[llvm] [HashRecognize] Recognize trunc-to-i1 little-endian bit check (PR #213883)

Vito Kortbeek via llvm-commits llvm-commits at lists.llvm.org
Thu Aug 13 05:41:00 PDT 2026


================
@@ -172,16 +173,25 @@ isSignificantBitCheckWellFormed(const RecurrenceInfo &ConditionalRecurrence,
   const Value *L;
   const APInt *R;
   Instruction *TV, *FV;
-  if (!match(SI, m_Select(m_ICmp(Pred, m_Value(L), m_APInt(R)),
-                          m_Instruction(TV), m_Instruction(FV))))
-    return false;
 
   // Match predicate with or without a SimpleRecurrence (the corresponding data
   // is LHSAux).
   auto MatchPred = m_CombineOr(
       m_Specific(ConditionalRecurrence.Phi),
       m_c_Xor(m_ZExtOrTruncOrSelf(m_Specific(ConditionalRecurrence.Phi)),
               m_ZExtOrTruncOrSelf(m_Specific(SimpleRecurrence.Phi))));
+
+  BinaryOperator *BitShift = ConditionalRecurrence.BO;
+  auto MatchBitShiftXorGenPoly = m_c_Xor(
+      m_Specific(BitShift), m_SpecificInt(*ConditionalRecurrence.ExtraConst));
+  if (!IsBigEndian &&
+      match(SI, m_Select(m_Trunc(MatchPred), MatchBitShiftXorGenPoly,
+                         m_Specific(BitShift))))
+    return true;
----------------
vkortbeek-gf wrote:

I think the current version is correct. These are now the two alternative little endian forms:
```llvm
; new trunc to i1 form
%cond = trunc i8 %x to i1
%shift = lshr i16 %crc, 1
%with.poly = xor i16 %shift, -24575
%next = select i1 %cond, i16 %with.poly, i16 %shift
```

```llvm
; (corresponding) original icmp form, there are more
%bit = and i8 %x, 1
%cond = icmp ne i8 %bit, 0
%shift = lshr i16 %crc, 1
%with.poly = xor i16 %shift, -24575
%next = select i1 %cond, i16 %with.poly, i16 %shift
```

The `trunc` matcher validates the complete first form. The later predicate/range checks validate the alternative (original) `icmp` form and don’t apply to `trunc`.

The `trunc` match is a significantly simpler because there are fewer produced `trunc` variations.

https://github.com/llvm/llvm-project/pull/213883


More information about the llvm-commits mailing list