[CIR][X86] Add support for kunpck builtins

clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp

@@ -68,6 +68,60 @@ static mlir::Value emitVectorFCmp(CIRGenBuilderTy &builder,
   return bitCast;
 }
 
+static mlir::Value convertKunpckMaskToVector(CIRGenBuilderTy &builder,
+                                             mlir::Value mask,
+                                             unsigned numElems) {
+  unsigned maskWidth = mlir::cast<cir::IntType>(mask.getType()).getWidth();
+  auto maskVecType =
+      cir::VectorType::get(builder.getContext(),
+                           cir::BoolType::get(builder.getContext()), maskWidth);
+  mlir::Value maskVec = builder.createBitcast(mask, maskVecType);
+
+  // If we have less than 8 elements, then the starting mask was an i8 and
+  // we need to extract down to the right number of elements.
+  if (numElems < 8) {
+    llvm::SmallVector<mlir::Attribute, 4> indices;
+    mlir::Type i32Ty = builder.getSInt32Ty();
+    for (auto i : llvm::seq<unsigned>(0, numElems))
+      indices.push_back(cir::IntAttr::get(i32Ty, i));
+    cir::ConstantOp poison = builder.getConstant(
+        mask.getLoc(), cir::PoisonAttr::get(maskVec.getType()));
+    maskVec = builder.createVecShuffle(mask.getLoc(), maskVec, poison, indices);
+  }
+  return maskVec;
+}
+
+static mlir::Value emitKunpckOp(CIRGenBuilderTy &builder, mlir::Value op0,
+                                mlir::Value op1, mlir::Location loc) {
+  auto maskIntType = mlir::cast<cir::IntType>(op0.getType());
+  unsigned numElems = maskIntType.getWidth();
+  mlir::Value lhs = convertKunpckMaskToVector(builder, op0, numElems);
+  mlir::Value rhs = convertKunpckMaskToVector(builder, op1, numElems);
+
+  // Build shuffle indices as attributes to avoid redundant conversion.
+  // createVecShuffle will loop through the vector again to convert if we use
+  // int64_t.
+  llvm::SmallVector<mlir::Attribute, 64> indices;
+  mlir::Type i32Ty = builder.getSInt32Ty();
+  for (auto i : llvm::seq<unsigned>(0, numElems))
+    indices.push_back(cir::IntAttr::get(i32Ty, i));
+
+  // First extract half of each vector. This gives better codegen than
+  // doing it in a single shuffle.
+  // Use poison as second operand for unary shuffle to avoid conversion loop.
+  cir::ConstantOp poison =
+      builder.getConstant(loc, cir::PoisonAttr::get(lhs.getType()));
+  lhs = builder.createVecShuffle(loc, lhs, poison,
+                                 llvm::ArrayRef(indices.data(), numElems / 2));
+  rhs = builder.createVecShuffle(loc, rhs, poison,
+                                 llvm::ArrayRef(indices.data(), numElems / 2));
+
+  // Concat the vectors.
+  // NOTE: Operands are swapped to match the intrinsic definition.
+  mlir::Value res = builder.createVecShuffle(loc, rhs, lhs, indices);
+  return builder.createBitcast(res, op0.getType());
+}
+
 mlir::Value CIRGenFunction::emitX86BuiltinExpr(unsigned builtinID,
                                                const CallExpr *expr) {
   if (builtinID == Builtin::BI__builtin_cpu_is) {
@@ -186,6 +240,12 @@ mlir::Value CIRGenFunction::emitX86BuiltinExpr(unsigned builtinID,
                  std::string("unimplemented X86 builtin call: ") +
                      getContext().BuiltinInfo.getName(builtinID));
     return {};
+
+  case X86::BI__builtin_ia32_kunpckdi:
+  case X86::BI__builtin_ia32_kunpcksi:
+  case X86::BI__builtin_ia32_kunpckhi:
+    return emitKunpckOp(builder, ops[0], ops[1], getLoc(expr->getExprLoc()));
+
   case X86::BI_mm_setcsr:
   case X86::BI__builtin_ia32_ldmxcsr: {
     mlir::Location loc = getLoc(expr->getExprLoc());
@@ -704,9 +764,6 @@ mlir::Value CIRGenFunction::emitX86BuiltinExpr(unsigned builtinID,
   case X86::BI__builtin_ia32_kmovw:
   case X86::BI__builtin_ia32_kmovd:
   case X86::BI__builtin_ia32_kmovq:
-  case X86::BI__builtin_ia32_kunpckdi:
-  case X86::BI__builtin_ia32_kunpcksi:
-  case X86::BI__builtin_ia32_kunpckhi:
   case X86::BI__builtin_ia32_sqrtsh_round_mask:
   case X86::BI__builtin_ia32_sqrtsd_round_mask:
   case X86::BI__builtin_ia32_sqrtss_round_mask:

clang/test/CodeGen/X86/avx512bw-builtins.c

@@ -3152,3 +3152,25 @@ void test_mm512_mask_cvtusepi16_storeu_epi8 (void * __P, __mmask32 __M, __m512i
  // CHECK: @llvm.x86.avx512.mask.pmovus.wb.mem.512
  _mm512_mask_cvtusepi16_storeu_epi8 ( __P, __M, __A);
 }
+
+__mmask32 test_mm512_kunpackw(__mmask32 a, __mmask32 b) {
+  // CHECK-LABEL: test_mm512_kunpackw
+  // CHECK: [[A_BITCAST:%.*]] = bitcast i32 %{{.*}} to <32 x i1>
+  // CHECK: [[A_SHUFFLE:%.*]] = shufflevector <32 x i1> [[A_BITCAST]], <32 x i1> [[A_BITCAST]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  // CHECK: [[B_BITCAST:%.*]] = bitcast i32 %{{.*}} to <32 x i1>
+  // CHECK: [[B_SHUFFLE:%.*]] = shufflevector <32 x i1> [[B_BITCAST]], <32 x i1> [[B_BITCAST]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  // CHECK: [[CONCAT:%.*]] = shufflevector <16 x i1> [[B_SHUFFLE]], <16 x i1> [[A_SHUFFLE]], <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
+  // CHECK: bitcast <32 x i1> [[CONCAT]] to i32
+  return _mm512_kunpackw(a, b);
+}
+
+__mmask64 test_mm512_kunpackd(__mmask64 a, __mmask64 b) {
+  // CHECK-LABEL: test_mm512_kunpackd
+  // CHECK: [[A_BITCAST:%.*]] = bitcast i64 %{{.*}} to <64 x i1>
+  // CHECK: [[A_SHUFFLE:%.*]] = shufflevector <64 x i1> [[A_BITCAST]], <64 x i1> [[A_BITCAST]], <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
+  // CHECK: [[B_BITCAST:%.*]] = bitcast i64 %{{.*}} to <64 x i1>
+  // CHECK: [[B_SHUFFLE:%.*]] = shufflevector <64 x i1> [[B_BITCAST]], <64 x i1> [[B_BITCAST]], <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
+  // CHECK: [[CONCAT:%.*]] = shufflevector <32 x i1> [[B_SHUFFLE]], <32 x i1> [[A_SHUFFLE]], <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
+  // CHECK: bitcast <64 x i1> [[CONCAT]] to i64
+  return _mm512_kunpackd(a, b);
+}