[X86] Distribute Certain Bitwise Operations over SELECT

InstCombine canonicalizes `(select P (and X (- X)) X)` to
`(and (select P (- X) umax) X)`. This is counterproductive for the X86
backend when BMI is available because we can encode `(and X (- X))`
using the `BLSI` instruction. A similar situation arises if we have
`(select P (and X (sub X 1)) X)` (prevents use of `BLSR` instruction) or
`(select P (xor X (sub X 1)) X)` (prevents use of `BLSMSK` instruction).

Trigger the inverse transformation in the X86 backend if BMI is
available and we can use the mentioned BMI instructions. This is done by
adjusting the `shouldFoldSelectWithIdentityConstant()` implementation
for the X86 backend. In this way, we get `(select P (and X (- X)) X)`
again, which enables the use of `BLSI` (similar for the other cases
described above).

Alive proofs: https://alive2.llvm.org/ce/z/MT_pKi

Fixes #131587, fixes #133848.

Author: mskamp

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -35552,8 +35552,29 @@ bool X86TargetLowering::isNarrowingProfitable(SDNode *N, EVT SrcVT,
 bool X86TargetLowering::shouldFoldSelectWithIdentityConstant(
     unsigned BinOpcode, EVT VT, unsigned SelectOpcode, SDValue X,
     SDValue Y) const {
-  if (SelectOpcode != ISD::VSELECT)
+  if (SelectOpcode == ISD::SELECT) {
+    if (VT.isVector())
+      return false;
+    if (!Subtarget.hasBMI() || (VT != MVT::i32 && VT != MVT::i64))
+      return false;
+    using namespace llvm::SDPatternMatch;
+    // BLSI
+    if (BinOpcode == ISD::AND && (sd_match(Y, m_Neg(m_Specific(X))) ||
+                                  sd_match(X, m_Neg(m_Specific(Y)))))
+      return true;
+    // BLSR
+    if (BinOpcode == ISD::AND &&
+        (sd_match(Y, m_Add(m_Specific(X), m_AllOnes())) ||
+         sd_match(X, m_Add(m_Specific(Y), m_AllOnes()))))
+      return true;
+    // BLSMSK
+    if (BinOpcode == ISD::XOR &&
+        (sd_match(Y, m_Add(m_Specific(X), m_AllOnes())) ||
+         sd_match(X, m_Add(m_Specific(Y), m_AllOnes()))))
+      return true;
+
     return false;
+  }
   // TODO: This is too general. There are cases where pre-AVX512 codegen would
   //       benefit. The transform may also be profitable for scalar code.
   if (!Subtarget.hasAVX512())