[PATCH] D154791: [InstCombine] Transform bitwise (A >> C - 1, zext(icmp)) -> zext (bitwise(A < 0, icmp)) fold.

[Nikita Popov via Phabricator via llvm-commits]

nikic added a comment.

This is still missing multi-use tests. We'll need some m_OneUse guards to prevent unprofitable transforms.



================
Comment at: llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp:1716
 
-  // ( A << (X - 1) ) | ((A > 0) zext to iX)
-  // <=> A < 0 | A > 0
-  // <=> (A != 0) zext to iX
-  Value *A;
-  ICmpInst::Predicate Pred;
+  // fold and/or(A << X - 1, zext(icmp))     (X is the scalar bits of the type of A)
+  //   -> and/or(zext(A < 0), zext(icmp))
----------------
X -> BW


================
Comment at: llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp:1732-1733
 
-  auto MatchOrZExtICmp = [&](Value *Op0, Value *Op1) -> bool {
-    return match(Op0, m_LShr(m_Value(A), m_SpecificInt(Op0->getType()->getScalarSizeInBits() - 1))) &&
-           match(Op1, m_ZExt(m_ICmp(Pred, m_Specific(A), m_Zero())));
+    return ZExtInst::Create(
+        ZExtInst::ZExt,
+        Builder.CreateBinOp(
----------------



================
Comment at: llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp:1742
+  if(auto *Ret = FoldBitwiseICmpZeroWithICmp(Op0, Op1)) return Ret;
+  if(auto *Ret = FoldBitwiseICmpZeroWithICmp(Op1, Op0)) return Ret;
 
----------------
Needs clang-format.


================
Comment at: llvm/test/Transforms/InstCombine/and-or-icmps.ll:2857
   ret i32 %D
 }
 
----------------
The main multi-use test I'm looking for is one where the resulting binop does not get folded. That's the case where your current code will increase instructions, I believe.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D154791/new/

https://reviews.llvm.org/D154791