[llvm] [InstCombine] Don't consider aligned_alloc removable if icmp uses result (PR #69474)

[Florian Hahn via llvm-commits]

================
@@ -2430,6 +2430,26 @@ static bool isAllocSiteRemovable(Instruction *AI,
         unsigned OtherIndex = (ICI->getOperand(0) == PI) ? 1 : 0;
         if (!isNeverEqualToUnescapedAlloc(ICI->getOperand(OtherIndex), TLI, AI))
           return false;
+
+        // Do not fold compares to aligned_alloc calls, as they may have to
+        // return null in case the required alignment cannot be satisfied,
+        // unless we can prove that both alignment and size are valid.
+        auto AlignmentAndSizeKnownValid = [](CallBase *CB) {
+          // Check if alignment and size of a call to aligned_alloc is valid,
+          // that is alignment is a power-of-2 and the size is a multiple of the
+          // alignment.
+          const APInt *Alignment;
+          const APInt *Size;
+          return match(CB->getArgOperand(0), m_APInt(Alignment)) &&
+                 match(CB->getArgOperand(1), m_APInt(Size)) &&
+                 Alignment->isPowerOf2() && Size->urem(*Alignment).isZero();
+        };
+        auto *CB = dyn_cast<CallBase>(AI);
----------------
fhahn wrote:

No, it's the same. Personally I prefer passing args explicitly in cases such as this one, as it makes things at the call site more explicit.

https://github.com/llvm/llvm-project/pull/69474