[X86][BtVer2] Fix the number of micro opcodes for a bunch of YMM instructions.

The Jaguar backend natively supports 128-bit data types. Operations on YMM
registers are split into two COPs (complex operations). Each COP consumes a slot
in the dispatch group, and in the reorder buffer.

The scheduling model for Jaguar should mark those instructions as `let
NumMicroOps = 2`.

This was found when testing AVX code for BtVer2 using llvm-mca.

llvm-svn: 328694

diff --git a/llvm/lib/Target/X86/X86ScheduleBtVer2.td b/llvm/lib/Target/X86/X86ScheduleBtVer2.td
index 4d88383..1b4d1ad 100644 (file)
--- a/llvm/lib/Target/X86/X86ScheduleBtVer2.td
+++ b/llvm/lib/Target/X86/X86ScheduleBtVer2.td
@@ -547,6 +547,7 @@ def : InstRW<[JWriteVDPPSYLd, ReadAfterLd], (instrs VDPPSYrmi)>;
 def JWriteFAddY: SchedWriteRes<[JFPU0, JFPA]> {
   let Latency = 3;
   let ResourceCycles = [2, 2];
+  let NumMicroOps = 2;
 }
 def : InstRW<[JWriteFAddY], (instrs VADDPDYrr, VADDPSYrr,
                                     VSUBPDYrr, VSUBPSYrr,
@@ -555,6 +556,7 @@ def : InstRW<[JWriteFAddY], (instrs VADDPDYrr, VADDPSYrr,
 def JWriteFAddYLd: SchedWriteRes<[JLAGU, JFPU0, JFPA]> {
   let Latency = 8;
   let ResourceCycles = [2, 2, 2];
+  let NumMicroOps = 2;
 }
 def : InstRW<[JWriteFAddYLd, ReadAfterLd], (instrs VADDPDYrm, VADDPSYrm,
                                                    VSUBPDYrm, VSUBPSYrm,
@@ -563,36 +565,42 @@ def : InstRW<[JWriteFAddYLd, ReadAfterLd], (instrs VADDPDYrm, VADDPSYrm,
 def JWriteFDivY: SchedWriteRes<[JFPU1, JFPM]> {
   let Latency = 38;
   let ResourceCycles = [2, 38];
+  let NumMicroOps = 2;
 }
 def : InstRW<[JWriteFDivY], (instrs VDIVPDYrr, VDIVPSYrr)>;
 
 def JWriteFDivYLd: SchedWriteRes<[JLAGU, JFPU1, JFPM]> {
   let Latency = 43;
   let ResourceCycles = [2, 2, 38];
+  let NumMicroOps = 2;
 }
 def : InstRW<[JWriteFDivYLd, ReadAfterLd], (instrs VDIVPDYrm, VDIVPSYrm)>;
 
 def JWriteVMULYPD: SchedWriteRes<[JFPU1, JFPM]> {
   let Latency = 4;
   let ResourceCycles = [2, 4];
+  let NumMicroOps = 2;
 }
 def : InstRW<[JWriteVMULYPD], (instrs VMULPDYrr)>;
 
 def JWriteVMULYPDLd: SchedWriteRes<[JLAGU, JFPU1, JFPM]> {
   let Latency = 9;
   let ResourceCycles = [2, 2, 4];
+  let NumMicroOps = 2;
 }
 def : InstRW<[JWriteVMULYPDLd, ReadAfterLd], (instrs VMULPDYrm)>;
 
 def JWriteVMULYPS: SchedWriteRes<[JFPU1, JFPM]> {
   let Latency = 2;
   let ResourceCycles = [2, 2];
+  let NumMicroOps = 2;
 }
 def : InstRW<[JWriteVMULYPS], (instrs VMULPSYrr, VRCPPSYr, VRSQRTPSYr)>;
 
 def JWriteVMULYPSLd: SchedWriteRes<[JLAGU, JFPU1, JFPM]> {
   let Latency = 7;
   let ResourceCycles = [2, 2, 2];
+  let NumMicroOps = 2;
 }
 def : InstRW<[JWriteVMULYPSLd, ReadAfterLd], (instrs VMULPSYrm, VRCPPSYm, VRSQRTPSYm)>;
 
@@ -611,6 +619,7 @@ def : InstRW<[JWriteVMULPDLd], (instrs MULPDrm, MULSDrm, VMULPDrm, VMULSDrm)>;
 def JWriteVCVTY: SchedWriteRes<[JFPU1, JSTC]> {
   let Latency = 3;
   let ResourceCycles = [2, 2];
+  let NumMicroOps = 2;
 }
 def : InstRW<[JWriteVCVTY], (instrs VCVTDQ2PDYrr, VCVTDQ2PSYrr,
                                     VCVTPS2DQYrr, VCVTTPS2DQYrr,
@@ -619,6 +628,7 @@ def : InstRW<[JWriteVCVTY], (instrs VCVTDQ2PDYrr, VCVTDQ2PSYrr,
 def JWriteVCVTYLd: SchedWriteRes<[JLAGU, JFPU1, JSTC]> {
   let Latency = 8;
   let ResourceCycles = [2, 2, 2];
+  let NumMicroOps = 2;
 }
 def : InstRW<[JWriteVCVTYLd, ReadAfterLd], (instrs VCVTDQ2PDYrm, VCVTDQ2PSYrm,
                                                    VCVTPS2DQYrm, VCVTTPS2DQYrm,
@@ -834,12 +844,14 @@ def : InstRW<[JWriteVSQRTYPDLd], (instrs VSQRTPDYm)>;
 def JWriteVSQRTYPS: SchedWriteRes<[JFPU1, JFPM]> {
   let Latency = 42;
   let ResourceCycles = [2, 42];
+  let NumMicroOps = 2;
 }
 def : InstRW<[JWriteVSQRTYPS], (instrs VSQRTPSYr)>;
 
 def JWriteVSQRTYPSLd: SchedWriteRes<[JLAGU, JFPU1, JFPM]> {
   let Latency = 47;
   let ResourceCycles = [2, 2, 42];
+  let NumMicroOps = 2;
 }
 def : InstRW<[JWriteVSQRTYPSLd], (instrs VSQRTPSYm)>;
 

diff --git a/llvm/test/tools/llvm-mca/X86/BtVer2/pipes-fpu.s b/llvm/test/tools/llvm-mca/X86/BtVer2/pipes-fpu.s
index 2fd3efe..89b9b7c 100644 (file)
--- a/llvm/test/tools/llvm-mca/X86/BtVer2/pipes-fpu.s
+++ b/llvm/test/tools/llvm-mca/X86/BtVer2/pipes-fpu.s
@@ -39,8 +39,8 @@ vsqrtps     %ymm0, %ymm2
 # CHECK-NEXT:  1      2     1.00                       vpclmulqdq      $0, %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      3     1.00                       vaddps  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      21    21.00                      vsqrtps %xmm0, %xmm2
-# CHECK-NEXT:  1      3     2.00                       vaddps  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      42    42.00                      vsqrtps %ymm0, %ymm2
+# CHECK-NEXT:  2      3     2.00                       vaddps  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      42    42.00                      vsqrtps %ymm0, %ymm2
 
 
 # CHECK:      Resources:
@@ -87,13 +87,13 @@ vsqrtps     %ymm0, %ymm2
 # CHECK-NEXT: [0,4]    . DeeeER  .    .    .    .    .    .    .    .    .    .    .    . .    vaddps  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT: [0,5]    . DeeeeeeeeeeeeeeeeeeeeeER    .    .    .    .    .    .    .    . .    vsqrtps %xmm0, %xmm2
 # CHECK-NEXT: [0,6]    .  DeeeE-----------------R    .    .    .    .    .    .    .    . .    vaddps  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT: [0,7]    .  D====================eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeER    vsqrtps %ymm0, %ymm2
+# CHECK-NEXT: [0,7]    .   D===================eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeER    vsqrtps %ymm0, %ymm2
 
-# CHECK:      [1,0]    .   D=eeE----------------------------------------------------------R    vpmulld %xmm0, %xmm1, %xmm2
-# CHECK-NEXT: [1,1]    .   DeE------------------------------------------------------------R    vpand   %xmm0, %xmm1, %xmm2
-# CHECK-NEXT: [1,2]    .    DeeeE---------------------------------------------------------R    vcvttps2dq      %xmm0, %xmm2
-# CHECK-NEXT: [1,3]    .    D=eeE---------------------------------------------------------R    vpclmulqdq      $0, %xmm0, %xmm1, %xmm2
-# CHECK-NEXT: [1,4]    .    .D=eeeE-------------------------------------------------------R    vaddps  %xmm0, %xmm1, %xmm2
+# CHECK:      [1,0]    .    DeeE----------------------------------------------------------R    vpmulld %xmm0, %xmm1, %xmm2
+# CHECK-NEXT: [1,1]    .    DeE-----------------------------------------------------------R    vpand   %xmm0, %xmm1, %xmm2
+# CHECK-NEXT: [1,2]    .    .DeeeE--------------------------------------------------------R    vcvttps2dq      %xmm0, %xmm2
+# CHECK-NEXT: [1,3]    .    .DeeE---------------------------------------------------------R    vpclmulqdq      $0, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT: [1,4]    .    . DeeeE-------------------------------------------------------R    vaddps  %xmm0, %xmm1, %xmm2
 
 
 # CHECK:      Average Wait times (based on the timeline view):
@@ -103,11 +103,11 @@ vsqrtps     %ymm0, %ymm2
 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 
 # CHECK:            [0]    [1]    [2]    [3]
-# CHECK-NEXT: 0.     2     1.5    1.5    29.0  vpmulld %xmm0, %xmm1, %xmm2
-# CHECK-NEXT: 1.     2     1.0    1.0    30.5  vpand   %xmm0, %xmm1, %xmm2
-# CHECK-NEXT: 2.     2     1.0    1.0    28.5  vcvttps2dq      %xmm0, %xmm2
-# CHECK-NEXT: 3.     2     1.5    1.5    29.0  vpclmulqdq      $0, %xmm0, %xmm1, %xmm2
-# CHECK-NEXT: 4.     2     1.5    1.5    27.5  vaddps  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT: 0.     2     1.0    1.0    29.0  vpmulld %xmm0, %xmm1, %xmm2
+# CHECK-NEXT: 1.     2     1.0    1.0    30.0  vpand   %xmm0, %xmm1, %xmm2
+# CHECK-NEXT: 2.     2     1.0    1.0    28.0  vcvttps2dq      %xmm0, %xmm2
+# CHECK-NEXT: 3.     2     1.0    1.0    29.0  vpclmulqdq      $0, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT: 4.     2     1.0    1.0    27.5  vaddps  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT: 5.     1     1.0    1.0    0.0   vsqrtps %xmm0, %xmm2
 # CHECK-NEXT: 6.     1     1.0    1.0    17.0  vaddps  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT: 7.     1     21.0   21.0   0.0   vsqrtps %ymm0, %ymm2
+# CHECK-NEXT: 7.     1     20.0   20.0   0.0   vsqrtps %ymm0, %ymm2

diff --git a/llvm/test/tools/llvm-mca/X86/BtVer2/resources-avx1.s b/llvm/test/tools/llvm-mca/X86/BtVer2/resources-avx1.s
index 8d79f19..d381fb3 100644 (file)
--- a/llvm/test/tools/llvm-mca/X86/BtVer2/resources-avx1.s
+++ b/llvm/test/tools/llvm-mca/X86/BtVer2/resources-avx1.s
@@ -1008,6 +1008,701 @@ vxorps            (%rax), %ymm1, %ymm2
 vzeroall
 vzeroupper
 
+
+# CHECK:      Instruction Info:
+# CHECK-NEXT: [1]: #uOps
+# CHECK-NEXT: [2]: Latency
+# CHECK-NEXT: [3]: RThroughput
+# CHECK-NEXT: [4]: MayLoad
+# CHECK-NEXT: [5]: MayStore
+# CHECK-NEXT: [6]: HasSideEffects
+
+# CHECK:      [1]    [2]    [3]    [4]    [5]    [6]   Instructions:
+# CHECK-NEXT:  1      3     1.00                       vaddpd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vaddpd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      3     2.00                       vaddpd  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      8     2.00    *                  vaddpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     1.00                       vaddps  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vaddps  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      3     2.00                       vaddps  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      8     2.00    *                  vaddps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     1.00                       vaddsd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vaddsd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                       vaddss  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vaddss  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                       vaddsubpd       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vaddsubpd       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      3     2.00                       vaddsubpd       %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      8     2.00    *                  vaddsubpd       (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     1.00                       vaddsubps       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vaddsubps       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      3     2.00                       vaddsubps       %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      8     2.00    *                  vaddsubps       (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     1.00                       vaesdec %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vaesdec (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                       vaesdeclast     %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vaesdeclast     (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                       vaesenc %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vaesenc (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                       vaesenclast     %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vaesenclast     (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vaesimc %xmm0, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vaesimc (%rax), %xmm2
+# CHECK-NEXT:  1      2     1.00                       vaeskeygenassist        $22, %xmm0, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vaeskeygenassist        $22, (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vandnpd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vandnpd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vandnpd %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vandnpd (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vandnps %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vandnps (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vandnps %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vandnps (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vandpd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vandpd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vandpd  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vandpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vandps  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vandps  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vandps  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vandps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vblendpd        $11, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vblendpd        $11, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vblendpd        $11, %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vblendpd        $11, (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vblendps        $11, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vblendps        $11, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vblendps        $11, %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vblendps        $11, (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  3      2     2.00                       vblendvpd       %xmm3, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  3      7     2.00    *                  vblendvpd       %xmm3, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  6      3     3.00                       vblendvpd       %ymm3, %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  6      8     3.00    *                  vblendvpd       %ymm3, (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  3      2     2.00                       vblendvps       %xmm3, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  3      7     2.00    *                  vblendvps       %xmm3, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  6      3     3.00                       vblendvps       %ymm3, %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  6      8     3.00    *                  vblendvps       %ymm3, (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      6     1.00    *                  vbroadcastf128  (%rax), %ymm2
+# CHECK-NEXT:  1      6     2.00    *                  vbroadcastsd    (%rax), %ymm2
+# CHECK-NEXT:  1      6     1.00    *                  vbroadcastss    (%rax), %xmm2
+# CHECK-NEXT:  1      6     2.00    *                  vbroadcastss    (%rax), %ymm2
+# CHECK-NEXT:  1      2     1.00                       vcmppd  $0, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vcmppd  $0, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     2.00                       vcmppd  $0, %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      7     2.00    *                  vcmppd  $0, (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      2     1.00                       vcmpps  $0, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vcmpps  $0, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     2.00                       vcmpps  $0, %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      7     2.00    *                  vcmpps  $0, (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      2     1.00                       vcmpsd  $0, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vcmpsd  $0, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vcmpss  $0, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vcmpss  $0, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                       vcomisd %xmm0, %xmm1
+# CHECK-NEXT:  1      8     1.00    *                  vcomisd (%rax), %xmm1
+# CHECK-NEXT:  1      3     1.00                       vcomiss %xmm0, %xmm1
+# CHECK-NEXT:  1      8     1.00    *                  vcomiss (%rax), %xmm1
+# CHECK-NEXT:  1      3     1.00                       vcvtdq2pd       %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vcvtdq2pd       (%rax), %xmm2
+# CHECK-NEXT:  2      3     2.00                       vcvtdq2pd       %xmm0, %ymm2
+# CHECK-NEXT:  2      8     2.00    *                  vcvtdq2pd       (%rax), %ymm2
+# CHECK-NEXT:  1      3     1.00                       vcvtdq2ps       %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vcvtdq2ps       (%rax), %xmm2
+# CHECK-NEXT:  2      3     2.00                       vcvtdq2ps       %ymm0, %ymm2
+# CHECK-NEXT:  2      8     2.00    *                  vcvtdq2ps       (%rax), %ymm2
+# CHECK-NEXT:  1      3     1.00                       vcvtpd2dq       %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vcvtpd2dqx      (%rax), %xmm2
+# CHECK-NEXT:  1      6     2.00                       vcvtpd2dq       %ymm0, %xmm2
+# CHECK-NEXT:  1      11    2.00    *                  vcvtpd2dqy      (%rax), %xmm2
+# CHECK-NEXT:  1      3     1.00                       vcvtpd2ps       %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vcvtpd2psx      (%rax), %xmm2
+# CHECK-NEXT:  1      6     2.00                       vcvtpd2ps       %ymm0, %xmm2
+# CHECK-NEXT:  1      11    2.00    *                  vcvtpd2psy      (%rax), %xmm2
+# CHECK-NEXT:  1      3     1.00                       vcvtps2dq       %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vcvtps2dq       (%rax), %xmm2
+# CHECK-NEXT:  2      3     2.00                       vcvtps2dq       %ymm0, %ymm2
+# CHECK-NEXT:  2      8     2.00    *                  vcvtps2dq       (%rax), %ymm2
+# CHECK-NEXT:  1      3     1.00                       vcvtps2pd       %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vcvtps2pd       (%rax), %xmm2
+# CHECK-NEXT:  1      3     1.00                       vcvtps2pd       %xmm0, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                  vcvtps2pd       (%rax), %ymm2
+# CHECK-NEXT:  2      7     1.00                       vcvtsd2si       %xmm0, %ecx
+# CHECK-NEXT:  2      7     1.00                       vcvtsd2si       %xmm0, %rcx
+# CHECK-NEXT:  2      12    1.00    *                  vcvtsd2si       (%rax), %ecx
+# CHECK-NEXT:  2      12    1.00    *                  vcvtsd2si       (%rax), %rcx
+# CHECK-NEXT:  2      7     2.00                       vcvtsd2ss       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  2      12    2.00    *                  vcvtsd2ss       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      9     1.00                       vcvtsi2sdl      %ecx, %xmm0, %xmm2
+# CHECK-NEXT:  2      9     1.00                       vcvtsi2sdq      %rcx, %xmm0, %xmm2
+# CHECK-NEXT:  2      14    1.00    *                  vcvtsi2sdl      (%rax), %xmm0, %xmm2
+# CHECK-NEXT:  2      14    1.00    *                  vcvtsi2sdq      (%rax), %xmm0, %xmm2
+# CHECK-NEXT:  2      9     1.00                       vcvtsi2ssl      %ecx, %xmm0, %xmm2
+# CHECK-NEXT:  2      9     1.00                       vcvtsi2ssq      %rcx, %xmm0, %xmm2
+# CHECK-NEXT:  2      14    1.00    *                  vcvtsi2ssl      (%rax), %xmm0, %xmm2
+# CHECK-NEXT:  2      14    1.00    *                  vcvtsi2ssq      (%rax), %xmm0, %xmm2
+# CHECK-NEXT:  2      7     2.00                       vcvtss2sd       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  2      12    2.00    *                  vcvtss2sd       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      7     1.00                       vcvtss2si       %xmm0, %ecx
+# CHECK-NEXT:  2      7     1.00                       vcvtss2si       %xmm0, %rcx
+# CHECK-NEXT:  2      12    1.00    *                  vcvtss2si       (%rax), %ecx
+# CHECK-NEXT:  2      12    1.00    *                  vcvtss2si       (%rax), %rcx
+# CHECK-NEXT:  1      3     1.00                       vcvttpd2dq      %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vcvttpd2dqx     (%rax), %xmm2
+# CHECK-NEXT:  1      6     2.00                       vcvttpd2dq      %ymm0, %xmm2
+# CHECK-NEXT:  1      11    2.00    *                  vcvttpd2dqy     (%rax), %xmm2
+# CHECK-NEXT:  1      3     1.00                       vcvttps2dq      %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vcvttps2dq      (%rax), %xmm2
+# CHECK-NEXT:  2      3     2.00                       vcvttps2dq      %ymm0, %ymm2
+# CHECK-NEXT:  2      8     2.00    *                  vcvttps2dq      (%rax), %ymm2
+# CHECK-NEXT:  2      7     1.00                       vcvttsd2si      %xmm0, %ecx
+# CHECK-NEXT:  2      7     1.00                       vcvttsd2si      %xmm0, %rcx
+# CHECK-NEXT:  2      12    1.00    *                  vcvttsd2si      (%rax), %ecx
+# CHECK-NEXT:  2      12    1.00    *                  vcvttsd2si      (%rax), %rcx
+# CHECK-NEXT:  2      7     1.00                       vcvttss2si      %xmm0, %ecx
+# CHECK-NEXT:  2      7     1.00                       vcvttss2si      %xmm0, %rcx
+# CHECK-NEXT:  2      12    1.00    *                  vcvttss2si      (%rax), %ecx
+# CHECK-NEXT:  2      12    1.00    *                  vcvttss2si      (%rax), %rcx
+# CHECK-NEXT:  1      19    19.00                      vdivpd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      24    19.00   *                  vdivpd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      38    38.00                      vdivpd  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      43    38.00   *                  vdivpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      19    19.00                      vdivps  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      24    19.00   *                  vdivps  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      38    38.00                      vdivps  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      43    38.00   *                  vdivps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      19    19.00                      vdivsd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      24    19.00   *                  vdivsd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      19    19.00                      vdivss  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      24    19.00   *                  vdivss  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  3      9     3.00                       vdppd   $22, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  3      14    3.00    *                  vdppd   $22, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  5      11    3.00                       vdpps   $22, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  5      16    3.00    *                  vdpps   $22, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  10     12    6.00                       vdpps   $22, %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  10     17    6.00    *                  vdpps   $22, (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vextractf128    $1, %ymm0, %xmm2
+# CHECK-NEXT:  1      1     1.00           *           vextractf128    $1, %ymm0, (%rax)
+# CHECK-NEXT:  1      1     0.50                       vextractps      $1, %xmm0, %ecx
+# CHECK-NEXT:  2      6     1.00           *           vextractps      $1, %xmm0, (%rax)
+# CHECK-NEXT:  1      3     1.00                       vhaddpd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vhaddpd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     2.00                       vhaddpd %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     2.00    *                  vhaddpd (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     1.00                       vhaddps %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vhaddps (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     2.00                       vhaddps %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     2.00    *                  vhaddps (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     1.00                       vhsubpd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vhsubpd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     2.00                       vhsubpd %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     2.00    *                  vhsubpd (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     1.00                       vhsubps %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vhsubps (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     2.00                       vhsubps %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     2.00    *                  vhsubps (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vinsertf128     $1, %xmm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      6     1.00    *                  vinsertf128     $1, (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vinsertps       $1, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vinsertps       $1, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      5     1.00    *                  vlddqu  (%rax), %xmm2
+# CHECK-NEXT:  1      5     1.00    *                  vlddqu  (%rax), %ymm2
+# CHECK-NEXT:  1      5     1.00    *      *      *    vldmxcsr        (%rax)
+# CHECK-NEXT:  1      1     1.00    *      *      *    vmaskmovdqu     %xmm0, %xmm1
+# CHECK-NEXT:  1      6     1.00    *                  vmaskmovpd      (%rax), %xmm0, %xmm2
+# CHECK-NEXT:  2      6     2.00    *                  vmaskmovpd      (%rax), %ymm0, %ymm2
+# CHECK-NEXT:  1      6     2.00    *      *           vmaskmovpd      %xmm0, %xmm1, (%rax)
+# CHECK-NEXT:  2      6     2.00    *      *           vmaskmovpd      %ymm0, %ymm1, (%rax)
+# CHECK-NEXT:  1      6     1.00    *                  vmaskmovps      (%rax), %xmm0, %xmm2
+# CHECK-NEXT:  2      6     2.00    *                  vmaskmovps      (%rax), %ymm0, %ymm2
+# CHECK-NEXT:  1      6     2.00    *      *           vmaskmovps      %xmm0, %xmm1, (%rax)
+# CHECK-NEXT:  2      6     2.00    *      *           vmaskmovps      %ymm0, %ymm1, (%rax)
+# CHECK-NEXT:  1      2     1.00                       vmaxpd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vmaxpd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     2.00                       vmaxpd  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      7     2.00    *                  vmaxpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      2     1.00                       vmaxps  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vmaxps  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     2.00                       vmaxps  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      7     2.00    *                  vmaxps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      2     1.00                       vmaxsd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vmaxsd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vmaxss  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vmaxss  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vminpd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vminpd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     2.00                       vminpd  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      7     2.00    *                  vminpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      2     1.00                       vminps  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vminps  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     2.00                       vminps  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      7     2.00    *                  vminps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      2     1.00                       vminsd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vminsd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vminss  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vminss  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vmovapd %xmm0, %xmm2
+# CHECK-NEXT:  1      1     1.00           *           vmovapd %xmm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovapd (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vmovapd %ymm0, %ymm2
+# CHECK-NEXT:  1      1     1.00           *           vmovapd %ymm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovapd (%rax), %ymm2
+# CHECK-NEXT:  1      1     0.50                       vmovaps %xmm0, %xmm2
+# CHECK-NEXT:  1      1     1.00           *           vmovaps %xmm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovaps (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vmovaps %ymm0, %ymm2
+# CHECK-NEXT:  1      1     1.00           *           vmovaps %ymm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovaps (%rax), %ymm2
+# CHECK-NEXT:  1      1     0.50                       vmovd   %eax, %xmm2
+# CHECK-NEXT:  1      5     1.00    *                  vmovd   (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vmovd   %xmm0, %ecx
+# CHECK-NEXT:  1      1     1.00           *           vmovd   %xmm0, (%rax)
+# CHECK-NEXT:  1      1     0.50                       vmovddup        %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vmovddup        (%rax), %xmm2
+# CHECK-NEXT:  2      1     1.00                       vmovddup        %ymm0, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vmovddup        (%rax), %ymm2
+# CHECK-NEXT:  1      1     0.50                       vmovdqa %xmm0, %xmm2
+# CHECK-NEXT:  1      1     1.00           *           vmovdqa %xmm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovdqa (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vmovdqa %ymm0, %ymm2
+# CHECK-NEXT:  1      1     1.00           *           vmovdqa %ymm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovdqa (%rax), %ymm2
+# CHECK-NEXT:  1      1     0.50                       vmovdqu %xmm0, %xmm2
+# CHECK-NEXT:  1      1     1.00           *           vmovdqu %xmm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovdqu (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vmovdqu %ymm0, %ymm2
+# CHECK-NEXT:  1      1     1.00           *           vmovdqu %ymm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovdqu (%rax), %ymm2
+# CHECK-NEXT:  1      1     0.50                       vmovhlps        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vmovlhps        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00           *           vmovhpd %xmm0, (%rax)
+# CHECK-NEXT:  1      6     1.00    *                  vmovhpd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00           *           vmovhps %xmm0, (%rax)
+# CHECK-NEXT:  1      6     1.00    *                  vmovhps (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00           *           vmovlpd %xmm0, (%rax)
+# CHECK-NEXT:  1      6     1.00    *                  vmovlpd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00           *           vmovlps %xmm0, (%rax)
+# CHECK-NEXT:  1      6     1.00    *                  vmovlps (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                       vmovmskpd       %xmm0, %ecx
+# CHECK-NEXT:  1      3     1.00                       vmovmskpd       %ymm0, %ecx
+# CHECK-NEXT:  1      3     1.00                       vmovmskps       %xmm0, %ecx
+# CHECK-NEXT:  1      3     1.00                       vmovmskps       %ymm0, %ecx
+# CHECK-NEXT:  1      2     1.00           *           vmovntdq        %xmm0, (%rax)
+# CHECK-NEXT:  1      3     2.00           *           vmovntdq        %ymm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovntdqa       (%rax), %xmm2
+# CHECK-NEXT:  1      5     1.00    *                  vmovntdqa       (%rax), %ymm2
+# CHECK-NEXT:  1      3     1.00           *           vmovntpd        %xmm0, (%rax)
+# CHECK-NEXT:  1      3     2.00           *           vmovntpd        %ymm0, (%rax)
+# CHECK-NEXT:  1      3     1.00           *           vmovntps        %xmm0, (%rax)
+# CHECK-NEXT:  1      3     2.00           *           vmovntps        %ymm0, (%rax)
+# CHECK-NEXT:  1      1     0.50                       vmovq   %xmm0, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vmovq   %rax, %xmm2
+# CHECK-NEXT:  1      5     1.00    *                  vmovq   (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vmovq   %xmm0, %rcx
+# CHECK-NEXT:  1      1     1.00           *           vmovq   %xmm0, (%rax)
+# CHECK-NEXT:  1      1     0.50                       vmovsd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00           *           vmovsd  %xmm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovsd  (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vmovshdup       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vmovshdup       (%rax), %xmm2
+# CHECK-NEXT:  2      1     1.00                       vmovshdup       %ymm0, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vmovshdup       (%rax), %ymm2
+# CHECK-NEXT:  1      1     0.50                       vmovsldup       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vmovsldup       (%rax), %xmm2
+# CHECK-NEXT:  2      1     1.00                       vmovsldup       %ymm0, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vmovsldup       (%rax), %ymm2
+# CHECK-NEXT:  1      1     0.50                       vmovss  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00           *           vmovss  %xmm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovss  (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vmovupd %xmm0, %xmm2
+# CHECK-NEXT:  1      1     1.00           *           vmovupd %xmm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovupd (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vmovupd %ymm0, %ymm2
+# CHECK-NEXT:  1      1     1.00           *           vmovupd %ymm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovupd (%rax), %ymm2
+# CHECK-NEXT:  1      1     0.50                       vmovups %xmm0, %xmm2
+# CHECK-NEXT:  1      1     1.00           *           vmovups %xmm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovups (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vmovups %ymm0, %ymm2
+# CHECK-NEXT:  1      1     1.00           *           vmovups %ymm0, (%rax)
+# CHECK-NEXT:  1      5     1.00    *                  vmovups (%rax), %ymm2
+# CHECK-NEXT:  1      3     2.00                       vmpsadbw        $1, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     2.00    *                  vmpsadbw        $1, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      4     2.00                       vmulpd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      9     2.00    *                  vmulpd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      4     4.00                       vmulpd  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      9     4.00    *                  vmulpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      2     1.00                       vmulps  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vmulps  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      2     2.00                       vmulps  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      7     2.00    *                  vmulps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      4     2.00                       vmulsd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      9     2.00    *                  vmulsd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vmulss  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vmulss  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vorpd   %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vorpd   (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vorpd   %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vorpd   (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vorps   %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vorps   (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vorps   %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vorps   (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vpabsb  %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpabsb  (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpabsd  %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpabsd  (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpabsw  %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpabsw  (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpackssdw       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpackssdw       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpacksswb       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpacksswb       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpackusdw       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpackusdw       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpackuswb       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpackuswb       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpaddb  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpaddb  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpaddd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpaddd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpaddq  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpaddq  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpaddsb %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpaddsb (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpaddsw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpaddsw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpaddusb        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpaddusb        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpaddusw        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpaddusw        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpaddw  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpaddw  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpalignr        $1, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpalignr        $1, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpand   %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpand   (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpandn  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpandn  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpavgb  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpavgb  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpavgw  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpavgw  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  3      2     2.00                       vpblendvb       %xmm3, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  3      7     2.00    *                  vpblendvb       %xmm3, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpblendw        $11, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpblendw        $11, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vpclmulqdq      $11, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vpclmulqdq      $11, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpcmpeqb        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpcmpeqb        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpcmpeqd        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpcmpeqd        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpcmpeqq        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpcmpeqq        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpcmpeqw        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpcmpeqw        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpcmpgtb        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpcmpgtb        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpcmpgtd        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpcmpgtd        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpcmpgtq        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpcmpgtq        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpcmpgtw        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpcmpgtw        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vperm2f128      $1, %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      6     1.00    *                  vperm2f128      $1, (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vpermilpd       $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpermilpd       $1, (%rax), %xmm2
+# CHECK-NEXT:  3      2     2.00                       vpermilpd       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  3      7     2.00    *                  vpermilpd       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vpermilpd       $1, %ymm0, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vpermilpd       $1, (%rax), %ymm2
+# CHECK-NEXT:  6      3     3.00                       vpermilpd       %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  6      8     3.00    *                  vpermilpd       (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vpermilps       $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpermilps       $1, (%rax), %xmm2
+# CHECK-NEXT:  3      2     2.00                       vpermilps       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  3      7     2.00    *                  vpermilps       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vpermilps       $1, %ymm0, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vpermilps       $1, (%rax), %ymm2
+# CHECK-NEXT:  6      3     3.00                       vpermilps       %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  6      8     3.00    *                  vpermilps       (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vpextrb $1, %xmm0, %ecx
+# CHECK-NEXT:  2      6     1.00           *           vpextrb $1, %xmm0, (%rax)
+# CHECK-NEXT:  1      1     0.50                       vpextrd $1, %xmm0, %ecx
+# CHECK-NEXT:  2      6     1.00           *           vpextrd $1, %xmm0, (%rax)
+# CHECK-NEXT:  1      1     0.50                       vpextrq $1, %xmm0, %rcx
+# CHECK-NEXT:  2      6     1.00           *           vpextrq $1, %xmm0, (%rax)
+# CHECK-NEXT:  1      1     0.50                       vpextrw $1, %xmm0, %ecx
+# CHECK-NEXT:  2      6     1.00           *           vpextrw $1, %xmm0, (%rax)
+# CHECK-NEXT:  1      1     0.50                       vphaddd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vphaddd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vphaddsw        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vphaddsw        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vphaddw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vphaddw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vphminposuw     %xmm0, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vphminposuw     (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vphsubd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vphsubd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vphsubsw        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vphsubsw        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vphsubw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vphsubw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpinsrb $1, %eax, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpinsrb $1, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpinsrd $1, %eax, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpinsrd $1, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpinsrq $1, %rax, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpinsrq $1, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpinsrw $1, %eax, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpinsrw $1, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vpmaddubsw      %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vpmaddubsw      (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vpmaddwd        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vpmaddwd        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmaxsb %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmaxsb (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmaxsd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmaxsd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmaxsw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmaxsw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmaxub %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmaxub (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmaxud %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmaxud (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmaxuw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmaxuw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpminsb %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpminsb (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpminsd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpminsd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpminsw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpminsw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpminub %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpminub (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpminud %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpminud (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpminuw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpminuw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                       vpmovmskb       %xmm0, %ecx
+# CHECK-NEXT:  1      1     0.50                       vpmovsxbd       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmovsxbd       (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmovsxbq       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmovsxbq       (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmovsxbw       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmovsxbw       (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmovsxdq       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmovsxdq       (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmovsxwd       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmovsxwd       (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmovsxwq       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmovsxwq       (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmovzxbd       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmovzxbd       (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmovzxbq       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmovzxbq       (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmovzxbw       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmovzxbw       (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmovzxdq       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmovzxdq       (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmovzxwd       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmovzxwd       (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpmovzxwq       %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpmovzxwq       (%rax), %xmm2
+# CHECK-NEXT:  1      2     1.00                       vpmuldq %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vpmuldq (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vpmulhrsw       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vpmulhrsw       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vpmulhuw        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vpmulhuw        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vpmulhw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vpmulhw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vpmulld %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vpmulld (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vpmullw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vpmullw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vpmuludq        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vpmuludq        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpor    %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpor    (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsadbw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsadbw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  3      2     2.00                       vpshufb %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  3      7     2.00    *                  vpshufb (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpshufd $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpshufd $1, (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpshufhw        $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpshufhw        $1, (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpshuflw        $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpshuflw        $1, (%rax), %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsignb %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsignb (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsignd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsignd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsignw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsignw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpslld  $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpslld  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpslld  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpslldq $1, %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsllq  $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsllq  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsllq  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsllw  $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsllw  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsllw  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsrad  $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsrad  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsrad  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsraw  $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsraw  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsraw  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsrld  $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsrld  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsrld  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsrldq $1, %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsrlq  $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsrlq  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsrlq  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsrlw  $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsrlw  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsrlw  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsubb  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsubb  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsubd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsubd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsubq  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsubq  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsubsb %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsubsb (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsubsw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsubsw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsubusb        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsubusb        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsubusw        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsubusw        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpsubw  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpsubw  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                       vptest  %xmm0, %xmm1
+# CHECK-NEXT:  1      8     1.00    *                  vptest  (%rax), %xmm1
+# CHECK-NEXT:  3      4     2.00                       vptest  %ymm0, %ymm1
+# CHECK-NEXT:  3      9     2.00    *                  vptest  (%rax), %ymm1
+# CHECK-NEXT:  1      1     0.50                       vpunpckhbw      %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpunpckhbw      (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpunpckhdq      %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpunpckhdq      (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpunpckhqdq     %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpunpckhqdq     (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpunpckhwd      %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpunpckhwd      (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpunpcklbw      %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpunpcklbw      (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpunpckldq      %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpunpckldq      (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpunpcklqdq     %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpunpcklqdq     (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpunpcklwd      %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpunpcklwd      (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vpxor   %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vpxor   (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vrcpps  %xmm0, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vrcpps  (%rax), %xmm2
+# CHECK-NEXT:  2      2     2.00                       vrcpps  %ymm0, %ymm2
+# CHECK-NEXT:  2      7     2.00    *                  vrcpps  (%rax), %ymm2
+# CHECK-NEXT:  1      2     1.00                       vrcpss  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vrcpss  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                       vroundpd        $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vroundpd        $1, (%rax), %xmm2
+# CHECK-NEXT:  2      3     2.00                       vroundpd        $1, %ymm0, %ymm2
+# CHECK-NEXT:  2      8     2.00    *                  vroundpd        $1, (%rax), %ymm2
+# CHECK-NEXT:  1      3     1.00                       vroundps        $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vroundps        $1, (%rax), %xmm2
+# CHECK-NEXT:  2      3     2.00                       vroundps        $1, %ymm0, %ymm2
+# CHECK-NEXT:  2      8     2.00    *                  vroundps        $1, (%rax), %ymm2
+# CHECK-NEXT:  1      3     1.00                       vroundsd        $1, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vroundsd        $1, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                       vroundss        $1, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vroundss        $1, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      2     1.00                       vrsqrtps        %xmm0, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vrsqrtps        (%rax), %xmm2
+# CHECK-NEXT:  2      2     2.00                       vrsqrtps        %ymm0, %ymm2
+# CHECK-NEXT:  2      7     2.00    *                  vrsqrtps        (%rax), %ymm2
+# CHECK-NEXT:  1      2     1.00                       vrsqrtss        %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      7     1.00    *                  vrsqrtss        (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     0.50                       vshufpd $1, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vshufpd $1, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vshufpd $1, %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vshufpd $1, (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vshufps $1, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vshufps $1, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vshufps $1, %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vshufps $1, (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      27    27.00                      vsqrtpd %xmm0, %xmm2
+# CHECK-NEXT:  1      32    27.00   *                  vsqrtpd (%rax), %xmm2
+# CHECK-NEXT:  1      54    54.00                      vsqrtpd %ymm0, %ymm2
+# CHECK-NEXT:  1      59    54.00   *                  vsqrtpd (%rax), %ymm2
+# CHECK-NEXT:  1      21    21.00                      vsqrtps %xmm0, %xmm2
+# CHECK-NEXT:  1      26    21.00   *                  vsqrtps (%rax), %xmm2
+# CHECK-NEXT:  2      42    42.00                      vsqrtps %ymm0, %ymm2
+# CHECK-NEXT:  2      47    42.00   *                  vsqrtps (%rax), %ymm2
+# CHECK-NEXT:  1      27    27.00                      vsqrtsd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      32    27.00   *                  vsqrtsd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      21    21.00                      vsqrtss %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      26    21.00   *                  vsqrtss (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00    *      *      *    vstmxcsr        (%rax)
+# CHECK-NEXT:  1      3     1.00                       vsubpd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vsubpd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      3     2.00                       vsubpd  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      8     2.00    *                  vsubpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     1.00                       vsubps  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vsubps  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      3     2.00                       vsubps  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      8     2.00    *                  vsubps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     1.00                       vsubsd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vsubsd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                       vsubss  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                  vsubss  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                       vtestpd %xmm0, %xmm1
+# CHECK-NEXT:  1      8     1.00    *                  vtestpd (%rax), %xmm1
+# CHECK-NEXT:  3      4     2.00                       vtestpd %ymm0, %ymm1
+# CHECK-NEXT:  3      9     2.00    *                  vtestpd (%rax), %ymm1
+# CHECK-NEXT:  1      3     1.00                       vtestps %xmm0, %xmm1
+# CHECK-NEXT:  1      8     1.00    *                  vtestps (%rax), %xmm1
+# CHECK-NEXT:  3      4     2.00                       vtestps %ymm0, %ymm1
+# CHECK-NEXT:  3      9     2.00    *                  vtestps (%rax), %ymm1
+# CHECK-NEXT:  1      3     1.00                       vucomisd        %xmm0, %xmm1
+# CHECK-NEXT:  1      8     1.00    *                  vucomisd        (%rax), %xmm1
+# CHECK-NEXT:  1      3     1.00                       vucomiss        %xmm0, %xmm1
+# CHECK-NEXT:  1      8     1.00    *                  vucomiss        (%rax), %xmm1
+# CHECK-NEXT:  1      1     0.50                       vunpckhpd       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vunpckhpd       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vunpckhpd       %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vunpckhpd       (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vunpckhps       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vunpckhps       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vunpckhps       %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vunpckhps       (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vunpcklpd       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vunpcklpd       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vunpcklpd       %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vunpcklpd       (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vunpcklps       %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vunpcklps       (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vunpcklps       %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vunpcklps       (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vxorpd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vxorpd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vxorpd  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vxorpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     0.50                       vxorps  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      6     1.00    *                  vxorps  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  2      1     1.00                       vxorps  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      6     2.00    *                  vxorps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  73     90     -      *      *      *    vzeroall
+# CHECK-NEXT:  37     46     -      *      *      *    vzeroupper
+
+
 # CHECK:      Resources:
 # CHECK-NEXT: [0] - JALU0
 # CHECK-NEXT: [1] - JALU1