[clang] [clang][clangIR]: X86 vperm2f128 builtin implementation (PR #208851)

[Lucas Ribeiro Lima via cfe-commits]

================
@@ -811,6 +811,48 @@ static mlir::Value emitX86MaskedLoad(CIRGenBuilderTy &builder,
   return builder.createMaskedLoad(loc, ty, ptr, alignment, maskVec, ops[1]);
 }
 
+static mlir::Value emitX86VPerm2f128(CIRGenBuilderTy &builder,
+                                     mlir::Location loc,
+                                     llvm::SmallVector<mlir::Value> ops) {
+  auto inputType = cast<cir::VectorType>(ops[0].getType());
+  assert(!inputType.getIsScalable() &&
+         "This is only intended for fixed-width vectors");
+
+  const uint8_t imm = CIRGenFunction::getZExtIntValueFromConstOp(ops[2]) & 0xFF;
+  mlir::Value zeroVec = builder.getZero(loc, inputType);
+
+  // Mirror hardware and OGCG behaviour returning a zero vector
+  if ((imm & 0x80) && (imm & 0x08))
+    return zeroVec;
+
+  mlir::Value lanes[2];
+  llvm::SmallVector<int64_t, 64> mask;
+  const unsigned numElts = inputType.getSize();
+
+  // We must evaluated each lane(128 bits) separetely
+  for (auto lane : llvm::seq(0, 2)) {
+    llvm::Boolean isZeroBit = imm & (1 << ((lane * 4) + 3)),
----------------
lucaslive974 wrote:

I thought was a good idea using llvm specific types than generics. Going to replace for a simple bool.

https://github.com/llvm/llvm-project/pull/208851