[AVX512] Added intrinsics for 128-, 256- and 512-bit versions of VCMPGT{BWDQ}.

Patch by Sergey Lisitsyn <sergey.lisitsyn@intel.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@218670 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/IR/IntrinsicsX86.td b/include/llvm/IR/IntrinsicsX86.td
index 1e366d17cd738369eaf8575308ed6ba383c88fd1..538d7a6cb89268d29cabcdaf123686bd94601828 100644 (file)
--- a/include/llvm/IR/IntrinsicsX86.td
+++ b/include/llvm/IR/IntrinsicsX86.td
@@ -3250,6 +3250,19 @@ let TargetPrefix = "x86" in {
             Intrinsic<[llvm_i8_ty], [llvm_v8i64_ty, llvm_v8i64_ty, llvm_i8_ty],
                       [IntrNoMem]>;
 
+  def int_x86_avx512_mask_pcmpgt_b_512: GCCBuiltin<"__builtin_ia32_pcmpgtb512_mask">,
+        Intrinsic<[llvm_i64_ty], [llvm_v64i8_ty, llvm_v64i8_ty, llvm_i64_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_w_512: GCCBuiltin<"__builtin_ia32_pcmpgtw512_mask">,
+        Intrinsic<[llvm_i32_ty], [llvm_v32i16_ty, llvm_v32i16_ty, llvm_i32_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_d_512: GCCBuiltin<"__builtin_ia32_pcmpgtd512_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i32_ty, llvm_v16i32_ty, llvm_i16_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_q_512: GCCBuiltin<"__builtin_ia32_pcmpgtq512_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i64_ty, llvm_v8i64_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+
   // 256-bit
   def int_x86_avx512_mask_pcmpeq_b_256 : GCCBuiltin<"__builtin_ia32_pcmpeqb256_mask">,
         Intrinsic<[llvm_i32_ty], [llvm_v32i8_ty, llvm_v32i8_ty, llvm_i32_ty],
@@ -3264,6 +3277,19 @@ let TargetPrefix = "x86" in {
         Intrinsic<[llvm_i8_ty], [llvm_v4i64_ty, llvm_v4i64_ty, llvm_i8_ty],
                   [IntrNoMem]>;
 
+  def int_x86_avx512_mask_pcmpgt_b_256: GCCBuiltin<"__builtin_ia32_pcmpgtb256_mask">,
+        Intrinsic<[llvm_i32_ty], [llvm_v32i8_ty, llvm_v32i8_ty, llvm_i32_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_w_256: GCCBuiltin<"__builtin_ia32_pcmpgtw256_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i16_ty, llvm_v16i16_ty, llvm_i16_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_d_256: GCCBuiltin<"__builtin_ia32_pcmpgtd256_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i32_ty, llvm_v8i32_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_q_256: GCCBuiltin<"__builtin_ia32_pcmpgtq256_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v4i64_ty, llvm_v4i64_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+
   // 128-bit
   def int_x86_avx512_mask_pcmpeq_b_128 : GCCBuiltin<"__builtin_ia32_pcmpeqb128_mask">,
         Intrinsic<[llvm_i16_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i16_ty],
@@ -3277,6 +3303,19 @@ let TargetPrefix = "x86" in {
   def int_x86_avx512_mask_pcmpeq_q_128 : GCCBuiltin<"__builtin_ia32_pcmpeqq128_mask">,
         Intrinsic<[llvm_i8_ty], [llvm_v2i64_ty, llvm_v2i64_ty, llvm_i8_ty],
                   [IntrNoMem]>;
+
+  def int_x86_avx512_mask_pcmpgt_b_128: GCCBuiltin<"__builtin_ia32_pcmpgtb128_mask">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i16_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_w_128: GCCBuiltin<"__builtin_ia32_pcmpgtw128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i16_ty, llvm_v8i16_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_d_128: GCCBuiltin<"__builtin_ia32_pcmpgtd128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v4i32_ty, llvm_v4i32_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
+  def int_x86_avx512_mask_pcmpgt_q_128: GCCBuiltin<"__builtin_ia32_pcmpgtq128_mask">,
+        Intrinsic<[llvm_i8_ty], [llvm_v2i64_ty, llvm_v2i64_ty, llvm_i8_ty],
+                  [IntrNoMem]>;
 }
 
 // Misc.

diff --git a/lib/Target/X86/X86IntrinsicsInfo.h b/lib/Target/X86/X86IntrinsicsInfo.h
index 35d6e89ce90bc8c04635ccd2976092141d7af9a5..8fa6cd2e13c2251b41cd25b530256e87f8c85e58 100644 (file)
--- a/lib/Target/X86/X86IntrinsicsInfo.h
+++ b/lib/Target/X86/X86IntrinsicsInfo.h
@@ -168,6 +168,18 @@ static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(avx512_mask_pcmpeq_w_128,  CMP_MASK,  X86ISD::PCMPEQM, 0),
   X86_INTRINSIC_DATA(avx512_mask_pcmpeq_w_256,  CMP_MASK,  X86ISD::PCMPEQM, 0),
   X86_INTRINSIC_DATA(avx512_mask_pcmpeq_w_512,  CMP_MASK,  X86ISD::PCMPEQM, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pcmpgt_b_128,  CMP_MASK,  X86ISD::PCMPGTM, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pcmpgt_b_256,  CMP_MASK,  X86ISD::PCMPGTM, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pcmpgt_b_512,  CMP_MASK,  X86ISD::PCMPGTM, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pcmpgt_d_128,  CMP_MASK,  X86ISD::PCMPGTM, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pcmpgt_d_256,  CMP_MASK,  X86ISD::PCMPGTM, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pcmpgt_d_512,  CMP_MASK,  X86ISD::PCMPGTM, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pcmpgt_q_128,  CMP_MASK,  X86ISD::PCMPGTM, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pcmpgt_q_256,  CMP_MASK,  X86ISD::PCMPGTM, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pcmpgt_q_512,  CMP_MASK,  X86ISD::PCMPGTM, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pcmpgt_w_128,  CMP_MASK,  X86ISD::PCMPGTM, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pcmpgt_w_256,  CMP_MASK,  X86ISD::PCMPGTM, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pcmpgt_w_512,  CMP_MASK,  X86ISD::PCMPGTM, 0),
   X86_INTRINSIC_DATA(avx_hadd_pd_256,   INTR_TYPE_2OP, X86ISD::FHADD, 0),
   X86_INTRINSIC_DATA(avx_hadd_ps_256,   INTR_TYPE_2OP, X86ISD::FHADD, 0),
   X86_INTRINSIC_DATA(avx_hsub_pd_256,   INTR_TYPE_2OP, X86ISD::FHSUB, 0),

diff --git a/test/CodeGen/X86/avx512-intrinsics.ll b/test/CodeGen/X86/avx512-intrinsics.ll
index 38d8da78a1fe3fb78044fc82fc0a3a035176a553..d63a324943466ddbe22a56f747b5af111462db9a 100644 (file)
--- a/test/CodeGen/X86/avx512-intrinsics.ll
+++ b/test/CodeGen/X86/avx512-intrinsics.ll
@@ -677,3 +677,35 @@ define i8 @test_mask_pcmpeq_q(<8 x i64> %a, <8 x i64> %b, i8 %mask) {
 }
 
 declare i8 @llvm.x86.avx512.mask.pcmpeq.q.512(<8 x i64>, <8 x i64>, i8)
+
+define i16 @test_pcmpgt_d(<16 x i32> %a, <16 x i32> %b) {
+; CHECK-LABEL: test_pcmpgt_d
+; CHECK: vpcmpgtd %zmm1, %zmm0, %k0 ##
+  %res = call i16 @llvm.x86.avx512.mask.pcmpgt.d.512(<16 x i32> %a, <16 x i32> %b, i16 -1)
+  ret i16 %res
+}
+
+define i16 @test_mask_pcmpgt_d(<16 x i32> %a, <16 x i32> %b, i16 %mask) {
+; CHECK-LABEL: test_mask_pcmpgt_d
+; CHECK: vpcmpgtd %zmm1, %zmm0, %k0 {%k1} ##
+  %res = call i16 @llvm.x86.avx512.mask.pcmpgt.d.512(<16 x i32> %a, <16 x i32> %b, i16 %mask)
+  ret i16 %res
+}
+
+declare i16 @llvm.x86.avx512.mask.pcmpgt.d.512(<16 x i32>, <16 x i32>, i16)
+
+define i8 @test_pcmpgt_q(<8 x i64> %a, <8 x i64> %b) {
+; CHECK-LABEL: test_pcmpgt_q
+; CHECK: vpcmpgtq %zmm1, %zmm0, %k0 ##
+  %res = call i8 @llvm.x86.avx512.mask.pcmpgt.q.512(<8 x i64> %a, <8 x i64> %b, i8 -1)
+  ret i8 %res
+}
+
+define i8 @test_mask_pcmpgt_q(<8 x i64> %a, <8 x i64> %b, i8 %mask) {
+; CHECK-LABEL: test_mask_pcmpgt_q
+; CHECK: vpcmpgtq %zmm1, %zmm0, %k0 {%k1} ##
+  %res = call i8 @llvm.x86.avx512.mask.pcmpgt.q.512(<8 x i64> %a, <8 x i64> %b, i8 %mask)
+  ret i8 %res
+}
+
+declare i8 @llvm.x86.avx512.mask.pcmpgt.q.512(<8 x i64>, <8 x i64>, i8)

diff --git a/test/CodeGen/X86/avx512bw-intrinsics.ll b/test/CodeGen/X86/avx512bw-intrinsics.ll
index a4ebaeb583415725a07d4602453abfb4bf38dbc5..d74f444b20d019728f28a40f8bf8412e5101376b 100644 (file)
--- a/test/CodeGen/X86/avx512bw-intrinsics.ll
+++ b/test/CodeGen/X86/avx512bw-intrinsics.ll
@@ -31,3 +31,35 @@ define i32 @test_mask_pcmpeq_w(<32 x i16> %a, <32 x i16> %b, i32 %mask) {
 }
 
 declare i32 @llvm.x86.avx512.mask.pcmpeq.w.512(<32 x i16>, <32 x i16>, i32)
+
+define i64 @test_pcmpgt_b(<64 x i8> %a, <64 x i8> %b) {
+; CHECK-LABEL: test_pcmpgt_b
+; CHECK: vpcmpgtb %zmm1, %zmm0, %k0 ##
+  %res = call i64 @llvm.x86.avx512.mask.pcmpgt.b.512(<64 x i8> %a, <64 x i8> %b, i64 -1)
+  ret i64 %res
+}
+
+define i64 @test_mask_pcmpgt_b(<64 x i8> %a, <64 x i8> %b, i64 %mask) {
+; CHECK-LABEL: test_mask_pcmpgt_b
+; CHECK: vpcmpgtb %zmm1, %zmm0, %k0 {%k1} ##
+  %res = call i64 @llvm.x86.avx512.mask.pcmpgt.b.512(<64 x i8> %a, <64 x i8> %b, i64 %mask)
+  ret i64 %res
+}
+
+declare i64 @llvm.x86.avx512.mask.pcmpgt.b.512(<64 x i8>, <64 x i8>, i64)
+
+define i32 @test_pcmpgt_w(<32 x i16> %a, <32 x i16> %b) {
+; CHECK-LABEL: test_pcmpgt_w
+; CHECK: vpcmpgtw %zmm1, %zmm0, %k0 ##
+  %res = call i32 @llvm.x86.avx512.mask.pcmpgt.w.512(<32 x i16> %a, <32 x i16> %b, i32 -1)
+  ret i32 %res
+}
+
+define i32 @test_mask_pcmpgt_w(<32 x i16> %a, <32 x i16> %b, i32 %mask) {
+; CHECK-LABEL: test_mask_pcmpgt_w
+; CHECK: vpcmpgtw %zmm1, %zmm0, %k0 {%k1} ##
+  %res = call i32 @llvm.x86.avx512.mask.pcmpgt.w.512(<32 x i16> %a, <32 x i16> %b, i32 %mask)
+  ret i32 %res
+}
+
+declare i32 @llvm.x86.avx512.mask.pcmpgt.w.512(<32 x i16>, <32 x i16>, i32)

diff --git a/test/CodeGen/X86/avx512bwvl-intrinsics.ll b/test/CodeGen/X86/avx512bwvl-intrinsics.ll
index 06cc91452a74475fa6d836e87bb0d9efdf4c07c3..798bc3598ec9d35ed772bebbe156d6b1b59d826e 100644 (file)
--- a/test/CodeGen/X86/avx512bwvl-intrinsics.ll
+++ b/test/CodeGen/X86/avx512bwvl-intrinsics.ll
@@ -34,6 +34,38 @@ define i16 @test_mask_pcmpeq_w_256(<16 x i16> %a, <16 x i16> %b, i16 %mask) {
 
 declare i16 @llvm.x86.avx512.mask.pcmpeq.w.256(<16 x i16>, <16 x i16>, i16)
 
+define i32 @test_pcmpgt_b_256(<32 x i8> %a, <32 x i8> %b) {
+; CHECK-LABEL: test_pcmpgt_b_256
+; CHECK: vpcmpgtb %ymm1, %ymm0, %k0 ##
+  %res = call i32 @llvm.x86.avx512.mask.pcmpgt.b.256(<32 x i8> %a, <32 x i8> %b, i32 -1)
+  ret i32 %res
+}
+
+define i32 @test_mask_pcmpgt_b_256(<32 x i8> %a, <32 x i8> %b, i32 %mask) {
+; CHECK-LABEL: test_mask_pcmpgt_b_256
+; CHECK: vpcmpgtb %ymm1, %ymm0, %k0 {%k1} ##
+  %res = call i32 @llvm.x86.avx512.mask.pcmpgt.b.256(<32 x i8> %a, <32 x i8> %b, i32 %mask)
+  ret i32 %res
+}
+
+declare i32 @llvm.x86.avx512.mask.pcmpgt.b.256(<32 x i8>, <32 x i8>, i32)
+
+define i16 @test_pcmpgt_w_256(<16 x i16> %a, <16 x i16> %b) {
+; CHECK-LABEL: test_pcmpgt_w_256
+; CHECK: vpcmpgtw %ymm1, %ymm0, %k0 ##
+  %res = call i16 @llvm.x86.avx512.mask.pcmpgt.w.256(<16 x i16> %a, <16 x i16> %b, i16 -1)
+  ret i16 %res
+}
+
+define i16 @test_mask_pcmpgt_w_256(<16 x i16> %a, <16 x i16> %b, i16 %mask) {
+; CHECK-LABEL: test_mask_pcmpgt_w_256
+; CHECK: vpcmpgtw %ymm1, %ymm0, %k0 {%k1} ##
+  %res = call i16 @llvm.x86.avx512.mask.pcmpgt.w.256(<16 x i16> %a, <16 x i16> %b, i16 %mask)
+  ret i16 %res
+}
+
+declare i16 @llvm.x86.avx512.mask.pcmpgt.w.256(<16 x i16>, <16 x i16>, i16)
+
 ; 128-bit
 
 define i16 @test_pcmpeq_b_128(<16 x i8> %a, <16 x i8> %b) {
@@ -68,3 +100,34 @@ define i8 @test_mask_pcmpeq_w_128(<8 x i16> %a, <8 x i16> %b, i8 %mask) {
 
 declare i8 @llvm.x86.avx512.mask.pcmpeq.w.128(<8 x i16>, <8 x i16>, i8)
 
+define i16 @test_pcmpgt_b_128(<16 x i8> %a, <16 x i8> %b) {
+; CHECK-LABEL: test_pcmpgt_b_128
+; CHECK: vpcmpgtb %xmm1, %xmm0, %k0 ##
+  %res = call i16 @llvm.x86.avx512.mask.pcmpgt.b.128(<16 x i8> %a, <16 x i8> %b, i16 -1)
+  ret i16 %res
+}
+
+define i16 @test_mask_pcmpgt_b_128(<16 x i8> %a, <16 x i8> %b, i16 %mask) {
+; CHECK-LABEL: test_mask_pcmpgt_b_128
+; CHECK: vpcmpgtb %xmm1, %xmm0, %k0 {%k1} ##
+  %res = call i16 @llvm.x86.avx512.mask.pcmpgt.b.128(<16 x i8> %a, <16 x i8> %b, i16 %mask)
+  ret i16 %res
+}
+
+declare i16 @llvm.x86.avx512.mask.pcmpgt.b.128(<16 x i8>, <16 x i8>, i16)
+
+define i8 @test_pcmpgt_w_128(<8 x i16> %a, <8 x i16> %b) {
+; CHECK-LABEL: test_pcmpgt_w_128
+; CHECK: vpcmpgtw %xmm1, %xmm0, %k0 ##
+  %res = call i8 @llvm.x86.avx512.mask.pcmpgt.w.128(<8 x i16> %a, <8 x i16> %b, i8 -1)
+  ret i8 %res
+}
+
+define i8 @test_mask_pcmpgt_w_128(<8 x i16> %a, <8 x i16> %b, i8 %mask) {
+; CHECK-LABEL: test_mask_pcmpgt_w_128
+; CHECK: vpcmpgtw %xmm1, %xmm0, %k0 {%k1} ##
+  %res = call i8 @llvm.x86.avx512.mask.pcmpgt.w.128(<8 x i16> %a, <8 x i16> %b, i8 %mask)
+  ret i8 %res
+}
+
+declare i8 @llvm.x86.avx512.mask.pcmpgt.w.128(<8 x i16>, <8 x i16>, i8)

diff --git a/test/CodeGen/X86/avx512vl-intrinsics.ll b/test/CodeGen/X86/avx512vl-intrinsics.ll
index ada13dea249ca295430cb40736d27a43ce910f00..c24b514c6c362aa72f5dccbbe88e448e0f86a2a8 100644 (file)
--- a/test/CodeGen/X86/avx512vl-intrinsics.ll
+++ b/test/CodeGen/X86/avx512vl-intrinsics.ll
@@ -34,6 +34,38 @@ define i8 @test_mask_pcmpeq_q_256(<4 x i64> %a, <4 x i64> %b, i8 %mask) {
 
 declare i8 @llvm.x86.avx512.mask.pcmpeq.q.256(<4 x i64>, <4 x i64>, i8)
 
+define i8 @test_pcmpgt_d_256(<8 x i32> %a, <8 x i32> %b) {
+; CHECK-LABEL: test_pcmpgt_d_256
+; CHECK: vpcmpgtd %ymm1, %ymm0, %k0 ##
+  %res = call i8 @llvm.x86.avx512.mask.pcmpgt.d.256(<8 x i32> %a, <8 x i32> %b, i8 -1)
+  ret i8 %res
+}
+
+define i8 @test_mask_pcmpgt_d_256(<8 x i32> %a, <8 x i32> %b, i8 %mask) {
+; CHECK-LABEL: test_mask_pcmpgt_d_256
+; CHECK: vpcmpgtd %ymm1, %ymm0, %k0 {%k1} ##
+  %res = call i8 @llvm.x86.avx512.mask.pcmpgt.d.256(<8 x i32> %a, <8 x i32> %b, i8 %mask)
+  ret i8 %res
+}
+
+declare i8 @llvm.x86.avx512.mask.pcmpgt.d.256(<8 x i32>, <8 x i32>, i8)
+
+define i8 @test_pcmpgt_q_256(<4 x i64> %a, <4 x i64> %b) {
+; CHECK-LABEL: test_pcmpgt_q_256
+; CHECK: vpcmpgtq %ymm1, %ymm0, %k0 ##
+  %res = call i8 @llvm.x86.avx512.mask.pcmpgt.q.256(<4 x i64> %a, <4 x i64> %b, i8 -1)
+  ret i8 %res
+}
+
+define i8 @test_mask_pcmpgt_q_256(<4 x i64> %a, <4 x i64> %b, i8 %mask) {
+; CHECK-LABEL: test_mask_pcmpgt_q_256
+; CHECK: vpcmpgtq %ymm1, %ymm0, %k0 {%k1} ##
+  %res = call i8 @llvm.x86.avx512.mask.pcmpgt.q.256(<4 x i64> %a, <4 x i64> %b, i8 %mask)
+  ret i8 %res
+}
+
+declare i8 @llvm.x86.avx512.mask.pcmpgt.q.256(<4 x i64>, <4 x i64>, i8)
+
 ; 128-bit
 
 define i8 @test_pcmpeq_d_128(<4 x i32> %a, <4 x i32> %b) {
@@ -67,3 +99,35 @@ define i8 @test_mask_pcmpeq_q_128(<2 x i64> %a, <2 x i64> %b, i8 %mask) {
 }
 
 declare i8 @llvm.x86.avx512.mask.pcmpeq.q.128(<2 x i64>, <2 x i64>, i8)
+
+define i8 @test_pcmpgt_d_128(<4 x i32> %a, <4 x i32> %b) {
+; CHECK-LABEL: test_pcmpgt_d_128
+; CHECK: vpcmpgtd %xmm1, %xmm0, %k0 ##
+  %res = call i8 @llvm.x86.avx512.mask.pcmpgt.d.128(<4 x i32> %a, <4 x i32> %b, i8 -1)
+  ret i8 %res
+}
+
+define i8 @test_mask_pcmpgt_d_128(<4 x i32> %a, <4 x i32> %b, i8 %mask) {
+; CHECK-LABEL: test_mask_pcmpgt_d_128
+; CHECK: vpcmpgtd %xmm1, %xmm0, %k0 {%k1} ##
+  %res = call i8 @llvm.x86.avx512.mask.pcmpgt.d.128(<4 x i32> %a, <4 x i32> %b, i8 %mask)
+  ret i8 %res
+}
+
+declare i8 @llvm.x86.avx512.mask.pcmpgt.d.128(<4 x i32>, <4 x i32>, i8)
+
+define i8 @test_pcmpgt_q_128(<2 x i64> %a, <2 x i64> %b) {
+; CHECK-LABEL: test_pcmpgt_q_128
+; CHECK: vpcmpgtq %xmm1, %xmm0, %k0 ##
+  %res = call i8 @llvm.x86.avx512.mask.pcmpgt.q.128(<2 x i64> %a, <2 x i64> %b, i8 -1)
+  ret i8 %res
+}
+
+define i8 @test_mask_pcmpgt_q_128(<2 x i64> %a, <2 x i64> %b, i8 %mask) {
+; CHECK-LABEL: test_mask_pcmpgt_q_128
+; CHECK: vpcmpgtq %xmm1, %xmm0, %k0 {%k1} ##
+  %res = call i8 @llvm.x86.avx512.mask.pcmpgt.q.128(<2 x i64> %a, <2 x i64> %b, i8 %mask)
+  ret i8 %res
+}
+
+declare i8 @llvm.x86.avx512.mask.pcmpgt.q.128(<2 x i64>, <2 x i64>, i8)