[llvm] [ConstantFolding] Consolidate poison propagation for intrinsics (PR #146878)

[Nikita Popov via llvm-commits]

================
@@ -7879,6 +7879,47 @@ bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
   llvm_unreachable("Instruction not contained in its own parent basic block.");
 }
 
+bool llvm::intrinsicPropagatesPoison(Intrinsic::ID IID) {
+  switch (IID) {
+  // TODO: Add more intrinsics.
+  case Intrinsic::sadd_with_overflow:
+  case Intrinsic::ssub_with_overflow:
+  case Intrinsic::smul_with_overflow:
+  case Intrinsic::uadd_with_overflow:
+  case Intrinsic::usub_with_overflow:
+  case Intrinsic::umul_with_overflow:
+    // If an input is a vector containing a poison element, the
+    // two output vectors (calculated results, overflow bits)'
+    // corresponding lanes are poison.
+    return true;
+  case Intrinsic::ctpop:
+  case Intrinsic::ctlz:
+  case Intrinsic::cttz:
+  case Intrinsic::abs:
+  case Intrinsic::smax:
+  case Intrinsic::smin:
+  case Intrinsic::umax:
+  case Intrinsic::umin:
+  case Intrinsic::scmp:
+  case Intrinsic::ucmp:
+  case Intrinsic::bitreverse:
+  case Intrinsic::bswap:
+  case Intrinsic::sadd_sat:
+  case Intrinsic::ssub_sat:
+  case Intrinsic::sshl_sat:
+  case Intrinsic::uadd_sat:
+  case Intrinsic::usub_sat:
+  case Intrinsic::ushl_sat:
+  case Intrinsic::smul_fix:
+  case Intrinsic::smul_fix_sat:
+  case Intrinsic::canonicalize:
+  case Intrinsic::sqrt:
----------------
nikic wrote:

Consolidating the frexp case is more complicated due to the struct return. I think it needs some refactorings to ConstantFoldStructCall first.

https://github.com/llvm/llvm-project/pull/146878