ISD::CTLZ_ZERO_UNDEF`. This commit separates codegen for the two cases to allow for the optimization for the latter case.

The details of the optimization are outlined in #82075

Fixes #82075

Co-authored-by: Manish Kausik H <hmamishkausik@gmail.com>

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

diff --git a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
index 86de1f3..3f1094e 100644
--- a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
@@ -2461,13 +2461,22 @@
       NewOpc = TargetOpcode::G_CTTZ_ZERO_UNDEF;
     }
 
+    unsigned SizeDiff = WideTy.getSizeInBits() - CurTy.getSizeInBits();
+
+    if (MI.getOpcode() == TargetOpcode::G_CTLZ_ZERO_UNDEF) {
+      // An optimization where the result is the CTLZ after the left shift by
+      // (Difference in widety and current ty), that is,
+      // MIBSrc = MIBSrc << (sizeinbits(WideTy) - sizeinbits(CurTy))
+      // Result = ctlz MIBSrc
+      MIBSrc = MIRBuilder.buildShl(WideTy, MIBSrc,
+                                   MIRBuilder.buildConstant(WideTy, SizeDiff));
+    }
+
     // Perform the operation at the larger size.
     auto MIBNewOp = MIRBuilder.buildInstr(NewOpc, {WideTy}, {MIBSrc});
     // This is already the correct result for CTPOP and CTTZs
-    if (MI.getOpcode() == TargetOpcode::G_CTLZ ||
-        MI.getOpcode() == TargetOpcode::G_CTLZ_ZERO_UNDEF) {
+    if (MI.getOpcode() == TargetOpcode::G_CTLZ) {
       // The correct result is NewOp - (Difference in widety and current ty).
-      unsigned SizeDiff = WideTy.getSizeInBits() - CurTy.getSizeInBits();
       MIBNewOp = MIRBuilder.buildSub(
           WideTy, MIBNewOp, MIRBuilder.buildConstant(WideTy, SizeDiff));
     }