[AVX512] Adding PTESTMB/W/D/Q instruction

Differential Revision: http://reviews.llvm.org/D16519

llvm-svn: 258686

diff --git a/llvm/include/llvm/IR/IntrinsicsX86.td b/llvm/include/llvm/IR/IntrinsicsX86.td
index d99eebd..2572f84 100644 (file)
--- a/llvm/include/llvm/IR/IntrinsicsX86.td
+++ b/llvm/include/llvm/IR/IntrinsicsX86.td
@@ -1854,6 +1854,38 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_avx512_mask_ptestm_q_512 : GCCBuiltin<"__builtin_ia32_ptestmq512">,
         Intrinsic<[llvm_i8_ty], [llvm_v8i64_ty, llvm_v8i64_ty,
                   llvm_i8_ty], [IntrNoMem]>;
+
+  def int_x86_avx512_ptestm_b_128 : GCCBuiltin<"__builtin_ia32_ptestmb128">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i8_ty,
+                   llvm_v16i8_ty, llvm_i16_ty], [IntrNoMem]>;
+  def int_x86_avx512_ptestm_b_256 : GCCBuiltin<"__builtin_ia32_ptestmb256">,
+        Intrinsic<[llvm_i32_ty], [llvm_v32i8_ty,
+                   llvm_v32i8_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx512_ptestm_b_512 : GCCBuiltin<"__builtin_ia32_ptestmb512">,
+        Intrinsic<[llvm_i64_ty], [llvm_v64i8_ty,
+                   llvm_v64i8_ty, llvm_i64_ty], [IntrNoMem]>;
+  def int_x86_avx512_ptestm_d_128 : GCCBuiltin<"__builtin_ia32_ptestmd128">,
+        Intrinsic<[llvm_i8_ty], [llvm_v4i32_ty,
+                   llvm_v4i32_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_ptestm_d_256 : GCCBuiltin<"__builtin_ia32_ptestmd256">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i32_ty,
+                   llvm_v8i32_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_ptestm_q_128 : GCCBuiltin<"__builtin_ia32_ptestmq128">,
+        Intrinsic<[llvm_i8_ty], [llvm_v2i64_ty,
+                   llvm_v2i64_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_ptestm_q_256 : GCCBuiltin<"__builtin_ia32_ptestmq256">,
+        Intrinsic<[llvm_i8_ty], [llvm_v4i64_ty,
+                   llvm_v4i64_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_ptestm_w_128 : GCCBuiltin<"__builtin_ia32_ptestmw128">,
+        Intrinsic<[llvm_i8_ty], [llvm_v8i16_ty,
+                   llvm_v8i16_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_ptestm_w_256 : GCCBuiltin<"__builtin_ia32_ptestmw256">,
+        Intrinsic<[llvm_i16_ty], [llvm_v16i16_ty,
+                   llvm_v16i16_ty, llvm_i16_ty], [IntrNoMem]>;
+  def int_x86_avx512_ptestm_w_512 : GCCBuiltin<"__builtin_ia32_ptestmw512">,
+        Intrinsic<[llvm_i32_ty], [llvm_v32i16_ty,
+                   llvm_v32i16_ty, llvm_i32_ty], [IntrNoMem]>;
+
   def int_x86_avx512_mask_fpclass_pd_128 : 
          GCCBuiltin<"__builtin_ia32_fpclasspd128_mask">,
           Intrinsic<[llvm_i8_ty], [llvm_v2f64_ty, llvm_i32_ty, llvm_i8_ty],

diff --git a/llvm/lib/Target/X86/X86IntrinsicsInfo.h b/llvm/lib/Target/X86/X86IntrinsicsInfo.h
index 4d6f7c8..fd66ff0 100644 (file)
--- a/llvm/lib/Target/X86/X86IntrinsicsInfo.h
+++ b/llvm/lib/Target/X86/X86IntrinsicsInfo.h
@@ -2030,6 +2030,16 @@ static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(avx512_psad_bw_512, INTR_TYPE_2OP, X86ISD::PSADBW, 0),
   X86_INTRINSIC_DATA(avx512_psll_dq_512, INTR_TYPE_2OP_IMM8, X86ISD::VSHLDQ, 0),
   X86_INTRINSIC_DATA(avx512_psrl_dq_512, INTR_TYPE_2OP_IMM8, X86ISD::VSRLDQ, 0),
+  X86_INTRINSIC_DATA(avx512_ptestm_b_128, CMP_MASK, X86ISD::TESTM, 0),
+  X86_INTRINSIC_DATA(avx512_ptestm_b_256, CMP_MASK, X86ISD::TESTM, 0),
+  X86_INTRINSIC_DATA(avx512_ptestm_b_512, CMP_MASK, X86ISD::TESTM, 0),
+  X86_INTRINSIC_DATA(avx512_ptestm_d_128, CMP_MASK, X86ISD::TESTM, 0),
+  X86_INTRINSIC_DATA(avx512_ptestm_d_256, CMP_MASK, X86ISD::TESTM, 0),
+  X86_INTRINSIC_DATA(avx512_ptestm_q_128, CMP_MASK, X86ISD::TESTM, 0),
+  X86_INTRINSIC_DATA(avx512_ptestm_q_256, CMP_MASK, X86ISD::TESTM, 0),
+  X86_INTRINSIC_DATA(avx512_ptestm_w_128, CMP_MASK, X86ISD::TESTM, 0),
+  X86_INTRINSIC_DATA(avx512_ptestm_w_256, CMP_MASK, X86ISD::TESTM, 0),
+  X86_INTRINSIC_DATA(avx512_ptestm_w_512, CMP_MASK, X86ISD::TESTM, 0),
   X86_INTRINSIC_DATA(avx512_rcp14_pd_128, INTR_TYPE_1OP_MASK, X86ISD::FRCP, 0),
   X86_INTRINSIC_DATA(avx512_rcp14_pd_256, INTR_TYPE_1OP_MASK, X86ISD::FRCP, 0),
   X86_INTRINSIC_DATA(avx512_rcp14_pd_512, INTR_TYPE_1OP_MASK, X86ISD::FRCP, 0),

diff --git a/llvm/test/CodeGen/X86/avx512bw-intrinsics.ll b/llvm/test/CodeGen/X86/avx512bw-intrinsics.ll
index c0125f3..58219b3 100644 (file)
--- a/llvm/test/CodeGen/X86/avx512bw-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512bw-intrinsics.ll
@@ -3444,3 +3444,39 @@ define <64 x i8>@test_int_x86_avx512_mask_movu_b_512(<64 x i8> %x0, <64 x i8> %x
        %res2 = add <64 x i8> %res, %res1
        ret <64 x i8> %res2
 }
+
+declare i64 @llvm.x86.avx512.ptestm.b.512(<64 x i8>, <64 x i8>, i64)
+
+define i64@test_int_x86_avx512_ptestm_b_512(<64 x i8> %x0, <64 x i8> %x1, i64 %x2) {
+; CHECK-LABEL: test_int_x86_avx512_ptestm_b_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovq %rdi, %k1 ## encoding: [0xc4,0xe1,0xfb,0x92,0xcf]
+; CHECK-NEXT:    vptestmb %zmm1, %zmm0, %k0 {%k1} ## encoding: [0x62,0xf2,0x7d,0x49,0x26,0xc1]
+; CHECK-NEXT:    kmovq %k0, %rcx ## encoding: [0xc4,0xe1,0xfb,0x93,0xc8]
+; CHECK-NEXT:    vptestmb %zmm1, %zmm0, %k0 ## encoding: [0x62,0xf2,0x7d,0x48,0x26,0xc1]
+; CHECK-NEXT:    kmovq %k0, %rax ## encoding: [0xc4,0xe1,0xfb,0x93,0xc0]
+; CHECK-NEXT:    addq %rcx, %rax ## encoding: [0x48,0x01,0xc8]
+; CHECK-NEXT:    retq ## encoding: [0xc3]
+  %res = call i64 @llvm.x86.avx512.ptestm.b.512(<64 x i8> %x0, <64 x i8> %x1, i64 %x2)
+  %res1 = call i64 @llvm.x86.avx512.ptestm.b.512(<64 x i8> %x0, <64 x i8> %x1, i64-1)
+  %res2 = add i64 %res, %res1
+  ret i64 %res2
+}
+
+declare i32 @llvm.x86.avx512.ptestm.w.512(<32 x i16>, <32 x i16>, i32)
+
+define i32@test_int_x86_avx512_ptestm_w_512(<32 x i16> %x0, <32 x i16> %x1, i32 %x2) {
+; CHECK-LABEL: test_int_x86_avx512_ptestm_w_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
+; CHECK-NEXT:    vptestmw %zmm1, %zmm0, %k0 {%k1} ## encoding: [0x62,0xf2,0xfd,0x49,0x26,0xc1]
+; CHECK-NEXT:    kmovd %k0, %ecx ## encoding: [0xc5,0xfb,0x93,0xc8]
+; CHECK-NEXT:    vptestmw %zmm1, %zmm0, %k0 ## encoding: [0x62,0xf2,0xfd,0x48,0x26,0xc1]
+; CHECK-NEXT:    kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
+; CHECK-NEXT:    addl %ecx, %eax ## encoding: [0x01,0xc8]
+; CHECK-NEXT:    retq ## encoding: [0xc3]
+  %res = call i32 @llvm.x86.avx512.ptestm.w.512(<32 x i16> %x0, <32 x i16> %x1, i32 %x2)
+  %res1 = call i32 @llvm.x86.avx512.ptestm.w.512(<32 x i16> %x0, <32 x i16> %x1, i32-1)
+  %res2 = add i32 %res, %res1
+  ret i32 %res2
+}

diff --git a/llvm/test/CodeGen/X86/avx512bwvl-intrinsics.ll b/llvm/test/CodeGen/X86/avx512bwvl-intrinsics.ll
index 4c624e5..4ec0e2f 100644 (file)
--- a/llvm/test/CodeGen/X86/avx512bwvl-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512bwvl-intrinsics.ll
@@ -5299,3 +5299,75 @@ define <32 x i8>@test_int_x86_avx512_mask_movu_b_256(<32 x i8> %x0, <32 x i8> %x
        %res2 = add <32 x i8> %res, %res1
        ret <32 x i8> %res2
 }
+
+declare i16 @llvm.x86.avx512.ptestm.b.128(<16 x i8>, <16 x i8>, i16)
+
+define i16@test_int_x86_avx512_ptestm_b_128(<16 x i8> %x0, <16 x i8> %x1, i16 %x2) {
+; CHECK-LABEL: test_int_x86_avx512_ptestm_b_128:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
+; CHECK-NEXT:    vptestmb %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x26,0xc1]
+; CHECK-NEXT:    kmovw %k0, %ecx ## encoding: [0xc5,0xf8,0x93,0xc8]
+; CHECK-NEXT:    vptestmb %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf2,0x7d,0x08,0x26,0xc1]
+; CHECK-NEXT:    kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
+; CHECK-NEXT:    addl %ecx, %eax ## encoding: [0x01,0xc8]
+; CHECK-NEXT:    retq ## encoding: [0xc3]
+  %res = call i16 @llvm.x86.avx512.ptestm.b.128(<16 x i8> %x0, <16 x i8> %x1, i16 %x2)
+  %res1 = call i16 @llvm.x86.avx512.ptestm.b.128(<16 x i8> %x0, <16 x i8> %x1, i16-1)
+  %res2 = add i16 %res, %res1
+  ret i16 %res2
+}
+
+declare i32 @llvm.x86.avx512.ptestm.b.256(<32 x i8>, <32 x i8>, i32)
+
+define i32@test_int_x86_avx512_ptestm_b_256(<32 x i8> %x0, <32 x i8> %x1, i32 %x2) {
+; CHECK-LABEL: test_int_x86_avx512_ptestm_b_256:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
+; CHECK-NEXT:    vptestmb %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x26,0xc1]
+; CHECK-NEXT:    kmovd %k0, %ecx ## encoding: [0xc5,0xfb,0x93,0xc8]
+; CHECK-NEXT:    vptestmb %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf2,0x7d,0x28,0x26,0xc1]
+; CHECK-NEXT:    kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
+; CHECK-NEXT:    addl %ecx, %eax ## encoding: [0x01,0xc8]
+; CHECK-NEXT:    retq ## encoding: [0xc3]
+  %res = call i32 @llvm.x86.avx512.ptestm.b.256(<32 x i8> %x0, <32 x i8> %x1, i32 %x2)
+  %res1 = call i32 @llvm.x86.avx512.ptestm.b.256(<32 x i8> %x0, <32 x i8> %x1, i32-1)
+  %res2 = add i32 %res, %res1
+  ret i32 %res2
+}
+
+declare i8 @llvm.x86.avx512.ptestm.w.128(<8 x i16>, <8 x i16>, i8)
+
+define i8@test_int_x86_avx512_ptestm_w_128(<8 x i16> %x0, <8 x i16> %x1, i8 %x2) {
+; CHECK-LABEL: test_int_x86_avx512_ptestm_w_128:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
+; CHECK-NEXT:    vptestmw %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf2,0xfd,0x09,0x26,0xc1]
+; CHECK-NEXT:    kmovw %k0, %ecx ## encoding: [0xc5,0xf8,0x93,0xc8]
+; CHECK-NEXT:    vptestmw %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf2,0xfd,0x08,0x26,0xc1]
+; CHECK-NEXT:    kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
+; CHECK-NEXT:    addb %cl, %al ## encoding: [0x00,0xc8]
+; CHECK-NEXT:    retq ## encoding: [0xc3]
+  %res = call i8 @llvm.x86.avx512.ptestm.w.128(<8 x i16> %x0, <8 x i16> %x1, i8 %x2)
+  %res1 = call i8 @llvm.x86.avx512.ptestm.w.128(<8 x i16> %x0, <8 x i16> %x1, i8-1)
+  %res2 = add i8 %res, %res1
+  ret i8 %res2
+}
+
+declare i16 @llvm.x86.avx512.ptestm.w.256(<16 x i16>, <16 x i16>, i16)
+
+define i16@test_int_x86_avx512_ptestm_w_256(<16 x i16> %x0, <16 x i16> %x1, i16 %x2) {
+; CHECK-LABEL: test_int_x86_avx512_ptestm_w_256:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
+; CHECK-NEXT:    vptestmw %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf2,0xfd,0x29,0x26,0xc1]
+; CHECK-NEXT:    kmovw %k0, %ecx ## encoding: [0xc5,0xf8,0x93,0xc8]
+; CHECK-NEXT:    vptestmw %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf2,0xfd,0x28,0x26,0xc1]
+; CHECK-NEXT:    kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
+; CHECK-NEXT:    addl %ecx, %eax ## encoding: [0x01,0xc8]
+; CHECK-NEXT:    retq ## encoding: [0xc3]
+  %res = call i16 @llvm.x86.avx512.ptestm.w.256(<16 x i16> %x0, <16 x i16> %x1, i16 %x2)
+  %res1 = call i16 @llvm.x86.avx512.ptestm.w.256(<16 x i16> %x0, <16 x i16> %x1, i16-1)
+  %res2 = add i16 %res, %res1
+  ret i16 %res2
+}

diff --git a/llvm/test/CodeGen/X86/avx512vl-intrinsics.ll b/llvm/test/CodeGen/X86/avx512vl-intrinsics.ll
index 3955f31..b814715 100644 (file)
--- a/llvm/test/CodeGen/X86/avx512vl-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512vl-intrinsics.ll
@@ -8061,3 +8061,75 @@ define <8 x float>@test_int_x86_avx512_maskz_fixupimm_ps_256(<8 x float> %x0, <8
   %res4 = fadd <8 x float> %res3, %res2
   ret <8 x float> %res4
 }
+
+declare i8 @llvm.x86.avx512.ptestm.d.128(<4 x i32>, <4 x i32>,i8)
+
+define i8@test_int_x86_avx512_ptestm_d_128(<4 x i32> %x0, <4 x i32> %x1, i8 %x2) {
+; CHECK-LABEL: test_int_x86_avx512_ptestm_d_128:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
+; CHECK-NEXT:    vptestmd %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x27,0xc1]
+; CHECK-NEXT:    kmovw %k0, %ecx ## encoding: [0xc5,0xf8,0x93,0xc8]
+; CHECK-NEXT:    vptestmd %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf2,0x7d,0x08,0x27,0xc1]
+; CHECK-NEXT:    kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
+; CHECK-NEXT:    addb %cl, %al ## encoding: [0x00,0xc8]
+; CHECK-NEXT:    retq ## encoding: [0xc3]
+  %res = call i8 @llvm.x86.avx512.ptestm.d.128(<4 x i32> %x0, <4 x i32> %x1, i8 %x2)
+  %res1 = call i8 @llvm.x86.avx512.ptestm.d.128(<4 x i32> %x0, <4 x i32> %x1, i8-1)
+  %res2 = add i8 %res, %res1
+  ret i8 %res2
+}
+
+declare i8 @llvm.x86.avx512.ptestm.d.256(<8 x i32>, <8 x i32>, i8)
+
+define i8@test_int_x86_avx512_ptestm_d_256(<8 x i32> %x0, <8 x i32> %x1, i8 %x2) {
+; CHECK-LABEL: test_int_x86_avx512_ptestm_d_256:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
+; CHECK-NEXT:    vptestmd %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x27,0xc1]
+; CHECK-NEXT:    kmovw %k0, %ecx ## encoding: [0xc5,0xf8,0x93,0xc8]
+; CHECK-NEXT:    vptestmd %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf2,0x7d,0x28,0x27,0xc1]
+; CHECK-NEXT:    kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
+; CHECK-NEXT:    addb %cl, %al ## encoding: [0x00,0xc8]
+; CHECK-NEXT:    retq ## encoding: [0xc3]
+  %res = call i8 @llvm.x86.avx512.ptestm.d.256(<8 x i32> %x0, <8 x i32> %x1, i8 %x2)
+  %res1 = call i8 @llvm.x86.avx512.ptestm.d.256(<8 x i32> %x0, <8 x i32> %x1, i8-1)
+  %res2 = add i8 %res, %res1
+  ret i8 %res2
+}
+
+declare i8 @llvm.x86.avx512.ptestm.q.128(<2 x i64>, <2 x i64>, i8)
+
+define i8@test_int_x86_avx512_ptestm_q_128(<2 x i64> %x0, <2 x i64> %x1, i8 %x2) {
+; CHECK-LABEL: test_int_x86_avx512_ptestm_q_128:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
+; CHECK-NEXT:    vptestmq %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf2,0xfd,0x09,0x27,0xc1]
+; CHECK-NEXT:    kmovw %k0, %ecx ## encoding: [0xc5,0xf8,0x93,0xc8]
+; CHECK-NEXT:    vptestmq %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf2,0xfd,0x08,0x27,0xc1]
+; CHECK-NEXT:    kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
+; CHECK-NEXT:    addb %cl, %al ## encoding: [0x00,0xc8]
+; CHECK-NEXT:    retq ## encoding: [0xc3]
+  %res = call i8 @llvm.x86.avx512.ptestm.q.128(<2 x i64> %x0, <2 x i64> %x1, i8 %x2)
+  %res1 = call i8 @llvm.x86.avx512.ptestm.q.128(<2 x i64> %x0, <2 x i64> %x1, i8-1)
+  %res2 = add i8 %res, %res1
+  ret i8 %res2
+}
+
+declare i8 @llvm.x86.avx512.ptestm.q.256(<4 x i64>, <4 x i64>, i8)
+
+define i8@test_int_x86_avx512_ptestm_q_256(<4 x i64> %x0, <4 x i64> %x1, i8 %x2) {
+; CHECK-LABEL: test_int_x86_avx512_ptestm_q_256:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
+; CHECK-NEXT:    vptestmq %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf2,0xfd,0x29,0x27,0xc1]
+; CHECK-NEXT:    kmovw %k0, %ecx ## encoding: [0xc5,0xf8,0x93,0xc8]
+; CHECK-NEXT:    vptestmq %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf2,0xfd,0x28,0x27,0xc1]
+; CHECK-NEXT:    kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
+; CHECK-NEXT:    addb %cl, %al ## encoding: [0x00,0xc8]
+; CHECK-NEXT:    retq ## encoding: [0xc3]
+  %res = call i8 @llvm.x86.avx512.ptestm.q.256(<4 x i64> %x0, <4 x i64> %x1, i8 %x2)
+  %res1 = call i8 @llvm.x86.avx512.ptestm.q.256(<4 x i64> %x0, <4 x i64> %x1, i8-1)
+  %res2 = add i8 %res, %res1
+  ret i8 %res2
+}