[X86] Add proper BTVER2 sched support for MOV

[avt77]

Summary

The current sched model uses JAny for mov instruction but it's wrong and should be changed on JALU01.

[spatel]

This isn't correct for moves between scalar and vector for example, but it's less wrong in those cases, so LGTM as a partial fix.

lib/Target/X86/X86ScheduleBtVer2.td

 
 ////////////////////////////////////////////////////////////////////////////////
 // Loads, stores, and moves, not folded with other operations.
 // FIXME: Split x86 and SSE load/store/moves
 ////////////////////////////////////////////////////////////////////////////////
 
 def : WriteRes<WriteLoad,  [JLAGU]> { let Latency = 5; }
 def : WriteRes<WriteStore, [JSAGU]>;
-def : WriteRes<WriteMove,  [JAny]>;
+def : WriteRes<WriteMove,  [JALU01]>;
 
 ////////////////////////////////////////////////////////////////////////////////
 // Idioms that clear a register, like xorps %xmm0, %xmm0.
 // These can often bypass execution ports completely.
 ////////////////////////////////////////////////////////////////////////////////
 
 def : WriteRes<WriteZero,  []>;
 

test/CodeGen/X86/mmx-schedule.ll

 ; SKX-NEXT:    movd %mm1, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_cvtpd2pi:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    cvtpd2pi (%rdi), %mm1 # sched: [8:1.00]
 ; BTVER2-NEXT:    cvtpd2pi %xmm0, %mm0 # sched: [3:1.00]
 ; BTVER2-NEXT:    por %mm0, %mm1 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm1, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm1, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_cvtpd2pi:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    cvtpd2pi (%rdi), %mm1 # sched: [12:1.00]
 ; ZNVER1-NEXT:    cvtpd2pi %xmm0, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    por %mm0, %mm1 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm1, %rax # sched: [2:1.00]

 ; SKX-NEXT:    movd %mm1, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_cvtps2pi:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    cvtps2pi (%rdi), %mm1 # sched: [8:1.00]
 ; BTVER2-NEXT:    cvtps2pi %xmm0, %mm0 # sched: [3:1.00]
 ; BTVER2-NEXT:    por %mm0, %mm1 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm1, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm1, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_cvtps2pi:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    cvtps2pi (%rdi), %mm1 # sched: [12:1.00]
 ; ZNVER1-NEXT:    cvtps2pi %xmm0, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    por %mm0, %mm1 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm1, %rax # sched: [2:1.00]

 ; SKX-NEXT:    movd %mm1, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_cvttpd2pi:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    cvttpd2pi (%rdi), %mm1 # sched: [8:1.00]
 ; BTVER2-NEXT:    cvttpd2pi %xmm0, %mm0 # sched: [3:1.00]
 ; BTVER2-NEXT:    por %mm0, %mm1 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm1, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm1, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_cvttpd2pi:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    cvttpd2pi (%rdi), %mm1 # sched: [12:1.00]
 ; ZNVER1-NEXT:    cvttpd2pi %xmm0, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    por %mm0, %mm1 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm1, %rax # sched: [2:1.00]

 ; SKX-NEXT:    movd %mm1, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_cvttps2pi:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    cvttps2pi (%rdi), %mm1 # sched: [8:1.00]
 ; BTVER2-NEXT:    cvttps2pi %xmm0, %mm0 # sched: [3:1.00]
 ; BTVER2-NEXT:    por %mm0, %mm1 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm1, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm1, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_cvttps2pi:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    cvttps2pi (%rdi), %mm1 # sched: [12:1.00]
 ; ZNVER1-NEXT:    cvttps2pi %xmm0, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    por %mm0, %mm1 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm1, %rax # sched: [2:1.00]

 ; SKX-NEXT:    paddd %mm1, %mm0 # sched: [1:0.50]
 ; SKX-NEXT:    movd %mm1, %ecx # sched: [2:1.00]
 ; SKX-NEXT:    movd %mm0, %eax # sched: [2:1.00]
 ; SKX-NEXT:    movl %ecx, (%rsi) # sched: [1:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_movd:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vmovd %edi, %xmm0 # sched: [1:0.17]
+; BTVER2-NEXT:    vmovd %edi, %xmm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    vmovq %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
 ; BTVER2-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:1.00]
 ; BTVER2-NEXT:    movq -{{[0-9]+}}(%rsp), %mm1 # sched: [5:1.00]
 ; BTVER2-NEXT:    vmovlps %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
 ; BTVER2-NEXT:    paddd -{{[0-9]+}}(%rsp), %mm1 # sched: [6:1.00]
 ; BTVER2-NEXT:    paddd %mm1, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm1, %ecx # sched: [1:0.17]
-; BTVER2-NEXT:    movd %mm0, %eax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm1, %ecx # sched: [1:0.50]
+; BTVER2-NEXT:    movd %mm0, %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    movl %ecx, (%rsi) # sched: [1:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_movd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vmovd %edi, %xmm0 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vmovq %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:0.50]
 ; ZNVER1-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [8:0.50]

 ; SKX:       # BB#0:
 ; SKX-NEXT:    movdq2q %xmm0, %mm0 # sched: [2:1.00]
 ; SKX-NEXT:    paddd %mm0, %mm0 # sched: [1:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_movdq2q:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    movdq2q %xmm0, %mm0 # sched: [1:0.17]
+; BTVER2-NEXT:    movdq2q %xmm0, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    paddd %mm0, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_movdq2q:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    movdq2q %xmm0, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    paddd %mm0, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ;
 ; SKX-LABEL: test_movq2dq:
 ; SKX:       # BB#0:
 ; SKX-NEXT:    movq2dq %mm0, %xmm0 # sched: [2:2.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_movq2dq:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    movq2dq %mm0, %xmm0 # sched: [1:0.17]
+; BTVER2-NEXT:    movq2dq %mm0, %xmm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_movq2dq:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    movq2dq %mm0, %xmm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = bitcast x86_mmx %a0 to i64
   %2 = insertelement <2 x i64> undef, i64 %1, i32 0

 ; SKX-NEXT:    pabsb %mm0, %mm0 # sched: [1:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pabsb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pabsb (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    pabsb %mm0, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pabsb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pabsb (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    pabsb %mm0, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pabsd %mm0, %mm0 # sched: [1:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pabsd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pabsd (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    pabsd %mm0, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pabsd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pabsd (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    pabsd %mm0, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pabsw %mm0, %mm0 # sched: [1:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pabsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pabsw (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    pabsw %mm0, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pabsw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pabsw (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    pabsw %mm0, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    packssdw (%rdi), %mm0 # sched: [7:2.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_packssdw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    packssdw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    packssdw (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_packssdw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    packssdw %mm1, %mm0 # sched: [1:0.50]
 ; ZNVER1-NEXT:    packssdw (%rdi), %mm0 # sched: [1:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    packsswb (%rdi), %mm0 # sched: [7:2.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_packsswb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    packsswb %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    packsswb (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_packsswb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    packsswb %mm1, %mm0 # sched: [1:0.50]
 ; ZNVER1-NEXT:    packsswb (%rdi), %mm0 # sched: [1:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    packuswb (%rdi), %mm0 # sched: [7:2.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_packuswb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    packuswb %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    packuswb (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_packuswb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    packuswb %mm1, %mm0 # sched: [1:0.50]
 ; ZNVER1-NEXT:    packuswb (%rdi), %mm0 # sched: [1:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    paddb (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_paddb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    paddb %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    paddb (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_paddb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    paddb %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    paddb (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    paddd (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_paddd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    paddd %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    paddd (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_paddd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    paddd %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    paddd (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    paddq (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_paddq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    paddq %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    paddq (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_paddq:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    paddq %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    paddq (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    paddsb (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_paddsb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    paddsb %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    paddsb (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_paddsb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    paddsb %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    paddsb (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    paddsw (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_paddsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    paddsw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    paddsw (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_paddsw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    paddsw %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    paddsw (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    paddusb (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_paddusb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    paddusb %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    paddusb (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_paddusb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    paddusb %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    paddusb (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    paddusw (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_paddusw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    paddusw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    paddusw (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_paddusw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    paddusw %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    paddusw (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    paddw (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_paddw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    paddw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    paddw (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_paddw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    paddw %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    paddw (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    palignr $1, (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_palignr:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    palignr $1, %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    palignr $1, (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_palignr:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    palignr $1, %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    palignr $1, (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pand (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pand:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pand %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    pand (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pand:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pand %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    pand (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pandn (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pandn:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pandn %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    pandn (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pandn:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pandn %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    pandn (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pavgb (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pavgb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pavgb %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    pavgb (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pavgb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pavgb %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    pavgb (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pavgw (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pavgw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pavgw %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    pavgw (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pavgw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pavgw %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    pavgw (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pcmpeqb (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pcmpeqb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pcmpeqb %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    pcmpeqb (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pcmpeqb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pcmpeqb %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    pcmpeqb (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pcmpeqd (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pcmpeqd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pcmpeqd %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    pcmpeqd (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pcmpeqd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pcmpeqd %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    pcmpeqd (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pcmpeqw (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pcmpeqw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pcmpeqw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    pcmpeqw (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pcmpeqw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pcmpeqw %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    pcmpeqw (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pcmpgtb (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pcmpgtb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pcmpgtb %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    pcmpgtb (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pcmpgtb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pcmpgtb %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    pcmpgtb (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pcmpgtd (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pcmpgtd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pcmpgtd %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    pcmpgtd (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pcmpgtd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pcmpgtd %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    pcmpgtd (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pcmpgtw (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pcmpgtw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pcmpgtw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    pcmpgtw (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pcmpgtw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pcmpgtw %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    pcmpgtw (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    phaddd (%rdi), %mm0 # sched: [8:2.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_phaddd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    phaddd %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    phaddd (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_phaddd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    phaddd %mm1, %mm0 # sched: [100:?]
 ; ZNVER1-NEXT:    phaddd (%rdi), %mm0 # sched: [100:?]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    phaddsw (%rdi), %mm0 # sched: [8:2.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_phaddsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    phaddsw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    phaddsw (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_phaddsw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    phaddsw %mm1, %mm0 # sched: [100:?]
 ; ZNVER1-NEXT:    phaddsw (%rdi), %mm0 # sched: [100:?]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    phaddw (%rdi), %mm0 # sched: [8:2.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_phaddw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    phaddw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    phaddw (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_phaddw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    phaddw %mm1, %mm0 # sched: [100:?]
 ; ZNVER1-NEXT:    phaddw (%rdi), %mm0 # sched: [100:?]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    phsubd (%rdi), %mm0 # sched: [8:2.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_phsubd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    phsubd %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    phsubd (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_phsubd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    phsubd %mm1, %mm0 # sched: [100:?]
 ; ZNVER1-NEXT:    phsubd (%rdi), %mm0 # sched: [100:?]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    phsubsw (%rdi), %mm0 # sched: [8:2.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_phsubsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    phsubsw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    phsubsw (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_phsubsw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    phsubsw %mm1, %mm0 # sched: [100:?]
 ; ZNVER1-NEXT:    phsubsw (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    phsubw (%rdi), %mm0 # sched: [8:2.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_phsubw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    phsubw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    phsubw (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_phsubw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    phsubw %mm1, %mm0 # sched: [100:?]
 ; ZNVER1-NEXT:    phsubw (%rdi), %mm0 # sched: [100:?]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pinsrw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    movswl (%rsi), %eax # sched: [4:1.00]
 ; BTVER2-NEXT:    pinsrw $0, %edi, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    pinsrw $1, %eax, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pinsrw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    movswl (%rsi), %eax # sched: [8:0.50]
 ; ZNVER1-NEXT:    pinsrw $0, %edi, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    pinsrw $1, %eax, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]

 ; SKX-NEXT:    pmaddwd (%rdi), %mm0 # sched: [9:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pmaddwd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pmaddwd %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    pmaddwd (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pmaddwd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pmaddwd %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    pmaddwd (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pmaddubsw (%rdi), %mm0 # sched: [9:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pmaddubsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pmaddubsw %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    pmaddubsw (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pmaddubsw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pmaddubsw %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    pmaddubsw (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pmaxsw (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pmaxsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pmaxsw %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    pmaxsw (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pmaxsw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pmaxsw %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    pmaxsw (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pmaxub (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pmaxub:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pmaxub %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    pmaxub (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pmaxub:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pmaxub %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    pmaxub (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pminsw (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pminsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pminsw %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    pminsw (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pminsw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pminsw %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    pminsw (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pminub (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pminub:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pminub %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    pminub (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pminub:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pminub %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    pminub (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pmulhrsw (%rdi), %mm0 # sched: [9:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pmulhrsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pmulhrsw %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    pmulhrsw (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pmulhrsw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pmulhrsw %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    pmulhrsw (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pmulhw (%rdi), %mm0 # sched: [9:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pmulhw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pmulhw %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    pmulhw (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pmulhw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pmulhw %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    pmulhw (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pmulhuw (%rdi), %mm0 # sched: [9:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pmulhuw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pmulhuw %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    pmulhuw (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pmulhuw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pmulhuw %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    pmulhuw (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pmullw (%rdi), %mm0 # sched: [9:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pmullw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pmullw %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    pmullw (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pmullw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pmullw %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    pmullw (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pmuludq (%rdi), %mm0 # sched: [9:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pmuludq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pmuludq %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    pmuludq (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pmuludq:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pmuludq %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    pmuludq (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    por (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_por:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    por %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    por (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_por:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    por %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    por (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    psadbw (%rdi), %mm0 # sched: [8:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psadbw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psadbw %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    psadbw (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psadbw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psadbw %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    psadbw (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pshufb (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pshufb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pshufb %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    pshufb (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pshufb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pshufb %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    pshufb (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pshufw $0, %mm0, %mm0 # mm0 = mm0[0,0,0,0] sched: [1:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pshufw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pshufw $0, (%rdi), %mm0 # mm0 = mem[0,0,0,0] sched: [6:1.00]
 ; BTVER2-NEXT:    pshufw $0, %mm0, %mm0 # mm0 = mm0[0,0,0,0] sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pshufw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pshufw $0, (%rdi), %mm0 # mm0 = mem[0,0,0,0] sched: [8:0.50]
 ; ZNVER1-NEXT:    pshufw $0, %mm0, %mm0 # mm0 = mm0[0,0,0,0] sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    psignb (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psignb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psignb %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    psignb (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psignb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psignb %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    psignb (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    psignd (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psignd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psignd %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    psignd (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psignd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psignd %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    psignd (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    psignw (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psignw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psignw %mm1, %mm0 # sched: [2:1.00]
 ; BTVER2-NEXT:    psignw (%rdi), %mm0 # sched: [7:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psignw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psignw %mm1, %mm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    psignw (%rdi), %mm0 # sched: [11:1.00]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pslld:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pslld %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    pslld (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    pslld $7, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pslld:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pslld %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    pslld (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    pslld $7, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]

 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psllq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psllq %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psllq (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    psllq $7, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psllq:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psllq %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psllq (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    psllq $7, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]

 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psllw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psllw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psllw (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    psllw $7, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psllw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psllw %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psllw (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    psllw $7, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]

 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psrad:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psrad %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psrad (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    psrad $7, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psrad:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psrad %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psrad (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    psrad $7, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]

 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psraw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psraw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psraw (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    psraw $7, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psraw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psraw %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psraw (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    psraw $7, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]

 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psrld:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psrld %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psrld (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    psrld $7, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psrld:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psrld %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psrld (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    psrld $7, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]

 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psrlq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psrlq %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psrlq (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    psrlq $7, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psrlq:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psrlq %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psrlq (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    psrlq $7, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]

 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psrlw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psrlw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psrlw (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    psrlw $7, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psrlw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psrlw %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psrlw (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    psrlw $7, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]

 ; SKX-NEXT:    psubb (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psubb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psubb %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psubb (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psubb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psubb %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psubb (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    psubd (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psubd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psubd %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psubd (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psubd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psubd %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psubd (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    psubq (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psubq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psubq %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psubq (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psubq:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psubq %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psubq (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    psubsb (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psubsb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psubsb %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psubsb (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psubsb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psubsb %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psubsb (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    psubsw (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psubsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psubsw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psubsw (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psubsw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psubsw %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psubsw (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    psubusb (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psubusb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psubusb %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psubusb (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psubusb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psubusb %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psubusb (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    psubusw (%rdi), %mm0 # sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psubusw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psubusw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psubusw (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psubusw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psubusw %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psubusw (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    psubw (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_psubw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    psubw %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    psubw (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_psubw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    psubw %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    psubw (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    punpckhbw (%rdi), %mm0 # mm0 = mm0[4],mem[4],mm0[5],mem[5],mm0[6],mem[6],mm0[7],mem[7] sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_punpckhbw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    punpckhbw %mm1, %mm0 # mm0 = mm0[4],mm1[4],mm0[5],mm1[5],mm0[6],mm1[6],mm0[7],mm1[7] sched: [1:0.50]
 ; BTVER2-NEXT:    punpckhbw (%rdi), %mm0 # mm0 = mm0[4],mem[4],mm0[5],mem[5],mm0[6],mem[6],mm0[7],mem[7] sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_punpckhbw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    punpckhbw %mm1, %mm0 # mm0 = mm0[4],mm1[4],mm0[5],mm1[5],mm0[6],mm1[6],mm0[7],mm1[7] sched: [1:0.25]
 ; ZNVER1-NEXT:    punpckhbw (%rdi), %mm0 # mm0 = mm0[4],mem[4],mm0[5],mem[5],mm0[6],mem[6],mm0[7],mem[7] sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    punpckhdq (%rdi), %mm0 # mm0 = mm0[1],mem[1] sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_punpckhdq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    punpckhdq %mm1, %mm0 # mm0 = mm0[1],mm1[1] sched: [1:0.50]
 ; BTVER2-NEXT:    punpckhdq (%rdi), %mm0 # mm0 = mm0[1],mem[1] sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_punpckhdq:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    punpckhdq %mm1, %mm0 # mm0 = mm0[1],mm1[1] sched: [1:0.25]
 ; ZNVER1-NEXT:    punpckhdq (%rdi), %mm0 # mm0 = mm0[1],mem[1] sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    punpckhwd (%rdi), %mm0 # mm0 = mm0[2],mem[2],mm0[3],mem[3] sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_punpckhwd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    punpckhwd %mm1, %mm0 # mm0 = mm0[2],mm1[2],mm0[3],mm1[3] sched: [1:0.50]
 ; BTVER2-NEXT:    punpckhwd (%rdi), %mm0 # mm0 = mm0[2],mem[2],mm0[3],mem[3] sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_punpckhwd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    punpckhwd %mm1, %mm0 # mm0 = mm0[2],mm1[2],mm0[3],mm1[3] sched: [1:0.25]
 ; ZNVER1-NEXT:    punpckhwd (%rdi), %mm0 # mm0 = mm0[2],mem[2],mm0[3],mem[3] sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    punpcklbw (%rdi), %mm0 # mm0 = mm0[0],mem[0],mm0[1],mem[1],mm0[2],mem[2],mm0[3],mem[3] sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_punpcklbw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    punpcklbw %mm1, %mm0 # mm0 = mm0[0],mm1[0],mm0[1],mm1[1],mm0[2],mm1[2],mm0[3],mm1[3] sched: [1:0.50]
 ; BTVER2-NEXT:    punpcklbw (%rdi), %mm0 # mm0 = mm0[0],mem[0],mm0[1],mem[1],mm0[2],mem[2],mm0[3],mem[3] sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_punpcklbw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    punpcklbw %mm1, %mm0 # mm0 = mm0[0],mm1[0],mm0[1],mm1[1],mm0[2],mm1[2],mm0[3],mm1[3] sched: [1:0.25]
 ; ZNVER1-NEXT:    punpcklbw (%rdi), %mm0 # mm0 = mm0[0],mem[0],mm0[1],mem[1],mm0[2],mem[2],mm0[3],mem[3] sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    punpckldq (%rdi), %mm0 # mm0 = mm0[0],mem[0] sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_punpckldq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    punpckldq %mm1, %mm0 # mm0 = mm0[0],mm1[0] sched: [1:0.50]
 ; BTVER2-NEXT:    punpckldq (%rdi), %mm0 # mm0 = mm0[0],mem[0] sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_punpckldq:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    punpckldq %mm1, %mm0 # mm0 = mm0[0],mm1[0] sched: [1:0.25]
 ; ZNVER1-NEXT:    punpckldq (%rdi), %mm0 # mm0 = mm0[0],mem[0] sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    punpcklwd (%rdi), %mm0 # mm0 = mm0[0],mem[0],mm0[1],mem[1] sched: [6:1.00]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_punpcklwd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    punpcklwd %mm1, %mm0 # mm0 = mm0[0],mm1[0],mm0[1],mm1[1] sched: [1:0.50]
 ; BTVER2-NEXT:    punpcklwd (%rdi), %mm0 # mm0 = mm0[0],mem[0],mm0[1],mem[1] sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_punpcklwd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    punpcklwd %mm1, %mm0 # mm0 = mm0[0],mm1[0],mm0[1],mm1[1] sched: [1:0.25]
 ; ZNVER1-NEXT:    punpcklwd (%rdi), %mm0 # mm0 = mm0[0],mem[0],mm0[1],mem[1] sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    pxor (%rdi), %mm0 # sched: [6:0.50]
 ; SKX-NEXT:    movd %mm0, %rax # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pxor:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    pxor %mm1, %mm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    pxor (%rdi), %mm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movd %mm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pxor:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pxor %mm1, %mm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    pxor (%rdi), %mm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    movd %mm0, %rax # sched: [2:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call x86_mmx @llvm.x86.mmx.pxor(x86_mmx %a0, x86_mmx %a1)
   %2 = load x86_mmx, x86_mmx *%a2, align 8
   %3 = call x86_mmx @llvm.x86.mmx.pxor(x86_mmx %1, x86_mmx %2)
   %4 = bitcast x86_mmx %3 to i64
   ret i64 %4
 }
 declare x86_mmx @llvm.x86.mmx.pxor(x86_mmx, x86_mmx) nounwind readnone

test/CodeGen/X86/mul-constant-i32.ll

 ;
 ; X64-HSW-LABEL: test_mul_by_1:
 ; X64-HSW:       # BB#0:
 ; X64-HSW-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; X64-HSW-NEXT:    retq # sched: [2:1.00]
 ;
 ; X64-JAG-LABEL: test_mul_by_1:
 ; X64-JAG:       # BB#0:
-; X64-JAG-NEXT:    movl %edi, %eax # sched: [1:0.17]
+; X64-JAG-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; X64-JAG-NEXT:    retq # sched: [4:1.00]
 ;
 ; X86-NOOPT-LABEL: test_mul_by_1:
 ; X86-NOOPT:       # BB#0:
 ; X86-NOOPT-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NOOPT-NEXT:    retl
 ;
 ; HSW-NOOPT-LABEL: test_mul_by_1:
 ; HSW-NOOPT:       # BB#0:
 ; HSW-NOOPT-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; HSW-NOOPT-NEXT:    retq # sched: [2:1.00]
 ;
 ; JAG-NOOPT-LABEL: test_mul_by_1:
 ; JAG-NOOPT:       # BB#0:
-; JAG-NOOPT-NEXT:    movl %edi, %eax # sched: [1:0.17]
+; JAG-NOOPT-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; JAG-NOOPT-NEXT:    retq # sched: [4:1.00]
 ;
 ; X64-SLM-LABEL: test_mul_by_1:
 ; X64-SLM:       # BB#0:
 ; X64-SLM-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; X64-SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SLM-NOOPT-LABEL: test_mul_by_1:

 ; X64-HSW:       # BB#0:
 ; X64-HSW-NEXT:    shll $4, %edi # sched: [1:0.50]
 ; X64-HSW-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; X64-HSW-NEXT:    retq # sched: [2:1.00]
 ;
 ; X64-JAG-LABEL: test_mul_by_16:
 ; X64-JAG:       # BB#0:
 ; X64-JAG-NEXT:    shll $4, %edi # sched: [1:0.50]
-; X64-JAG-NEXT:    movl %edi, %eax # sched: [1:0.17]
+; X64-JAG-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; X64-JAG-NEXT:    retq # sched: [4:1.00]
 ;
 ; X86-NOOPT-LABEL: test_mul_by_16:
 ; X86-NOOPT:       # BB#0:
 ; X86-NOOPT-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NOOPT-NEXT:    shll $4, %eax
 ; X86-NOOPT-NEXT:    retl
 ;
 ; HSW-NOOPT-LABEL: test_mul_by_16:
 ; HSW-NOOPT:       # BB#0:
 ; HSW-NOOPT-NEXT:    shll $4, %edi # sched: [1:0.50]
 ; HSW-NOOPT-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; HSW-NOOPT-NEXT:    retq # sched: [2:1.00]
 ;
 ; JAG-NOOPT-LABEL: test_mul_by_16:
 ; JAG-NOOPT:       # BB#0:
 ; JAG-NOOPT-NEXT:    shll $4, %edi # sched: [1:0.50]
-; JAG-NOOPT-NEXT:    movl %edi, %eax # sched: [1:0.17]
+; JAG-NOOPT-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; JAG-NOOPT-NEXT:    retq # sched: [4:1.00]
 ;
 ; X64-SLM-LABEL: test_mul_by_16:
 ; X64-SLM:       # BB#0:
 ; X64-SLM-NEXT:    shll $4, %edi # sched: [1:1.00]
 ; X64-SLM-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; X64-SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; X64-HSW-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; X64-HSW-NEXT:    shll $4, %eax # sched: [1:0.50]
 ; X64-HSW-NEXT:    leal (%rax,%rdi), %eax # sched: [1:0.50]
 ; X64-HSW-NEXT:    retq # sched: [2:1.00]
 ;
 ; X64-JAG-LABEL: test_mul_by_17:
 ; X64-JAG:       # BB#0:
 ; X64-JAG-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
-; X64-JAG-NEXT:    movl %edi, %eax # sched: [1:0.17]
+; X64-JAG-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; X64-JAG-NEXT:    shll $4, %eax # sched: [1:0.50]
 ; X64-JAG-NEXT:    leal (%rax,%rdi), %eax # sched: [1:0.50]
 ; X64-JAG-NEXT:    retq # sched: [4:1.00]
 ;
 ; X86-NOOPT-LABEL: test_mul_by_17:
 ; X86-NOOPT:       # BB#0:
 ; X86-NOOPT-NEXT:    imull $17, {{[0-9]+}}(%esp), %eax
 ; X86-NOOPT-NEXT:    retl

 ; X64-HSW-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; X64-HSW-NEXT:    shll $5, %eax # sched: [1:0.50]
 ; X64-HSW-NEXT:    subl %edi, %eax # sched: [1:0.25]
 ; X64-HSW-NEXT:    subl %edi, %eax # sched: [1:0.25]
 ; X64-HSW-NEXT:    retq # sched: [2:1.00]
 ;
 ; X64-JAG-LABEL: test_mul_by_30:
 ; X64-JAG:       # BB#0:
-; X64-JAG-NEXT:    movl %edi, %eax # sched: [1:0.17]
+; X64-JAG-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; X64-JAG-NEXT:    shll $5, %eax # sched: [1:0.50]
 ; X64-JAG-NEXT:    subl %edi, %eax # sched: [1:0.50]
 ; X64-JAG-NEXT:    subl %edi, %eax # sched: [1:0.50]
 ; X64-JAG-NEXT:    retq # sched: [4:1.00]
 ;
 ; X86-NOOPT-LABEL: test_mul_by_30:
 ; X86-NOOPT:       # BB#0:
 ; X86-NOOPT-NEXT:    imull $30, {{[0-9]+}}(%esp), %eax

 ; X64-HSW:       # BB#0:
 ; X64-HSW-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; X64-HSW-NEXT:    shll $5, %eax # sched: [1:0.50]
 ; X64-HSW-NEXT:    subl %edi, %eax # sched: [1:0.25]
 ; X64-HSW-NEXT:    retq # sched: [2:1.00]
 ;
 ; X64-JAG-LABEL: test_mul_by_31:
 ; X64-JAG:       # BB#0:
-; X64-JAG-NEXT:    movl %edi, %eax # sched: [1:0.17]
+; X64-JAG-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; X64-JAG-NEXT:    shll $5, %eax # sched: [1:0.50]
 ; X64-JAG-NEXT:    subl %edi, %eax # sched: [1:0.50]
 ; X64-JAG-NEXT:    retq # sched: [4:1.00]
 ;
 ; X86-NOOPT-LABEL: test_mul_by_31:
 ; X86-NOOPT:       # BB#0:
 ; X86-NOOPT-NEXT:    imull $31, {{[0-9]+}}(%esp), %eax
 ; X86-NOOPT-NEXT:    retl

 ; X64-HSW:       # BB#0:
 ; X64-HSW-NEXT:    shll $5, %edi # sched: [1:0.50]
 ; X64-HSW-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; X64-HSW-NEXT:    retq # sched: [2:1.00]
 ;
 ; X64-JAG-LABEL: test_mul_by_32:
 ; X64-JAG:       # BB#0:
 ; X64-JAG-NEXT:    shll $5, %edi # sched: [1:0.50]
-; X64-JAG-NEXT:    movl %edi, %eax # sched: [1:0.17]
+; X64-JAG-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; X64-JAG-NEXT:    retq # sched: [4:1.00]
 ;
 ; X86-NOOPT-LABEL: test_mul_by_32:
 ; X86-NOOPT:       # BB#0:
 ; X86-NOOPT-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NOOPT-NEXT:    shll $5, %eax
 ; X86-NOOPT-NEXT:    retl
 ;
 ; HSW-NOOPT-LABEL: test_mul_by_32:
 ; HSW-NOOPT:       # BB#0:
 ; HSW-NOOPT-NEXT:    shll $5, %edi # sched: [1:0.50]
 ; HSW-NOOPT-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; HSW-NOOPT-NEXT:    retq # sched: [2:1.00]
 ;
 ; JAG-NOOPT-LABEL: test_mul_by_32:
 ; JAG-NOOPT:       # BB#0:
 ; JAG-NOOPT-NEXT:    shll $5, %edi # sched: [1:0.50]
-; JAG-NOOPT-NEXT:    movl %edi, %eax # sched: [1:0.17]
+; JAG-NOOPT-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; JAG-NOOPT-NEXT:    retq # sched: [4:1.00]
 ;
 ; X64-SLM-LABEL: test_mul_by_32:
 ; X64-SLM:       # BB#0:
 ; X64-SLM-NEXT:    shll $5, %edi # sched: [1:1.00]
 ; X64-SLM-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; X64-SLM-NEXT:    retq # sched: [4:1.00]
 ;

test/CodeGen/X86/mul-constant-i64.ll

 ;
 ; X64-HSW-LABEL: test_mul_by_1:
 ; X64-HSW:       # BB#0:
 ; X64-HSW-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; X64-HSW-NEXT:    retq # sched: [2:1.00]
 ;
 ; X64-JAG-LABEL: test_mul_by_1:
 ; X64-JAG:       # BB#0:
-; X64-JAG-NEXT:    movq %rdi, %rax # sched: [1:0.17]
+; X64-JAG-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; X64-JAG-NEXT:    retq # sched: [4:1.00]
 ;
 ; X86-NOOPT-LABEL: test_mul_by_1:
 ; X86-NOOPT:       # BB#0:
 ; X86-NOOPT-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NOOPT-NEXT:    movl {{[0-9]+}}(%esp), %edx
 ; X86-NOOPT-NEXT:    retl
 ;
 ; HSW-NOOPT-LABEL: test_mul_by_1:
 ; HSW-NOOPT:       # BB#0:
 ; HSW-NOOPT-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; HSW-NOOPT-NEXT:    retq # sched: [2:1.00]
 ;
 ; JAG-NOOPT-LABEL: test_mul_by_1:
 ; JAG-NOOPT:       # BB#0:
-; JAG-NOOPT-NEXT:    movq %rdi, %rax # sched: [1:0.17]
+; JAG-NOOPT-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; JAG-NOOPT-NEXT:    retq # sched: [4:1.00]
 ;
 ; X64-SLM-LABEL: test_mul_by_1:
 ; X64-SLM:       # BB#0:
 ; X64-SLM-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; X64-SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SLM-NOOPT-LABEL: test_mul_by_1:

 ; X64-HSW:       # BB#0:
 ; X64-HSW-NEXT:    shlq $4, %rdi # sched: [1:0.50]
 ; X64-HSW-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; X64-HSW-NEXT:    retq # sched: [2:1.00]
 ;
 ; X64-JAG-LABEL: test_mul_by_16:
 ; X64-JAG:       # BB#0:
 ; X64-JAG-NEXT:    shlq $4, %rdi # sched: [1:0.50]
-; X64-JAG-NEXT:    movq %rdi, %rax # sched: [1:0.17]
+; X64-JAG-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; X64-JAG-NEXT:    retq # sched: [4:1.00]
 ;
 ; X86-NOOPT-LABEL: test_mul_by_16:
 ; X86-NOOPT:       # BB#0:
 ; X86-NOOPT-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NOOPT-NEXT:    movl {{[0-9]+}}(%esp), %edx
 ; X86-NOOPT-NEXT:    shldl $4, %eax, %edx
 ; X86-NOOPT-NEXT:    shll $4, %eax
 ; X86-NOOPT-NEXT:    retl
 ;
 ; HSW-NOOPT-LABEL: test_mul_by_16:
 ; HSW-NOOPT:       # BB#0:
 ; HSW-NOOPT-NEXT:    shlq $4, %rdi # sched: [1:0.50]
 ; HSW-NOOPT-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; HSW-NOOPT-NEXT:    retq # sched: [2:1.00]
 ;
 ; JAG-NOOPT-LABEL: test_mul_by_16:
 ; JAG-NOOPT:       # BB#0:
 ; JAG-NOOPT-NEXT:    shlq $4, %rdi # sched: [1:0.50]
-; JAG-NOOPT-NEXT:    movq %rdi, %rax # sched: [1:0.17]
+; JAG-NOOPT-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; JAG-NOOPT-NEXT:    retq # sched: [4:1.00]
 ;
 ; X64-SLM-LABEL: test_mul_by_16:
 ; X64-SLM:       # BB#0:
 ; X64-SLM-NEXT:    shlq $4, %rdi # sched: [1:1.00]
 ; X64-SLM-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; X64-SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; X64-HSW:       # BB#0:
 ; X64-HSW-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; X64-HSW-NEXT:    shlq $4, %rax # sched: [1:0.50]
 ; X64-HSW-NEXT:    leaq (%rax,%rdi), %rax # sched: [1:0.50]
 ; X64-HSW-NEXT:    retq # sched: [2:1.00]
 ;
 ; X64-JAG-LABEL: test_mul_by_17:
 ; X64-JAG:       # BB#0:
-; X64-JAG-NEXT:    movq %rdi, %rax # sched: [1:0.17]
+; X64-JAG-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; X64-JAG-NEXT:    shlq $4, %rax # sched: [1:0.50]
 ; X64-JAG-NEXT:    leaq (%rax,%rdi), %rax # sched: [1:0.50]
 ; X64-JAG-NEXT:    retq # sched: [4:1.00]
 ;
 ; X86-NOOPT-LABEL: test_mul_by_17:
 ; X86-NOOPT:       # BB#0:
 ; X86-NOOPT-NEXT:    movl $17, %eax
 ; X86-NOOPT-NEXT:    mull {{[0-9]+}}(%esp)

 ; X64-HSW-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; X64-HSW-NEXT:    shlq $5, %rax # sched: [1:0.50]
 ; X64-HSW-NEXT:    subq %rdi, %rax # sched: [1:0.25]
 ; X64-HSW-NEXT:    subq %rdi, %rax # sched: [1:0.25]
 ; X64-HSW-NEXT:    retq # sched: [2:1.00]
 ;
 ; X64-JAG-LABEL: test_mul_by_30:
 ; X64-JAG:       # BB#0:
-; X64-JAG-NEXT:    movq %rdi, %rax # sched: [1:0.17]
+; X64-JAG-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; X64-JAG-NEXT:    shlq $5, %rax # sched: [1:0.50]
 ; X64-JAG-NEXT:    subq %rdi, %rax # sched: [1:0.50]
 ; X64-JAG-NEXT:    subq %rdi, %rax # sched: [1:0.50]
 ; X64-JAG-NEXT:    retq # sched: [4:1.00]
 ;
 ; X86-NOOPT-LABEL: test_mul_by_30:
 ; X86-NOOPT:       # BB#0:
 ; X86-NOOPT-NEXT:    movl $30, %eax

 ; X64-HSW:       # BB#0:
 ; X64-HSW-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; X64-HSW-NEXT:    shlq $5, %rax # sched: [1:0.50]
 ; X64-HSW-NEXT:    subq %rdi, %rax # sched: [1:0.25]
 ; X64-HSW-NEXT:    retq # sched: [2:1.00]
 ;
 ; X64-JAG-LABEL: test_mul_by_31:
 ; X64-JAG:       # BB#0:
-; X64-JAG-NEXT:    movq %rdi, %rax # sched: [1:0.17]
+; X64-JAG-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; X64-JAG-NEXT:    shlq $5, %rax # sched: [1:0.50]
 ; X64-JAG-NEXT:    subq %rdi, %rax # sched: [1:0.50]
 ; X64-JAG-NEXT:    retq # sched: [4:1.00]
 ;
 ; X86-NOOPT-LABEL: test_mul_by_31:
 ; X86-NOOPT:       # BB#0:
 ; X86-NOOPT-NEXT:    movl $31, %eax
 ; X86-NOOPT-NEXT:    mull {{[0-9]+}}(%esp)

 ; X64-HSW:       # BB#0:
 ; X64-HSW-NEXT:    shlq $5, %rdi # sched: [1:0.50]
 ; X64-HSW-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; X64-HSW-NEXT:    retq # sched: [2:1.00]
 ;
 ; X64-JAG-LABEL: test_mul_by_32:
 ; X64-JAG:       # BB#0:
 ; X64-JAG-NEXT:    shlq $5, %rdi # sched: [1:0.50]
-; X64-JAG-NEXT:    movq %rdi, %rax # sched: [1:0.17]
+; X64-JAG-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; X64-JAG-NEXT:    retq # sched: [4:1.00]
 ;
 ; X86-NOOPT-LABEL: test_mul_by_32:
 ; X86-NOOPT:       # BB#0:
 ; X86-NOOPT-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NOOPT-NEXT:    movl {{[0-9]+}}(%esp), %edx
 ; X86-NOOPT-NEXT:    shldl $5, %eax, %edx
 ; X86-NOOPT-NEXT:    shll $5, %eax
 ; X86-NOOPT-NEXT:    retl
 ;
 ; HSW-NOOPT-LABEL: test_mul_by_32:
 ; HSW-NOOPT:       # BB#0:
 ; HSW-NOOPT-NEXT:    shlq $5, %rdi # sched: [1:0.50]
 ; HSW-NOOPT-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; HSW-NOOPT-NEXT:    retq # sched: [2:1.00]
 ;
 ; JAG-NOOPT-LABEL: test_mul_by_32:
 ; JAG-NOOPT:       # BB#0:
 ; JAG-NOOPT-NEXT:    shlq $5, %rdi # sched: [1:0.50]
-; JAG-NOOPT-NEXT:    movq %rdi, %rax # sched: [1:0.17]
+; JAG-NOOPT-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; JAG-NOOPT-NEXT:    retq # sched: [4:1.00]
 ;
 ; X64-SLM-LABEL: test_mul_by_32:
 ; X64-SLM:       # BB#0:
 ; X64-SLM-NEXT:    shlq $5, %rdi # sched: [1:1.00]
 ; X64-SLM-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; X64-SLM-NEXT:    retq # sched: [4:1.00]
 ;

test/CodeGen/X86/schedule-x86_64.ll

 ; SKX:       # BB#0:
 ; SKX-NEXT:    bswapl %edi # sched: [2:0.50]
 ; SKX-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_bswap32:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    bswapl %edi # sched: [1:0.50]
-; BTVER2-NEXT:    movl %edi, %eax # sched: [1:0.17]
+; BTVER2-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_bswap32:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    bswapl %edi # sched: [1:1.00]
 ; ZNVER1-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = tail call i32 asm "bswap $0", "=r,0"(i32 %a0) nounwind

 ; SKX:       # BB#0:
 ; SKX-NEXT:    bswapq %rdi # sched: [2:0.50]
 ; SKX-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_bswap64:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    bswapq %rdi # sched: [1:0.50]
-; BTVER2-NEXT:    movq %rdi, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_bswap64:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    bswapq %rdi # sched: [1:1.00]
 ; ZNVER1-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = tail call i64 asm "bswap $0", "=r,0"(i64 %a0) nounwind

test/CodeGen/X86/sse2-schedule.ll

 ; SKX-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
 ; SKX-NEXT:    vmovd %xmm0, %eax # sched: [2:1.00]
 ; SKX-NEXT:    vmovd %xmm2, (%rsi) # sched: [1:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_movd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [5:1.00]
-; BTVER2-NEXT:    vmovd %edi, %xmm1 # sched: [1:0.17]
+; BTVER2-NEXT:    vmovd %edi, %xmm1 # sched: [1:0.50]
 ; BTVER2-NEXT:    vpaddd %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
 ; BTVER2-NEXT:    vmovd %xmm1, (%rsi) # sched: [1:1.00]
 ; BTVER2-NEXT:    vpaddd %xmm2, %xmm0, %xmm0 # sched: [1:0.50]
-; BTVER2-NEXT:    vmovd %xmm0, %eax # sched: [1:0.17]
+; BTVER2-NEXT:    vmovd %xmm0, %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_movd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [8:0.50]
 ; ZNVER1-NEXT:    vmovd %edi, %xmm1 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vpaddd %xmm1, %xmm0, %xmm1 # sched: [1:0.25]
 ; ZNVER1-NEXT:    vmovd %xmm1, (%rsi) # sched: [1:0.50]

 ; SKX-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
 ; SKX-NEXT:    vmovq %xmm0, %rax # sched: [2:1.00]
 ; SKX-NEXT:    vmovq %xmm2, (%rsi) # sched: [1:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_movd_64:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero sched: [5:1.00]
-; BTVER2-NEXT:    vmovq %rdi, %xmm1 # sched: [1:0.17]
+; BTVER2-NEXT:    vmovq %rdi, %xmm1 # sched: [1:0.50]
 ; BTVER2-NEXT:    vpaddq %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
 ; BTVER2-NEXT:    vmovq %xmm1, (%rsi) # sched: [1:1.00]
 ; BTVER2-NEXT:    vpaddq %xmm2, %xmm0, %xmm0 # sched: [1:0.50]
-; BTVER2-NEXT:    vmovq %xmm0, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    vmovq %xmm0, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_movd_64:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero sched: [8:0.50]
 ; ZNVER1-NEXT:    vmovq %rdi, %xmm1 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vpaddq %xmm1, %xmm0, %xmm1 # sched: [1:0.25]
 ; ZNVER1-NEXT:    vmovq %xmm1, (%rsi) # sched: [1:0.50]

test/CodeGen/X86/sse3-schedule.ll

 ; SKX-NEXT:    leaq (%rdi), %rax # sched: [1:0.50]
 ; SKX-NEXT:    movl %esi, %ecx # sched: [1:0.25]
 ; SKX-NEXT:    monitor # sched: [100:0.25]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_monitor:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    leaq (%rdi), %rax # sched: [1:0.50]
-; BTVER2-NEXT:    movl %esi, %ecx # sched: [1:0.17]
+; BTVER2-NEXT:    movl %esi, %ecx # sched: [1:0.50]
 ; BTVER2-NEXT:    monitor # sched: [100:0.17]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_monitor:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    leaq (%rdi), %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    movl %esi, %ecx # sched: [1:0.25]
 ; ZNVER1-NEXT:    monitor # sched: [100:?]

 ; SKX:       # BB#0:
 ; SKX-NEXT:    movl %edi, %ecx # sched: [1:0.25]
 ; SKX-NEXT:    movl %esi, %eax # sched: [1:0.25]
 ; SKX-NEXT:    mwait # sched: [20:2.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_mwait:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    movl %edi, %ecx # sched: [1:0.17]
-; BTVER2-NEXT:    movl %esi, %eax # sched: [1:0.17]
+; BTVER2-NEXT:    movl %edi, %ecx # sched: [1:0.50]
+; BTVER2-NEXT:    movl %esi, %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    mwait # sched: [100:0.17]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_mwait:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    movl %edi, %ecx # sched: [1:0.25]
 ; ZNVER1-NEXT:    movl %esi, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    mwait # sched: [100:?]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   tail call void @llvm.x86.sse3.mwait(i32 %a0, i32 %a1)
   ret void
 }
 declare void @llvm.x86.sse3.mwait(i32, i32)

test/CodeGen/X86/sse42-schedule.ll

 ; SKX-NEXT:    crc32b (%rdx), %edi # sched: [8:1.00]
 ; SKX-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: crc32_32_8:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    crc32b %sil, %edi # sched: [3:1.00]
 ; BTVER2-NEXT:    crc32b (%rdx), %edi # sched: [8:1.00]
-; BTVER2-NEXT:    movl %edi, %eax # sched: [1:0.17]
+; BTVER2-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: crc32_32_8:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    crc32b %sil, %edi # sched: [3:1.00]
 ; ZNVER1-NEXT:    crc32b (%rdx), %edi # sched: [10:1.00]
 ; ZNVER1-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    crc32w (%rdx), %edi # sched: [8:1.00]
 ; SKX-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: crc32_32_16:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    crc32w %si, %edi # sched: [3:1.00]
 ; BTVER2-NEXT:    crc32w (%rdx), %edi # sched: [8:1.00]
-; BTVER2-NEXT:    movl %edi, %eax # sched: [1:0.17]
+; BTVER2-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: crc32_32_16:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    crc32w %si, %edi # sched: [3:1.00]
 ; ZNVER1-NEXT:    crc32w (%rdx), %edi # sched: [10:1.00]
 ; ZNVER1-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    crc32l (%rdx), %edi # sched: [8:1.00]
 ; SKX-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: crc32_32_32:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    crc32l %esi, %edi # sched: [3:1.00]
 ; BTVER2-NEXT:    crc32l (%rdx), %edi # sched: [8:1.00]
-; BTVER2-NEXT:    movl %edi, %eax # sched: [1:0.17]
+; BTVER2-NEXT:    movl %edi, %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: crc32_32_32:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    crc32l %esi, %edi # sched: [3:1.00]
 ; ZNVER1-NEXT:    crc32l (%rdx), %edi # sched: [10:1.00]
 ; ZNVER1-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    crc32b (%rdx), %edi # sched: [8:1.00]
 ; SKX-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: crc32_64_8:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    crc32b %sil, %edi # sched: [3:1.00]
 ; BTVER2-NEXT:    crc32b (%rdx), %edi # sched: [8:1.00]
-; BTVER2-NEXT:    movq %rdi, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: crc32_64_8:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    crc32b %sil, %edi # sched: [3:1.00]
 ; ZNVER1-NEXT:    crc32b (%rdx), %edi # sched: [10:1.00]
 ; ZNVER1-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    crc32q (%rdx), %rdi # sched: [8:1.00]
 ; SKX-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: crc32_64_64:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    crc32q %rsi, %rdi # sched: [3:1.00]
 ; BTVER2-NEXT:    crc32q (%rdx), %rdi # sched: [8:1.00]
-; BTVER2-NEXT:    movq %rdi, %rax # sched: [1:0.17]
+; BTVER2-NEXT:    movq %rdi, %rax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: crc32_64_64:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    crc32q %rsi, %rdi # sched: [3:1.00]
 ; ZNVER1-NEXT:    crc32q (%rdx), %rdi # sched: [10:1.00]
 ; ZNVER1-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ; SKX-NEXT:    movl $7, %edx # sched: [1:0.25]
 ; SKX-NEXT:    vpcmpestri $7, (%rdi), %xmm0 # sched: [24:4.00]
 ; SKX-NEXT:    # kill: %ECX<def> %ECX<kill> %RCX<def>
 ; SKX-NEXT:    leal (%rcx,%rsi), %eax # sched: [1:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pcmpestri:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    movl $7, %eax # sched: [1:0.17]
-; BTVER2-NEXT:    movl $7, %edx # sched: [1:0.17]
+; BTVER2-NEXT:    movl $7, %eax # sched: [1:0.50]
+; BTVER2-NEXT:    movl $7, %edx # sched: [1:0.50]
 ; BTVER2-NEXT:    vpcmpestri $7, %xmm1, %xmm0 # sched: [13:2.50]
-; BTVER2-NEXT:    movl $7, %eax # sched: [1:0.17]
-; BTVER2-NEXT:    movl $7, %edx # sched: [1:0.17]
-; BTVER2-NEXT:    movl %ecx, %esi # sched: [1:0.17]
+; BTVER2-NEXT:    movl $7, %eax # sched: [1:0.50]
+; BTVER2-NEXT:    movl $7, %edx # sched: [1:0.50]
+; BTVER2-NEXT:    movl %ecx, %esi # sched: [1:0.50]
 ; BTVER2-NEXT:    vpcmpestri $7, (%rdi), %xmm0 # sched: [18:2.50]
 ; BTVER2-NEXT:    # kill: %ECX<def> %ECX<kill> %RCX<def>
 ; BTVER2-NEXT:    leal (%rcx,%rsi), %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pcmpestri:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    movl $7, %eax # sched: [1:0.25]

 ; SKX-NEXT:    vpcmpestrm $7, %xmm1, %xmm0 # sched: [19:4.00]
 ; SKX-NEXT:    movl $7, %eax # sched: [1:0.25]
 ; SKX-NEXT:    movl $7, %edx # sched: [1:0.25]
 ; SKX-NEXT:    vpcmpestrm $7, (%rdi), %xmm0 # sched: [25:4.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pcmpestrm:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    movl $7, %eax # sched: [1:0.17]
-; BTVER2-NEXT:    movl $7, %edx # sched: [1:0.17]
+; BTVER2-NEXT:    movl $7, %eax # sched: [1:0.50]
+; BTVER2-NEXT:    movl $7, %edx # sched: [1:0.50]
 ; BTVER2-NEXT:    vpcmpestrm $7, %xmm1, %xmm0 # sched: [13:2.50]
-; BTVER2-NEXT:    movl $7, %eax # sched: [1:0.17]
-; BTVER2-NEXT:    movl $7, %edx # sched: [1:0.17]
+; BTVER2-NEXT:    movl $7, %eax # sched: [1:0.50]
+; BTVER2-NEXT:    movl $7, %edx # sched: [1:0.50]
 ; BTVER2-NEXT:    vpcmpestrm $7, (%rdi), %xmm0 # sched: [18:2.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pcmpestrm:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    movl $7, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    movl $7, %edx # sched: [1:0.25]
 ; ZNVER1-NEXT:    vpcmpestrm $7, %xmm1, %xmm0 # sched: [100:?]

 ; SKX-NEXT:    vpcmpistri $7, (%rdi), %xmm0 # sched: [16:3.00]
 ; SKX-NEXT:    # kill: %ECX<def> %ECX<kill> %RCX<def>
 ; SKX-NEXT:    leal (%rcx,%rax), %eax # sched: [1:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pcmpistri:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpcmpistri $7, %xmm1, %xmm0 # sched: [6:1.00]
-; BTVER2-NEXT:    movl %ecx, %eax # sched: [1:0.17]
+; BTVER2-NEXT:    movl %ecx, %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    vpcmpistri $7, (%rdi), %xmm0 # sched: [11:1.00]
 ; BTVER2-NEXT:    # kill: %ECX<def> %ECX<kill> %RCX<def>
 ; BTVER2-NEXT:    leal (%rcx,%rax), %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pcmpistri:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpcmpistri $7, %xmm1, %xmm0 # sched: [100:?]