Allow Builtin.bitcast to do VecNxInt1 -> IntN (swiftlang#36252)

This is necessary to support the idiom for movmsk operations on x86.

Author: stephentyrone

lib/AST/Builtins.cpp

@@ -577,6 +577,15 @@ static ValueDecl *getCastOperation(ASTContext &Context, Identifier Id,
       if (auto *BIT = CheckOutput->getAs<BuiltinIntegerType>())
         if (BIT->isFixedWidth() && BIT->getFixedWidth() == BFT->getBitWidth())
           break;
+      
+    // Support VecNxInt1 -> IntN bitcast for SIMD comparison results.
+    if (auto *Vec = CheckInput->getAs<BuiltinVectorType>())
+      if (auto *BIT = CheckOutput->getAs<BuiltinIntegerType>())
+        if (auto *Element = Vec->getElementType()->getAs<BuiltinIntegerType>())
+          if (Element->getFixedWidth() == 1 &&
+              BIT->isFixedWidth() &&
+              BIT->getFixedWidth() == Vec->getNumElements())
+            break;
 
     // FIXME: Implement bitcast typechecking.
     llvm_unreachable("Bitcast not supported yet!");

test/IRGen/builtins.swift

@@ -158,6 +158,14 @@ func cast_test(_ ptr: inout Builtin.RawPointer, i8: inout Builtin.Int8,
   d = Builtin.bitcast_Int64_FPIEEE64(i64)   // CHECK: bitcast
 }
 
+func vector_bitcast_test(_ src: Builtin.Vec16xInt8) -> Builtin.Int16 {
+  // CHECK: vector_bitcast_test
+  // This is the idiom for pmovmskb on x86 targets:
+  let zero: Builtin.Vec16xInt8 = Builtin.zeroInitializer()
+  let mask = Builtin.cmp_slt_Vec16xInt8(src, zero)
+  return Builtin.bitcast_Vec16xInt1_Int16(mask) // CHECK: bitcast
+}
+
 func intrinsic_test(_ i32: inout Builtin.Int32, i16: inout Builtin.Int16,
                     _ v8i16: Builtin.Vec8xInt16) {
   // CHECK: intrinsic_test