[llvm] InstCombine: Handle canonicalize in SimplifyDemandedFPClass (PR #173189)

[Serge Pavlov via llvm-commits]

================
@@ -87,6 +87,54 @@ void KnownFPClass::propagateDenormal(const KnownFPClass &Src,
   }
 }
 
+void KnownFPClass::canonicalize(DenormalMode DenormMode) {
+  KnownFPClass KnownSrc = *this;
+  *this = KnownFPClass();
+
+  // This is essentially a stronger form of
+  // propagateCanonicalizingSrc. Other "canonicalizing" operations don't
+  // actually have an IR canonicalization guarantee.
+
+  // Canonicalize may flush denormals to zero, so we have to consider the
+  // denormal mode to preserve known-not-0 knowledge.
+  KnownFPClasses = KnownSrc.KnownFPClasses | fcZero | fcQNan;
+
+  // Stronger version of propagateNaN
+  // Canonicalize is guaranteed to quiet signaling nans.
+  if (KnownSrc.isKnownNeverNaN())
+    knownNot(fcNan);
+  else
+    knownNot(fcSNan);
+
+  // FIXME: Missing check of IEEE like types.
+
+  // If the parent function flushes denormals, the canonical output cannot be a
+  // denormal.
+  if (DenormMode == DenormalMode::getIEEE()) {
+    if (KnownSrc.isKnownNever(fcPosZero))
+      knownNot(fcPosZero);
+    if (KnownSrc.isKnownNever(fcNegZero))
+      knownNot(fcNegZero);
+    return;
+  }
+
+  if (DenormMode.inputsAreZero() || DenormMode.outputsAreZero())
----------------
spavloff wrote:

IIUC this function canonicalize FP class of the result of some operation. Output denormal mode is OK here, but how the input mode is used here? If a subnormal operand of an instruction is replaced with zero, it is visible only to the operation.

https://github.com/llvm/llvm-project/pull/173189