AVX-512: changed intrinsics for mask operations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@196918 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/IR/IntrinsicsX86.td

@@ -2643,37 +2643,30 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   // Mask instructions
   // 16-bit mask
-  def int_x86_kadd_v16i1 : GCCBuiltin<"__builtin_ia32_kaddw">,
+  def int_x86_avx512_kand_w : GCCBuiltin<"__builtin_ia32_kandhi">,
-              Intrinsic<[llvm_v16i1_ty], [llvm_v16i1_ty, llvm_v16i1_ty],
+              Intrinsic<[llvm_i16_ty], [llvm_i16_ty, llvm_i16_ty],
                          [IntrNoMem]>;
-  def int_x86_kand_v16i1 : GCCBuiltin<"__builtin_ia32_kandw">,
+  def int_x86_avx512_kandn_w : GCCBuiltin<"__builtin_ia32_kandnhi">,
-              Intrinsic<[llvm_v16i1_ty], [llvm_v16i1_ty, llvm_v16i1_ty],
+              Intrinsic<[llvm_i16_ty], [llvm_i16_ty, llvm_i16_ty],
                          [IntrNoMem]>;
-  def int_x86_kandn_v16i1 : GCCBuiltin<"__builtin_ia32_kandnw">,
+  def int_x86_avx512_knot_w : GCCBuiltin<"__builtin_ia32_knothi">,
-              Intrinsic<[llvm_v16i1_ty], [llvm_v16i1_ty, llvm_v16i1_ty],
+              Intrinsic<[llvm_i16_ty], [llvm_i16_ty], [IntrNoMem]>;
+  def int_x86_avx512_kor_w : GCCBuiltin<"__builtin_ia32_korhi">,
+              Intrinsic<[llvm_i16_ty], [llvm_i16_ty, llvm_i16_ty],
                          [IntrNoMem]>;
-  def int_x86_knot_v16i1 : GCCBuiltin<"__builtin_ia32_knotw">,
+  def int_x86_avx512_kxor_w : GCCBuiltin<"__builtin_ia32_kxorhi">,
-              Intrinsic<[llvm_v16i1_ty], [llvm_v16i1_ty], [IntrNoMem]>;
+              Intrinsic<[llvm_i16_ty], [llvm_i16_ty, llvm_i16_ty],
-  def int_x86_kor_v16i1 : GCCBuiltin<"__builtin_ia32_korw">,
-              Intrinsic<[llvm_v16i1_ty], [llvm_v16i1_ty, llvm_v16i1_ty],
                          [IntrNoMem]>;
-  def int_x86_kxor_v16i1 : GCCBuiltin<"__builtin_ia32_kxorw">,
+  def int_x86_avx512_kxnor_w : GCCBuiltin<"__builtin_ia32_kxnorhi">,
-              Intrinsic<[llvm_v16i1_ty], [llvm_v16i1_ty, llvm_v16i1_ty],
+              Intrinsic<[llvm_i16_ty], [llvm_i16_ty, llvm_i16_ty],
                          [IntrNoMem]>;
-  def int_x86_kxnor_v16i1 : GCCBuiltin<"__builtin_ia32_kxnorw">,
+  def int_x86_avx512_kunpck_bw : GCCBuiltin<"__builtin_ia32_kunpckhi">,
-              Intrinsic<[llvm_v16i1_ty], [llvm_v16i1_ty, llvm_v16i1_ty],
+              Intrinsic<[llvm_i16_ty], [llvm_i16_ty, llvm_i16_ty],
                          [IntrNoMem]>;
-  def int_x86_mask2int_v16i1 : GCCBuiltin<"__builtin_ia32_mask2intw">,
+  def int_x86_avx512_kortestz_w : GCCBuiltin<"__builtin_ia32_kortestzhi">,
-              Intrinsic<[llvm_i32_ty], [llvm_v16i1_ty], [IntrNoMem]>;
-  def int_x86_int2mask_v16i1 : GCCBuiltin<"__builtin_ia32_int2maskw">,
-              Intrinsic<[llvm_v16i1_ty], [llvm_i32_ty], [IntrNoMem]>;
-  def int_x86_kunpck_v16i1 : GCCBuiltin<"__builtin_ia32_kunpckbw">,
-              Intrinsic<[llvm_v16i1_ty], [llvm_v8i1_ty, llvm_v8i1_ty],
-                         [IntrNoMem]>;
-  def int_x86_avx512_kortestz : GCCBuiltin<"__builtin_ia32_kortestz">,
               Intrinsic<[llvm_i32_ty], [llvm_i16_ty, llvm_i16_ty],
                         [IntrNoMem]>;
-  def int_x86_avx512_kortestc : GCCBuiltin<"__builtin_ia32_kortestc">,
+  def int_x86_avx512_kortestc_w : GCCBuiltin<"__builtin_ia32_kortestchi">,
               Intrinsic<[llvm_i32_ty], [llvm_i16_ty, llvm_i16_ty],
                         [IntrNoMem]>;
 }

lib/Target/X86/X86ISelLowering.cpp

@@ -11490,9 +11490,9 @@ static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
     SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
     return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
   }
-  case Intrinsic::x86_avx512_kortestz:
+  case Intrinsic::x86_avx512_kortestz_w:
-  case Intrinsic::x86_avx512_kortestc: {
+  case Intrinsic::x86_avx512_kortestc_w: {
-    unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz)? X86::COND_E: X86::COND_B;
+    unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
     SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
     SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
     SDValue CC = DAG.getConstant(X86CC, MVT::i8);

lib/Target/X86/X86InstrAVX512.td

@@ -898,6 +898,15 @@ multiclass avx512_mask_unop_w<bits<8> opc, string OpcodeStr,
 
 defm KNOT : avx512_mask_unop_w<0x44, "knot", not>;
 
+multiclass avx512_mask_unop_int<string IntName, string InstName> {
+  let Predicates = [HasAVX512] in
+    def : Pat<(!cast<Intrinsic>("int_x86_avx512_"##IntName##"_w")
+                (i16 GR16:$src)),
+              (COPY_TO_REGCLASS (!cast<Instruction>(InstName##"Wrr")
+              (v16i1 (COPY_TO_REGCLASS GR16:$src, VK16))), GR16)>;
+}
+defm : avx512_mask_unop_int<"knot", "KNOT">;
+
 def : Pat<(xor VK16:$src1, (v16i1 immAllOnesV)), (KNOTWrr VK16:$src1)>;
 def : Pat<(xor VK8:$src1,  (v8i1 immAllOnesV)),
           (COPY_TO_REGCLASS (KNOTWrr (COPY_TO_REGCLASS VK8:$src1, VK16)), VK8)>;
@@ -908,7 +917,7 @@ def : Pat<(not VK8:$src),
             (KNOTWrr (COPY_TO_REGCLASS VK8:$src, VK16)), VK8)>;
 
 // Mask binary operation
-// - KADD, KAND, KANDN, KOR, KXNOR, KXOR
+// - KAND, KANDN, KOR, KXNOR, KXOR
 multiclass avx512_mask_binop<bits<8> opc, string OpcodeStr,
                            RegisterClass KRC, SDPatternOperator OpNode> {
   let Predicates = [HasAVX512] in
@@ -928,7 +937,6 @@ def andn : PatFrag<(ops node:$i0, node:$i1), (and (not node:$i0), node:$i1)>;
 def xnor : PatFrag<(ops node:$i0, node:$i1), (not (xor node:$i0, node:$i1))>;
 
 let isCommutable = 1 in {
-  defm KADD  : avx512_mask_binop_w<0x4a, "kadd",  add>;
   defm KAND  : avx512_mask_binop_w<0x41, "kand",  and>;
   let isCommutable = 0 in
   defm KANDN : avx512_mask_binop_w<0x42, "kandn", andn>;
@@ -939,17 +947,19 @@ let isCommutable = 1 in {
 
 multiclass avx512_mask_binop_int<string IntName, string InstName> {
   let Predicates = [HasAVX512] in
-    def : Pat<(!cast<Intrinsic>("int_x86_"##IntName##"_v16i1")
+    def : Pat<(!cast<Intrinsic>("int_x86_avx512_"##IntName##"_w")
-                VK16:$src1, VK16:$src2),
+                (i16 GR16:$src1), (i16 GR16:$src2)),
-              (!cast<Instruction>(InstName##"Wrr") VK16:$src1, VK16:$src2)>;
+              (COPY_TO_REGCLASS (!cast<Instruction>(InstName##"Wrr")
+              (v16i1 (COPY_TO_REGCLASS GR16:$src1, VK16)),
+              (v16i1 (COPY_TO_REGCLASS GR16:$src2, VK16))), GR16)>;
 }
 
-defm : avx512_mask_binop_int<"kadd",  "KADD">;
 defm : avx512_mask_binop_int<"kand",  "KAND">;
 defm : avx512_mask_binop_int<"kandn", "KANDN">;
 defm : avx512_mask_binop_int<"kor",   "KOR">;
 defm : avx512_mask_binop_int<"kxnor", "KXNOR">;
 defm : avx512_mask_binop_int<"kxor",  "KXOR">;
+
 // With AVX-512, 8-bit mask is promoted to 16-bit mask.
 multiclass avx512_binop_pat<SDPatternOperator OpNode, Instruction Inst> {
   let Predicates = [HasAVX512] in
@@ -967,15 +977,15 @@ defm : avx512_binop_pat<xor,  KXORWrr>;
 
 // Mask unpacking
 multiclass avx512_mask_unpck<bits<8> opc, string OpcodeStr,
-                           RegisterClass KRC1, RegisterClass KRC2> {
+                           RegisterClass KRC> {
   let Predicates = [HasAVX512] in
-    def rr : I<opc, MRMSrcReg, (outs KRC1:$dst), (ins KRC2:$src1, KRC2:$src2),
+    def rr : I<opc, MRMSrcReg, (outs KRC:$dst), (ins KRC:$src1, KRC:$src2),
                !strconcat(OpcodeStr,
                           "\t{$src2, $src1, $dst|$dst, $src1, $src2}"), []>;
 }
 
 multiclass avx512_mask_unpck_bw<bits<8> opc, string OpcodeStr> {
-  defm BW : avx512_mask_unpck<opc, !strconcat(OpcodeStr, "bw"), VK16, VK8>,
+  defm BW : avx512_mask_unpck<opc, !strconcat(OpcodeStr, "bw"), VK16>,
                             VEX_4V, VEX_L, OpSize, TB;
 }
 
@@ -983,12 +993,14 @@ defm KUNPCK : avx512_mask_unpck_bw<0x4b, "kunpck">;
 
 multiclass avx512_mask_unpck_int<string IntName, string InstName> {
   let Predicates = [HasAVX512] in
-    def : Pat<(!cast<Intrinsic>("int_x86_"##IntName##"_v16i1")
+    def : Pat<(!cast<Intrinsic>("int_x86_avx512_"##IntName##"_bw")
-                VK8:$src1, VK8:$src2),
+                (i16 GR16:$src1), (i16 GR16:$src2)),
-              (!cast<Instruction>(InstName##"BWrr") VK8:$src1, VK8:$src2)>;
+              (COPY_TO_REGCLASS (!cast<Instruction>(InstName##"BWrr")
+              (v16i1 (COPY_TO_REGCLASS GR16:$src1, VK16)),
+              (v16i1 (COPY_TO_REGCLASS GR16:$src2, VK16))), GR16)>;
 }
+defm : avx512_mask_unpck_int<"kunpck",  "KUNPCK">;
 
-defm : avx512_mask_unpck_int<"kunpck", "KUNPCK">;
 // Mask bit testing
 multiclass avx512_mask_testop<bits<8> opc, string OpcodeStr, RegisterClass KRC,
                             SDNode OpNode> {

test/CodeGen/X86/avx512-intrinsics.ll

@@ -1,23 +1,51 @@
 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl | FileCheck %s
 
-declare i32 @llvm.x86.avx512.kortestz(i16, i16) nounwind readnone
+declare i32 @llvm.x86.avx512.kortestz.w(i16, i16) nounwind readnone
-; CHECK: test_kortestz
+; CHECK-LABEL: test_kortestz
 ; CHECK: kortestw
 ; CHECK: sete
 define i32 @test_kortestz(i16 %a0, i16 %a1) {
-  %res = call i32 @llvm.x86.avx512.kortestz(i16 %a0, i16 %a1) 
+  %res = call i32 @llvm.x86.avx512.kortestz.w(i16 %a0, i16 %a1) 
   ret i32 %res
 }
 
-declare i32 @llvm.x86.avx512.kortestc(i16, i16) nounwind readnone
+declare i32 @llvm.x86.avx512.kortestc.w(i16, i16) nounwind readnone
-; CHECK: test_kortestc
+; CHECK-LABEL: test_kortestc
 ; CHECK: kortestw
 ; CHECK: sbbl
 define i32 @test_kortestc(i16 %a0, i16 %a1) {
-  %res = call i32 @llvm.x86.avx512.kortestc(i16 %a0, i16 %a1) 
+  %res = call i32 @llvm.x86.avx512.kortestc.w(i16 %a0, i16 %a1) 
   ret i32 %res
 }
 
+declare i16 @llvm.x86.avx512.kand.w(i16, i16) nounwind readnone
+; CHECK-LABEL: test_kand
+; CHECK: kandw
+; CHECK: kandw
+define i16 @test_kand(i16 %a0, i16 %a1) {
+  %t1 = call i16 @llvm.x86.avx512.kand.w(i16 %a0, i16 8)
+  %t2 = call i16 @llvm.x86.avx512.kand.w(i16 %t1, i16 %a1)
+  ret i16 %t2
+}
+
+declare i16 @llvm.x86.avx512.knot.w(i16) nounwind readnone
+; CHECK-LABEL: test_knot
+; CHECK: knotw
+define i16 @test_knot(i16 %a0) {
+  %res = call i16 @llvm.x86.avx512.knot.w(i16 %a0)
+  ret i16 %res
+}
+
+declare i16 @llvm.x86.avx512.kunpck.bw(i16, i16) nounwind readnone
+
+; CHECK-LABEL: unpckbw_test
+; CHECK: kunpckbw
+; CHECK:ret
+define i16 @unpckbw_test(i16 %a0, i16 %a1) {
+  %res = call i16 @llvm.x86.avx512.kunpck.bw(i16 %a0, i16 %a1)
+  ret i16 %res
+}
+
 define <16 x float> @test_rcp_ps_512(<16 x float> %a0) {
   ; CHECK: vrcp14ps
   %res = call <16 x float> @llvm.x86.avx512.rcp14.ps.512(<16 x float> %a0) ; <<16 x float>> [#uses=1]

test/CodeGen/X86/avx512-mask-op.ll

@@ -33,19 +33,6 @@ define i16 @mand16(i16 %x, i16 %y) {
   ret i16 %ret
 }
 
-; CHECK: unpckbw_test
-; CHECK: kunpckbw
-; CHECK:ret
-declare <16 x i1> @llvm.x86.kunpck.v16i1(<8 x i1>, <8 x i1>) nounwind readnone
-
-define i16 @unpckbw_test(i8 %x, i8 %y) {
-  %m0 = bitcast i8 %x to <8 x i1>
-  %m1 = bitcast i8 %y to <8 x i1>
-  %k = tail call <16 x i1> @llvm.x86.kunpck.v16i1(<8 x i1> %m0, <8 x i1> %m1)
-  %r = bitcast <16 x i1> %k to i16
-  ret i16 %r
-}
-
 ; CHECK: shuf_test1
 ; CHECK: kshiftrw        $8
 ; CHECK:ret