From: Uriel Korach <uriel.korach@intel.com>
Date: Mon, 6 Nov 2017 09:22:38 +0000 (+0000)
Subject: [X86][AVX512] Improve lowering of AVX512 test intrinsics
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=bb86686a8b6f9433954a535f99f23e11b78ca348;p=platform%2Fupstream%2Fllvm.git

[X86][AVX512] Improve lowering of AVX512 test intrinsics

Added TESTM and TESTNM to the list of instructions that already zeroing unused upper bits
and does not need the redundant shift left and shift right instructions afterwards.
Added a pattern for TESTM and TESTNM in iselLowering, so now icmp(neq,and(X,Y), 0) goes folds into TESTM
and icmp(eq,and(X,Y), 0) goes folds into TESTNM
This commit is a preparation for lowering the test and testn X86 intrinsics to IR.

Differential Revision: https://reviews.llvm.org/D38732

llvm-svn: 317465
---

diff --git a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
index 01d6d0b..0cbf760 100644
--- a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -449,10 +449,10 @@ namespace {
 // Returns true if this masked compare can be implemented legally with this
 // type.
 static bool isLegalMaskCompare(SDNode *N, const X86Subtarget *Subtarget) {
-  if (N->getOpcode() == X86ISD::PCMPEQM ||
-      N->getOpcode() == X86ISD::PCMPGTM ||
-      N->getOpcode() == X86ISD::CMPM ||
-      N->getOpcode() == X86ISD::CMPMU) {
+  unsigned Opcode = N->getOpcode();
+  if (Opcode == X86ISD::PCMPEQM || Opcode == X86ISD::PCMPGTM ||
+      Opcode == X86ISD::CMPM || Opcode == X86ISD::TESTM ||
+      Opcode == X86ISD::TESTNM || Opcode == X86ISD::CMPMU) {
     // We can get 256-bit 8 element types here without VLX being enabled. When
     // this happens we will use 512-bit operations and the mask will not be
     // zero extended.
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 7b88887..85fee29 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -4998,6 +4998,8 @@ static bool isMaskedZeroUpperBitsvXi1(unsigned int Opcode) {
   switch (Opcode) {
   default:
     return false;
+  case X86ISD::TESTM:
+  case X86ISD::TESTNM:
   case X86ISD::PCMPEQM:
   case X86ISD::PCMPGTM:
   case X86ISD::CMPM:
@@ -17469,6 +17471,20 @@ static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
 
   if (Swap)
     std::swap(Op0, Op1);
+
+  //  See if it is the case of CMP(EQ|NEQ,AND(A,B),ZERO) and change it to TESTM|NM.
+  if ((!Opc && SSECC == 4) || Opc == X86ISD::PCMPEQM) {
+    SDValue A = peekThroughBitcasts(Op0);
+    if ((A.getOpcode() == ISD::AND || A.getOpcode() == X86ISD::FAND) &&
+        ISD::isBuildVectorAllZeros(Op1.getNode())) {
+      MVT VT0 = Op0.getSimpleValueType();
+      SDValue RHS = DAG.getBitcast(VT0, A.getOperand(0));
+      SDValue LHS = DAG.getBitcast(VT0, A.getOperand(1));
+      return DAG.getNode(Opc == X86ISD::PCMPEQM ? X86ISD::TESTNM : X86ISD::TESTM,
+                         dl, VT, RHS, LHS);
+    }
+  }
+
   if (Opc)
     return DAG.getNode(Opc, dl, VT, Op0, Op1);
   Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
diff --git a/llvm/test/CodeGen/X86/avx512-skx-insert-subvec.ll b/llvm/test/CodeGen/X86/avx512-skx-insert-subvec.ll
index 23d6645..ff25c00 100644
--- a/llvm/test/CodeGen/X86/avx512-skx-insert-subvec.ll
+++ b/llvm/test/CodeGen/X86/avx512-skx-insert-subvec.ll
@@ -46,8 +46,6 @@ define <8 x i1> @test3(<4 x i1> %a) {
 ; CHECK:       # BB#0:
 ; CHECK-NEXT:    vpslld $31, %xmm0, %xmm0
 ; CHECK-NEXT:    vptestmd %xmm0, %xmm0, %k0
-; CHECK-NEXT:    kshiftlb $4, %k0, %k0
-; CHECK-NEXT:    kshiftrb $4, %k0, %k0
 ; CHECK-NEXT:    vpmovm2w %k0, %xmm0
 ; CHECK-NEXT:    retq
 
diff --git a/llvm/test/CodeGen/X86/avx512bw-vec-test-testn.ll b/llvm/test/CodeGen/X86/avx512bw-vec-test-testn.ll
index 6dd6440..82d0b88 100644
--- a/llvm/test/CodeGen/X86/avx512bw-vec-test-testn.ll
+++ b/llvm/test/CodeGen/X86/avx512bw-vec-test-testn.ll
@@ -5,9 +5,7 @@
 define zeroext i32 @TEST_mm512_test_epi16_mask(<8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_test_epi16_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpneqw %zmm1, %zmm0, %k0
+; CHECK-NEXT:    vptestmw %zmm0, %zmm1, %k0
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
@@ -24,9 +22,7 @@ entry:
 define zeroext i64 @TEST_mm512_test_epi8_mask(<8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_test_epi8_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpneqb %zmm1, %zmm0, %k0
+; CHECK-NEXT:    vptestmb %zmm0, %zmm1, %k0
 ; CHECK-NEXT:    kmovq %k0, %rax
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
@@ -42,10 +38,8 @@ entry:
 define zeroext i32 @TEST_mm512_mask_test_epi16_mask(i32 %__U, <8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_mask_test_epi16_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovd %edi, %k1
-; CHECK-NEXT:    vpcmpneqw %zmm1, %zmm0, %k0 {%k1}
+; CHECK-NEXT:    vptestmw %zmm0, %zmm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
@@ -63,10 +57,8 @@ entry:
 define zeroext i64 @TEST_mm512_mask_test_epi8_mask(i64 %__U, <8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_mask_test_epi8_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovq %rdi, %k1
-; CHECK-NEXT:    vpcmpneqb %zmm1, %zmm0, %k0 {%k1}
+; CHECK-NEXT:    vptestmb %zmm0, %zmm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovq %k0, %rax
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
@@ -84,9 +76,7 @@ entry:
 define zeroext i32 @TEST_mm512_testn_epi16_mask(<8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_testn_epi16_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpeqw %zmm1, %zmm0, %k0
+; CHECK-NEXT:    vptestnmw %zmm0, %zmm1, %k0
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
@@ -103,9 +93,7 @@ entry:
 define zeroext i64 @TEST_mm512_testn_epi8_mask(<8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_testn_epi8_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpeqb %zmm1, %zmm0, %k0
+; CHECK-NEXT:    vptestnmb %zmm0, %zmm1, %k0
 ; CHECK-NEXT:    kmovq %k0, %rax
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
@@ -121,10 +109,8 @@ entry:
 define zeroext i32 @TEST_mm512_mask_testn_epi16_mask(i32 %__U, <8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_mask_testn_epi16_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovd %edi, %k1
-; CHECK-NEXT:    vpcmpeqw %zmm1, %zmm0, %k0 {%k1}
+; CHECK-NEXT:    vptestnmw %zmm0, %zmm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
@@ -142,10 +128,8 @@ entry:
 define zeroext i64 @TEST_mm512_mask_testn_epi8_mask(i64 %__U, <8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_mask_testn_epi8_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovq %rdi, %k1
-; CHECK-NEXT:    vpcmpeqb %zmm1, %zmm0, %k0 {%k1}
+; CHECK-NEXT:    vptestnmb %zmm0, %zmm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovq %k0, %rax
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
diff --git a/llvm/test/CodeGen/X86/avx512bwvl-vec-test-testn.ll b/llvm/test/CodeGen/X86/avx512bwvl-vec-test-testn.ll
index f67ceb2..44075de 100644
--- a/llvm/test/CodeGen/X86/avx512bwvl-vec-test-testn.ll
+++ b/llvm/test/CodeGen/X86/avx512bwvl-vec-test-testn.ll
@@ -5,9 +5,7 @@
 define zeroext i16 @TEST_mm_test_epi8_mask(<2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm_test_epi8_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpneqb %xmm1, %xmm0, %k0
+; CHECK-NEXT:    vptestmb %xmm0, %xmm1, %k0
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; CHECK-NEXT:    retq
@@ -23,10 +21,8 @@ entry:
 define zeroext i16 @TEST_mm_mask_test_epi8_mask(i16 zeroext %__U, <2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm_mask_test_epi8_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovd %edi, %k1
-; CHECK-NEXT:    vpcmpneqb %xmm1, %xmm0, %k0 {%k1}
+; CHECK-NEXT:    vptestmb %xmm0, %xmm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; CHECK-NEXT:    retq
@@ -44,9 +40,7 @@ entry:
 define zeroext i8 @TEST_mm_test_epi16_mask(<2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm_test_epi16_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpneqw %xmm1, %xmm0, %k0
+; CHECK-NEXT:    vptestmw %xmm0, %xmm1, %k0
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; CHECK-NEXT:    retq
@@ -62,10 +56,8 @@ entry:
 define zeroext i8 @TEST_mm_mask_test_epi16_mask(i8 zeroext %__U, <2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm_mask_test_epi16_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovd %edi, %k1
-; CHECK-NEXT:    vpcmpneqw %xmm1, %xmm0, %k0 {%k1}
+; CHECK-NEXT:    vptestmw %xmm0, %xmm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; CHECK-NEXT:    retq
@@ -83,9 +75,7 @@ entry:
 define zeroext i16 @TEST_mm_testn_epi8_mask(<2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm_testn_epi8_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpeqb %xmm1, %xmm0, %k0
+; CHECK-NEXT:    vptestnmb %xmm0, %xmm1, %k0
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; CHECK-NEXT:    retq
@@ -101,10 +91,8 @@ entry:
 define zeroext i16 @TEST_mm_mask_testn_epi8_mask(i16 zeroext %__U, <2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm_mask_testn_epi8_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovd %edi, %k1
-; CHECK-NEXT:    vpcmpeqb %xmm1, %xmm0, %k0 {%k1}
+; CHECK-NEXT:    vptestnmb %xmm0, %xmm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; CHECK-NEXT:    retq
@@ -122,9 +110,7 @@ entry:
 define zeroext i8 @TEST_mm_testn_epi16_mask(<2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm_testn_epi16_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpeqw %xmm1, %xmm0, %k0
+; CHECK-NEXT:    vptestnmw %xmm0, %xmm1, %k0
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; CHECK-NEXT:    retq
@@ -140,10 +126,8 @@ entry:
 define zeroext i8 @TEST_mm_mask_testn_epi16_mask(i8 zeroext %__U, <2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm_mask_testn_epi16_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovd %edi, %k1
-; CHECK-NEXT:    vpcmpeqw %xmm1, %xmm0, %k0 {%k1}
+; CHECK-NEXT:    vptestnmw %xmm0, %xmm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; CHECK-NEXT:    retq
@@ -161,9 +145,7 @@ entry:
 define i32 @TEST_mm256_test_epi8_mask(<4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm256_test_epi8_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpneqb %ymm1, %ymm0, %k0
+; CHECK-NEXT:    vptestmb %ymm0, %ymm1, %k0
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
@@ -179,10 +161,8 @@ entry:
 define i32 @TEST_mm256_mask_test_epi8_mask(i32 %__U, <4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm256_mask_test_epi8_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovd %edi, %k1
-; CHECK-NEXT:    vpcmpneqb %ymm1, %ymm0, %k0 {%k1}
+; CHECK-NEXT:    vptestmb %ymm0, %ymm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
@@ -200,9 +180,7 @@ entry:
 define zeroext i16 @TEST_mm256_test_epi16_mask(<4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm256_test_epi16_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpneqw %ymm1, %ymm0, %k0
+; CHECK-NEXT:    vptestmw %ymm0, %ymm1, %k0
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; CHECK-NEXT:    vzeroupper
@@ -219,10 +197,8 @@ entry:
 define zeroext i16 @TEST_mm256_mask_test_epi16_mask(i16 zeroext %__U, <4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm256_mask_test_epi16_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovd %edi, %k1
-; CHECK-NEXT:    vpcmpneqw %ymm1, %ymm0, %k0 {%k1}
+; CHECK-NEXT:    vptestmw %ymm0, %ymm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; CHECK-NEXT:    vzeroupper
@@ -241,9 +217,7 @@ entry:
 define i32 @TEST_mm256_testn_epi8_mask(<4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm256_testn_epi8_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpeqb %ymm1, %ymm0, %k0
+; CHECK-NEXT:    vptestnmb %ymm0, %ymm1, %k0
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
@@ -259,10 +233,8 @@ entry:
 define i32 @TEST_mm256_mask_testn_epi8_mask(i32 %__U, <4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm256_mask_testn_epi8_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovd %edi, %k1
-; CHECK-NEXT:    vpcmpeqb %ymm1, %ymm0, %k0 {%k1}
+; CHECK-NEXT:    vptestnmb %ymm0, %ymm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
@@ -280,9 +252,7 @@ entry:
 define zeroext i16 @TEST_mm256_testn_epi16_mask(<4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm256_testn_epi16_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpeqw %ymm1, %ymm0, %k0
+; CHECK-NEXT:    vptestnmw %ymm0, %ymm1, %k0
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; CHECK-NEXT:    vzeroupper
@@ -299,10 +269,8 @@ entry:
 define zeroext i16 @TEST_mm256_mask_testn_epi16_mask(i16 zeroext %__U, <4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm256_mask_testn_epi16_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovd %edi, %k1
-; CHECK-NEXT:    vpcmpeqw %ymm1, %ymm0, %k0 {%k1}
+; CHECK-NEXT:    vptestnmw %ymm0, %ymm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovd %k0, %eax
 ; CHECK-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; CHECK-NEXT:    vzeroupper
diff --git a/llvm/test/CodeGen/X86/avx512f-vec-test-testn.ll b/llvm/test/CodeGen/X86/avx512f-vec-test-testn.ll
index c9c0c22..e9cdacc 100644
--- a/llvm/test/CodeGen/X86/avx512f-vec-test-testn.ll
+++ b/llvm/test/CodeGen/X86/avx512f-vec-test-testn.ll
@@ -5,9 +5,7 @@
 define zeroext i8 @TEST_mm512_test_epi64_mask(<8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_test_epi64_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpneqq %zmm1, %zmm0, %k0
+; CHECK-NEXT:    vptestmq %zmm0, %zmm1, %k0
 ; CHECK-NEXT:    kmovw %k0, %eax
 ; CHECK-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; CHECK-NEXT:    vzeroupper
@@ -23,9 +21,7 @@ entry:
 define zeroext i16 @TEST_mm512_test_epi32_mask(<8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_test_epi32_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpneqd %zmm1, %zmm0, %k0
+; CHECK-NEXT:    vptestmd %zmm0, %zmm1, %k0
 ; CHECK-NEXT:    kmovw %k0, %eax
 ; CHECK-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; CHECK-NEXT:    vzeroupper
@@ -42,10 +38,8 @@ entry:
 define zeroext i8 @TEST_mm512_mask_test_epi64_mask(i8 %__U, <8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_mask_test_epi64_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovw %edi, %k1
-; CHECK-NEXT:    vpcmpneqq %zmm1, %zmm0, %k0 {%k1}
+; CHECK-NEXT:    vptestmq %zmm0, %zmm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovw %k0, %eax
 ; CHECK-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; CHECK-NEXT:    vzeroupper
@@ -63,10 +57,8 @@ entry:
 define zeroext i16 @TEST_mm512_mask_test_epi32_mask(i16 %__U, <8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_mask_test_epi32_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovw %edi, %k1
-; CHECK-NEXT:    vpcmpneqd %zmm1, %zmm0, %k0 {%k1}
+; CHECK-NEXT:    vptestmd %zmm0, %zmm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovw %k0, %eax
 ; CHECK-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; CHECK-NEXT:    vzeroupper
@@ -85,9 +77,7 @@ entry:
 define zeroext i8 @TEST_mm512_testn_epi64_mask(<8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_testn_epi64_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0
+; CHECK-NEXT:    vptestnmq %zmm0, %zmm1, %k0
 ; CHECK-NEXT:    kmovw %k0, %eax
 ; CHECK-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; CHECK-NEXT:    vzeroupper
@@ -103,9 +93,7 @@ entry:
 define zeroext i16 @TEST_mm512_testn_epi32_mask(<8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_testn_epi32_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
+; CHECK-NEXT:    vptestnmd %zmm0, %zmm1, %k0
 ; CHECK-NEXT:    kmovw %k0, %eax
 ; CHECK-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; CHECK-NEXT:    vzeroupper
@@ -122,10 +110,8 @@ entry:
 define zeroext i8 @TEST_mm512_mask_testn_epi64_mask(i8 %__U, <8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_mask_testn_epi64_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovw %edi, %k1
-; CHECK-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
+; CHECK-NEXT:    vptestnmq %zmm0, %zmm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovw %k0, %eax
 ; CHECK-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; CHECK-NEXT:    vzeroupper
@@ -143,10 +129,8 @@ entry:
 define zeroext i16 @TEST_mm512_mask_testn_epi32_mask(i16 %__U, <8 x i64> %__A, <8 x i64> %__B) local_unnamed_addr #0 {
 ; CHECK-LABEL: TEST_mm512_mask_testn_epi32_mask:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    vpandq %zmm0, %zmm1, %zmm0
-; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovw %edi, %k1
-; CHECK-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
+; CHECK-NEXT:    vptestnmd %zmm0, %zmm1, %k0 {%k1}
 ; CHECK-NEXT:    kmovw %k0, %eax
 ; CHECK-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; CHECK-NEXT:    vzeroupper
diff --git a/llvm/test/CodeGen/X86/avx512vl-vec-masked-cmp.ll b/llvm/test/CodeGen/X86/avx512vl-vec-masked-cmp.ll
index 5ee06fd..26a7d83 100644
--- a/llvm/test/CodeGen/X86/avx512vl-vec-masked-cmp.ll
+++ b/llvm/test/CodeGen/X86/avx512vl-vec-masked-cmp.ll
@@ -1217,8 +1217,6 @@ define zeroext i16 @test_vpcmpeqw_v8i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -1246,8 +1244,6 @@ define zeroext i16 @test_vpcmpeqw_v8i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>*
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -1278,8 +1274,6 @@ define zeroext i16 @test_masked_vpcmpeqw_v8i1_v16i1_mask(i8 zeroext %__u, <2 x i
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -1311,8 +1305,6 @@ define zeroext i16 @test_masked_vpcmpeqw_v8i1_v16i1_mask_mem(i8 zeroext %__u, <2
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -13586,8 +13578,6 @@ define zeroext i16 @test_vpcmpsgtw_v8i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -13615,8 +13605,6 @@ define zeroext i16 @test_vpcmpsgtw_v8i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -13647,8 +13635,6 @@ define zeroext i16 @test_masked_vpcmpsgtw_v8i1_v16i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -13680,8 +13666,6 @@ define zeroext i16 @test_masked_vpcmpsgtw_v8i1_v16i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -25987,8 +25971,6 @@ define zeroext i16 @test_vpcmpsgew_v8i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -26019,8 +26001,6 @@ define zeroext i16 @test_vpcmpsgew_v8i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -26053,8 +26033,6 @@ define zeroext i16 @test_masked_vpcmpsgew_v8i1_v16i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -26089,8 +26067,6 @@ define zeroext i16 @test_masked_vpcmpsgew_v8i1_v16i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -38587,8 +38563,6 @@ define zeroext i16 @test_vpcmpultw_v8i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -38619,8 +38593,6 @@ define zeroext i16 @test_vpcmpultw_v8i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -38654,8 +38626,6 @@ define zeroext i16 @test_masked_vpcmpultw_v8i1_v16i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
@@ -38690,8 +38660,6 @@ define zeroext i16 @test_masked_vpcmpultw_v8i1_v16i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kunpckbw %k0, %k1, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; NoVLX-NEXT:    vzeroupper
diff --git a/llvm/test/CodeGen/X86/avx512vl-vec-test-testn.ll b/llvm/test/CodeGen/X86/avx512vl-vec-test-testn.ll
index f1919cb..32de025 100644
--- a/llvm/test/CodeGen/X86/avx512vl-vec-test-testn.ll
+++ b/llvm/test/CodeGen/X86/avx512vl-vec-test-testn.ll
@@ -6,18 +6,14 @@
 define zeroext i8 @TEST_mm_test_epi64_mask(<2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm_test_epi64_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; X86_64-NEXT:    vpcmpneqq %xmm1, %xmm0, %k0
+; X86_64-NEXT:    vptestmq %xmm0, %xmm1, %k0
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    retq
 ;
 ; I386-LABEL: TEST_mm_test_epi64_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; I386-NEXT:    vpcmpneqq %xmm1, %xmm0, %k0
+; I386-NEXT:    vptestmq %xmm0, %xmm1, %k0
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    retl
@@ -33,18 +29,14 @@ entry:
 define zeroext i8 @TEST_mm_test_epi32_mask(<2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm_test_epi32_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; X86_64-NEXT:    vpcmpneqd %xmm1, %xmm0, %k0
+; X86_64-NEXT:    vptestmd %xmm0, %xmm1, %k0
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    retq
 ;
 ; I386-LABEL: TEST_mm_test_epi32_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; I386-NEXT:    vpcmpneqd %xmm1, %xmm0, %k0
+; I386-NEXT:    vptestmd %xmm0, %xmm1, %k0
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    retl
@@ -61,9 +53,7 @@ entry:
 define zeroext i8 @TEST_mm256_test_epi64_mask(<4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm256_test_epi64_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; X86_64-NEXT:    vpcmpneqq %ymm1, %ymm0, %k0
+; X86_64-NEXT:    vptestmq %ymm0, %ymm1, %k0
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    vzeroupper
@@ -71,9 +61,7 @@ define zeroext i8 @TEST_mm256_test_epi64_mask(<4 x i64> %__A, <4 x i64> %__B) lo
 ;
 ; I386-LABEL: TEST_mm256_test_epi64_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; I386-NEXT:    vpcmpneqq %ymm1, %ymm0, %k0
+; I386-NEXT:    vptestmq %ymm0, %ymm1, %k0
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    vzeroupper
@@ -90,9 +78,7 @@ entry:
 define zeroext i8 @TEST_mm256_test_epi32_mask(<4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm256_test_epi32_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; X86_64-NEXT:    vpcmpneqd %ymm1, %ymm0, %k0
+; X86_64-NEXT:    vptestmd %ymm0, %ymm1, %k0
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    vzeroupper
@@ -100,9 +86,7 @@ define zeroext i8 @TEST_mm256_test_epi32_mask(<4 x i64> %__A, <4 x i64> %__B) lo
 ;
 ; I386-LABEL: TEST_mm256_test_epi32_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; I386-NEXT:    vpcmpneqd %ymm1, %ymm0, %k0
+; I386-NEXT:    vptestmd %ymm0, %ymm1, %k0
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    vzeroupper
@@ -119,21 +103,17 @@ entry:
 define zeroext i8 @TEST_mm_mask_test_epi64_mask(i8 %__U, <2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm_mask_test_epi64_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; X86_64-NEXT:    kmovw %edi, %k1
-; X86_64-NEXT:    vpcmpneqq %xmm1, %xmm0, %k0 {%k1}
+; X86_64-NEXT:    vptestmq %xmm0, %xmm1, %k0 {%k1}
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    retq
 ;
 ; I386-LABEL: TEST_mm_mask_test_epi64_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; I386-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
 ; I386-NEXT:    kmovw %eax, %k1
-; I386-NEXT:    vpcmpneqq %xmm1, %xmm0, %k0 {%k1}
+; I386-NEXT:    vptestmq %xmm0, %xmm1, %k0 {%k1}
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    retl
@@ -152,21 +132,17 @@ entry:
 define zeroext i8 @TEST_mm_mask_test_epi32_mask(i8 %__U, <2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm_mask_test_epi32_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; X86_64-NEXT:    kmovw %edi, %k1
-; X86_64-NEXT:    vpcmpneqd %xmm1, %xmm0, %k0 {%k1}
+; X86_64-NEXT:    vptestmd %xmm0, %xmm1, %k0 {%k1}
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    retq
 ;
 ; I386-LABEL: TEST_mm_mask_test_epi32_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; I386-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
 ; I386-NEXT:    kmovw %eax, %k1
-; I386-NEXT:    vpcmpneqd %xmm1, %xmm0, %k0 {%k1}
+; I386-NEXT:    vptestmd %xmm0, %xmm1, %k0 {%k1}
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    retl
@@ -187,10 +163,8 @@ entry:
 define zeroext i8 @TEST_mm256_mask_test_epi64_mask(i8 %__U, <4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm256_mask_test_epi64_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; X86_64-NEXT:    kmovw %edi, %k1
-; X86_64-NEXT:    vpcmpneqq %ymm1, %ymm0, %k0 {%k1}
+; X86_64-NEXT:    vptestmq %ymm0, %ymm1, %k0 {%k1}
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    vzeroupper
@@ -198,11 +172,9 @@ define zeroext i8 @TEST_mm256_mask_test_epi64_mask(i8 %__U, <4 x i64> %__A, <4 x
 ;
 ; I386-LABEL: TEST_mm256_mask_test_epi64_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; I386-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
 ; I386-NEXT:    kmovw %eax, %k1
-; I386-NEXT:    vpcmpneqq %ymm1, %ymm0, %k0 {%k1}
+; I386-NEXT:    vptestmq %ymm0, %ymm1, %k0 {%k1}
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    vzeroupper
@@ -222,10 +194,8 @@ entry:
 define zeroext i8 @TEST_mm256_mask_test_epi32_mask(i8 %__U, <4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm256_mask_test_epi32_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; X86_64-NEXT:    kmovw %edi, %k1
-; X86_64-NEXT:    vpcmpneqd %ymm1, %ymm0, %k0 {%k1}
+; X86_64-NEXT:    vptestmd %ymm0, %ymm1, %k0 {%k1}
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    vzeroupper
@@ -233,11 +203,9 @@ define zeroext i8 @TEST_mm256_mask_test_epi32_mask(i8 %__U, <4 x i64> %__A, <4 x
 ;
 ; I386-LABEL: TEST_mm256_mask_test_epi32_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; I386-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
 ; I386-NEXT:    kmovw %eax, %k1
-; I386-NEXT:    vpcmpneqd %ymm1, %ymm0, %k0 {%k1}
+; I386-NEXT:    vptestmd %ymm0, %ymm1, %k0 {%k1}
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    vzeroupper
@@ -256,18 +224,14 @@ entry:
 define zeroext i8 @TEST_mm_testn_epi64_mask(<2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm_testn_epi64_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; X86_64-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0
+; X86_64-NEXT:    vptestnmq %xmm0, %xmm1, %k0
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    retq
 ;
 ; I386-LABEL: TEST_mm_testn_epi64_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; I386-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0
+; I386-NEXT:    vptestnmq %xmm0, %xmm1, %k0
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    retl
@@ -283,18 +247,14 @@ entry:
 define zeroext i8 @TEST_mm_testn_epi32_mask(<2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm_testn_epi32_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; X86_64-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0
+; X86_64-NEXT:    vptestnmd %xmm0, %xmm1, %k0
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    retq
 ;
 ; I386-LABEL: TEST_mm_testn_epi32_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; I386-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0
+; I386-NEXT:    vptestnmd %xmm0, %xmm1, %k0
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    retl
@@ -311,9 +271,7 @@ entry:
 define zeroext i8 @TEST_mm256_testn_epi64_mask(<4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm256_testn_epi64_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; X86_64-NEXT:    vpcmpeqq %ymm1, %ymm0, %k0
+; X86_64-NEXT:    vptestnmq %ymm0, %ymm1, %k0
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    vzeroupper
@@ -321,9 +279,7 @@ define zeroext i8 @TEST_mm256_testn_epi64_mask(<4 x i64> %__A, <4 x i64> %__B) l
 ;
 ; I386-LABEL: TEST_mm256_testn_epi64_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; I386-NEXT:    vpcmpeqq %ymm1, %ymm0, %k0
+; I386-NEXT:    vptestnmq %ymm0, %ymm1, %k0
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    vzeroupper
@@ -340,9 +296,7 @@ entry:
 define zeroext i8 @TEST_mm256_testn_epi32_mask(<4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm256_testn_epi32_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; X86_64-NEXT:    vpcmpeqd %ymm1, %ymm0, %k0
+; X86_64-NEXT:    vptestnmd %ymm0, %ymm1, %k0
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    vzeroupper
@@ -350,9 +304,7 @@ define zeroext i8 @TEST_mm256_testn_epi32_mask(<4 x i64> %__A, <4 x i64> %__B) l
 ;
 ; I386-LABEL: TEST_mm256_testn_epi32_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; I386-NEXT:    vpcmpeqd %ymm1, %ymm0, %k0
+; I386-NEXT:    vptestnmd %ymm0, %ymm1, %k0
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    vzeroupper
@@ -369,21 +321,17 @@ entry:
 define zeroext i8 @TEST_mm_mask_testn_epi64_mask(i8 %__U, <2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm_mask_testn_epi64_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; X86_64-NEXT:    kmovw %edi, %k1
-; X86_64-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0 {%k1}
+; X86_64-NEXT:    vptestnmq %xmm0, %xmm1, %k0 {%k1}
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    retq
 ;
 ; I386-LABEL: TEST_mm_mask_testn_epi64_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; I386-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
 ; I386-NEXT:    kmovw %eax, %k1
-; I386-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0 {%k1}
+; I386-NEXT:    vptestnmq %xmm0, %xmm1, %k0 {%k1}
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    retl
@@ -402,21 +350,17 @@ entry:
 define zeroext i8 @TEST_mm_mask_testn_epi32_mask(i8 %__U, <2 x i64> %__A, <2 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm_mask_testn_epi32_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; X86_64-NEXT:    kmovw %edi, %k1
-; X86_64-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0 {%k1}
+; X86_64-NEXT:    vptestnmd %xmm0, %xmm1, %k0 {%k1}
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    retq
 ;
 ; I386-LABEL: TEST_mm_mask_testn_epi32_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; I386-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
 ; I386-NEXT:    kmovw %eax, %k1
-; I386-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0 {%k1}
+; I386-NEXT:    vptestnmd %xmm0, %xmm1, %k0 {%k1}
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    retl
@@ -437,10 +381,8 @@ entry:
 define zeroext i8 @TEST_mm256_mask_testn_epi64_mask(i8 %__U, <4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm256_mask_testn_epi64_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; X86_64-NEXT:    kmovw %edi, %k1
-; X86_64-NEXT:    vpcmpeqq %ymm1, %ymm0, %k0 {%k1}
+; X86_64-NEXT:    vptestnmq %ymm0, %ymm1, %k0 {%k1}
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    vzeroupper
@@ -448,11 +390,9 @@ define zeroext i8 @TEST_mm256_mask_testn_epi64_mask(i8 %__U, <4 x i64> %__A, <4
 ;
 ; I386-LABEL: TEST_mm256_mask_testn_epi64_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; I386-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
 ; I386-NEXT:    kmovw %eax, %k1
-; I386-NEXT:    vpcmpeqq %ymm1, %ymm0, %k0 {%k1}
+; I386-NEXT:    vptestnmq %ymm0, %ymm1, %k0 {%k1}
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    vzeroupper
@@ -472,10 +412,8 @@ entry:
 define zeroext i8 @TEST_mm256_mask_testn_epi32_mask(i8 %__U, <4 x i64> %__A, <4 x i64> %__B) local_unnamed_addr #0 {
 ; X86_64-LABEL: TEST_mm256_mask_testn_epi32_mask:
 ; X86_64:       # BB#0: # %entry
-; X86_64-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; X86_64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; X86_64-NEXT:    kmovw %edi, %k1
-; X86_64-NEXT:    vpcmpeqd %ymm1, %ymm0, %k0 {%k1}
+; X86_64-NEXT:    vptestnmd %ymm0, %ymm1, %k0 {%k1}
 ; X86_64-NEXT:    kmovw %k0, %eax
 ; X86_64-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; X86_64-NEXT:    vzeroupper
@@ -483,11 +421,9 @@ define zeroext i8 @TEST_mm256_mask_testn_epi32_mask(i8 %__U, <4 x i64> %__A, <4
 ;
 ; I386-LABEL: TEST_mm256_mask_testn_epi32_mask:
 ; I386:       # BB#0: # %entry
-; I386-NEXT:    vpand %ymm0, %ymm1, %ymm0
-; I386-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; I386-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
 ; I386-NEXT:    kmovw %eax, %k1
-; I386-NEXT:    vpcmpeqd %ymm1, %ymm0, %k0 {%k1}
+; I386-NEXT:    vptestnmd %ymm0, %ymm1, %k0 {%k1}
 ; I386-NEXT:    kmovw %k0, %eax
 ; I386-NEXT:    # kill: %AL<def> %AL<kill> %EAX<kill>
 ; I386-NEXT:    vzeroupper
diff --git a/llvm/test/CodeGen/X86/compress_expand.ll b/llvm/test/CodeGen/X86/compress_expand.ll
index c6a1c07..9237544 100644
--- a/llvm/test/CodeGen/X86/compress_expand.ll
+++ b/llvm/test/CodeGen/X86/compress_expand.ll
@@ -140,9 +140,7 @@ define void @test7(float* %base, <8 x float> %V, <8 x i1> %mask) {
 ; KNL-NEXT:    # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
 ; KNL-NEXT:    vpmovsxwq %xmm1, %zmm1
 ; KNL-NEXT:    vpsllq $63, %zmm1, %zmm1
-; KNL-NEXT:    vptestmq %zmm1, %zmm1, %k0
-; KNL-NEXT:    kshiftlw $8, %k0, %k0
-; KNL-NEXT:    kshiftrw $8, %k0, %k1
+; KNL-NEXT:    vptestmq %zmm1, %zmm1, %k1
 ; KNL-NEXT:    vcompressps %zmm0, (%rdi) {%k1}
 ; KNL-NEXT:    retq
   call void @llvm.masked.compressstore.v8f32(<8 x float> %V, float* %base, <8 x i1> %mask)
diff --git a/llvm/test/CodeGen/X86/masked_gather_scatter.ll b/llvm/test/CodeGen/X86/masked_gather_scatter.ll
index 8983c3a..94057f9 100644
--- a/llvm/test/CodeGen/X86/masked_gather_scatter.ll
+++ b/llvm/test/CodeGen/X86/masked_gather_scatter.ll
@@ -1057,9 +1057,7 @@ define void @test20(<2 x float>%a1, <2 x float*> %ptr, <2 x i1> %mask) {
 ; SKX:       # BB#0:
 ; SKX-NEXT:    # kill: %XMM1<def> %XMM1<kill> %YMM1<def>
 ; SKX-NEXT:    vpsllq $63, %xmm2, %xmm2
-; SKX-NEXT:    vptestmq %xmm2, %xmm2, %k0
-; SKX-NEXT:    kshiftlb $6, %k0, %k0
-; SKX-NEXT:    kshiftrb $6, %k0, %k1
+; SKX-NEXT:    vptestmq %xmm2, %xmm2, %k1
 ; SKX-NEXT:    vscatterqps %xmm0, (,%ymm1) {%k1}
 ; SKX-NEXT:    vzeroupper
 ; SKX-NEXT:    retq
@@ -1068,9 +1066,7 @@ define void @test20(<2 x float>%a1, <2 x float*> %ptr, <2 x i1> %mask) {
 ; SKX_32:       # BB#0:
 ; SKX_32-NEXT:    vpermilps {{.*#+}} xmm1 = xmm1[0,2,2,3]
 ; SKX_32-NEXT:    vpsllq $63, %xmm2, %xmm2
-; SKX_32-NEXT:    vptestmq %xmm2, %xmm2, %k0
-; SKX_32-NEXT:    kshiftlb $6, %k0, %k0
-; SKX_32-NEXT:    kshiftrb $6, %k0, %k1
+; SKX_32-NEXT:    vptestmq %xmm2, %xmm2, %k1
 ; SKX_32-NEXT:    vscatterdps %xmm0, (,%xmm1) {%k1}
 ; SKX_32-NEXT:    retl
   call void @llvm.masked.scatter.v2f32.v2p0f32(<2 x float> %a1, <2 x float*> %ptr, i32 4, <2 x i1> %mask)
@@ -1105,9 +1101,7 @@ define void @test21(<2 x i32>%a1, <2 x i32*> %ptr, <2 x i1>%mask) {
 ; SKX:       # BB#0:
 ; SKX-NEXT:    # kill: %XMM1<def> %XMM1<kill> %YMM1<def>
 ; SKX-NEXT:    vpsllq $63, %xmm2, %xmm2
-; SKX-NEXT:    vptestmq %xmm2, %xmm2, %k0
-; SKX-NEXT:    kshiftlb $6, %k0, %k0
-; SKX-NEXT:    kshiftrb $6, %k0, %k1
+; SKX-NEXT:    vptestmq %xmm2, %xmm2, %k1
 ; SKX-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; SKX-NEXT:    vpscatterqd %xmm0, (,%ymm1) {%k1}
 ; SKX-NEXT:    vzeroupper
@@ -1117,9 +1111,7 @@ define void @test21(<2 x i32>%a1, <2 x i32*> %ptr, <2 x i1>%mask) {
 ; SKX_32:       # BB#0:
 ; SKX_32-NEXT:    # kill: %XMM1<def> %XMM1<kill> %YMM1<def>
 ; SKX_32-NEXT:    vpsllq $63, %xmm2, %xmm2
-; SKX_32-NEXT:    vptestmq %xmm2, %xmm2, %k0
-; SKX_32-NEXT:    kshiftlb $6, %k0, %k0
-; SKX_32-NEXT:    kshiftrb $6, %k0, %k1
+; SKX_32-NEXT:    vptestmq %xmm2, %xmm2, %k1
 ; SKX_32-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; SKX_32-NEXT:    vpscatterqd %xmm0, (,%ymm1) {%k1}
 ; SKX_32-NEXT:    vzeroupper
@@ -1165,9 +1157,7 @@ define <2 x float> @test22(float* %base, <2 x i32> %ind, <2 x i1> %mask, <2 x fl
 ; SKX:       # BB#0:
 ; SKX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; SKX-NEXT:    vpsllq $63, %xmm1, %xmm1
-; SKX-NEXT:    vptestmq %xmm1, %xmm1, %k0
-; SKX-NEXT:    kshiftlb $6, %k0, %k0
-; SKX-NEXT:    kshiftrb $6, %k0, %k1
+; SKX-NEXT:    vptestmq %xmm1, %xmm1, %k1
 ; SKX-NEXT:    vgatherdps (%rdi,%xmm0,4), %xmm2 {%k1}
 ; SKX-NEXT:    vmovaps %xmm2, %xmm0
 ; SKX-NEXT:    retq
@@ -1176,9 +1166,7 @@ define <2 x float> @test22(float* %base, <2 x i32> %ind, <2 x i1> %mask, <2 x fl
 ; SKX_32:       # BB#0:
 ; SKX_32-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; SKX_32-NEXT:    vpsllq $63, %xmm1, %xmm1
-; SKX_32-NEXT:    vptestmq %xmm1, %xmm1, %k0
-; SKX_32-NEXT:    kshiftlb $6, %k0, %k0
-; SKX_32-NEXT:    kshiftrb $6, %k0, %k1
+; SKX_32-NEXT:    vptestmq %xmm1, %xmm1, %k1
 ; SKX_32-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; SKX_32-NEXT:    vgatherdps (%eax,%xmm0,4), %xmm2 {%k1}
 ; SKX_32-NEXT:    vmovaps %xmm2, %xmm0
diff --git a/llvm/test/CodeGen/X86/masked_memop.ll b/llvm/test/CodeGen/X86/masked_memop.ll
index 3e257f5..f43e3f6 100644
--- a/llvm/test/CodeGen/X86/masked_memop.ll
+++ b/llvm/test/CodeGen/X86/masked_memop.ll
@@ -285,9 +285,7 @@ define <8 x i32> @test11b(<8 x i1> %mask, <8 x i32>* %addr, <8 x i32> %dst) {
 ; AVX512F-NEXT:    ## kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
 ; AVX512F-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; AVX512F-NEXT:    vpsllq $63, %zmm0, %zmm0
-; AVX512F-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; AVX512F-NEXT:    kshiftlw $8, %k0, %k0
-; AVX512F-NEXT:    kshiftrw $8, %k0, %k1
+; AVX512F-NEXT:    vptestmq %zmm0, %zmm0, %k1
 ; AVX512F-NEXT:    vpblendmd (%rdi), %zmm1, %zmm0 {%k1}
 ; AVX512F-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
 ; AVX512F-NEXT:    retq
@@ -327,9 +325,7 @@ define <8 x float> @test11c(<8 x i1> %mask, <8 x float>* %addr) {
 ; AVX512F:       ## BB#0:
 ; AVX512F-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; AVX512F-NEXT:    vpsllq $63, %zmm0, %zmm0
-; AVX512F-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; AVX512F-NEXT:    kshiftlw $8, %k0, %k0
-; AVX512F-NEXT:    kshiftrw $8, %k0, %k1
+; AVX512F-NEXT:    vptestmq %zmm0, %zmm0, %k1
 ; AVX512F-NEXT:    vmovups (%rdi), %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
 ; AVX512F-NEXT:    retq
@@ -369,9 +365,7 @@ define <8 x i32> @test11d(<8 x i1> %mask, <8 x i32>* %addr) {
 ; AVX512F:       ## BB#0:
 ; AVX512F-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; AVX512F-NEXT:    vpsllq $63, %zmm0, %zmm0
-; AVX512F-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; AVX512F-NEXT:    kshiftlw $8, %k0, %k0
-; AVX512F-NEXT:    kshiftrw $8, %k0, %k1
+; AVX512F-NEXT:    vptestmq %zmm0, %zmm0, %k1
 ; AVX512F-NEXT:    vmovdqu32 (%rdi), %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
 ; AVX512F-NEXT:    retq
diff --git a/llvm/test/CodeGen/X86/setcc-lowering.ll b/llvm/test/CodeGen/X86/setcc-lowering.ll
index 20c77a4..359b8d6 100644
--- a/llvm/test/CodeGen/X86/setcc-lowering.ll
+++ b/llvm/test/CodeGen/X86/setcc-lowering.ll
@@ -23,10 +23,9 @@ define <8 x i16> @pr25080(<8 x i32> %a) {
 ;
 ; KNL-32-LABEL: pr25080:
 ; KNL-32:       # BB#0: # %entry
-; KNL-32-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [8388607,8388607,8388607,8388607,8388607,8388607,8388607,8388607]
-; KNL-32-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; KNL-32-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; KNL-32-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
+; KNL-32-NEXT:    # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
+; KNL-32-NEXT:    vbroadcastss {{.*#+}} ymm1 = [8388607,8388607,8388607,8388607,8388607,8388607,8388607,8388607]
+; KNL-32-NEXT:    vptestnmd %zmm1, %zmm0, %k0
 ; KNL-32-NEXT:    movb $15, %al
 ; KNL-32-NEXT:    kmovw %eax, %k1
 ; KNL-32-NEXT:    korw %k1, %k0, %k1