[llvm] [ValueTracking] Extend LHS/RHS with matching operand to work without constants. (PR #85557)

[Yingwei Zheng via llvm-commits]

================
@@ -8729,11 +8729,30 @@ static std::optional<bool> isImpliedCondICmps(const ICmpInst *LHS,
     }
   }
 
-  // Can we infer anything when the 0-operands match and the 1-operands are
-  // constants (not necessarily matching)?
-  const APInt *LC, *RC;
-  if (L0 == R0 && match(L1, m_APInt(LC)) && match(R1, m_APInt(RC)))
-    return isImpliedCondCommonOperandWithConstants(LPred, *LC, RPred, *RC);
+  // See if we can infer anything if operand-0 matches and we have at least one
+  // constant.
+  const APInt *Unused;
+  if (L0 == R0 && L0->getType()->isIntOrIntVectorTy() &&
+      (match(L1, m_APInt(Unused)) || match(R1, m_APInt(Unused)))) {
+    // Potential TODO: We could also further use the constant range of L0/R0 to
+    // further constraint the constant ranges. At the moment this leads to
+    // several regressions related to not transforming `multi_use(A + C0) eq/ne
+    // C1` (see discussion: D58633).
+    ConstantRange LCR = computeConstantRange(
+        L1, ICmpInst::isSigned(LPred), /* UseInstrInfo=*/true, /*AC=*/nullptr,
+        /*CxtI=*/nullptr, /*DT=*/nullptr, MaxAnalysisRecursionDepth - 1);
----------------
dtcxzyw wrote:

Should we pass `Depth` to `isImpliedCondICmps`?


https://github.com/llvm/llvm-project/pull/85557