[X86][Sched] A bunch of fixes to the Zen2 sched model latencies.

[courbet]

Summary

As determined with llvm-exegesis.

Some of these look like typos/misunderstandings of the sched model td
spec:

• latency defaults to 1 when not set => Maybe we can avoid having a default ?
• problems with regexps not being anchored by default (XCHG matching CMPXHG)

Note that this is not complete, it fixes only the most obvious mistakes,
and only for latency (not uops).

[courbet]

update mca tests

[RKSimon]

There seems to be a load of whitespace changes (maybe copy+paste issues in D66088?) - please could you regenerate against trunk, commit those NFC changes and then rebase?

[courbet]

Rebase on fixed tests.

[courbet]

In D73172#1833515, @RKSimon wrote:

There seems to be a load of whitespace changes (maybe copy+paste issues in D66088?) - please could you regenerate against trunk, commit those NFC changes and then rebase?

Sure, done.

[GGanesh]

Changes done are obvious. Are you using libpfm with znver2 for verifying these?

[courbet]

In D73172#1841351, @GGanesh wrote:

Changes done are obvious. Are you using libpfm with znver2 for verifying these?

Yes, I'm using llvm-exegesis:

llvm-exegesis -mode=latency -opcode-index=-1 -dump-object-to-disk=0 > /tmp/zen2-latency
llvm-exegesis -mode=analysis -analysis-inconsistencies-output-file=/tmp/zen2-latency.html  -benchmarks-file=/tmp/zen2-latency

[courbet]

@GGanesh does this look reasonable to you ?

[GGanesh]

Yes looks good to me.

[courbet]

Thanks.

[xbolva00]

Backport to 10 branch?

[courbet]

In D73172#1854678, @xbolva00 wrote:

Backport to 10 branch?

No strong opinion: I don't know if there ave a lot of users of llvm-exegesis that are not using LLVM at head.

llvm/lib/Target/X86/X86ScheduleZnver2.td

 def : WriteRes<WriteLEA, [Zn2ALU]>;
 defm : Zn2WriteResPair<WriteALU,   [Zn2ALU], 1>;
 defm : Zn2WriteResPair<WriteADC,   [Zn2ALU], 1>;
 
 defm : Zn2WriteResPair<WriteIMul8,     [Zn2ALU1, Zn2Multiplier], 4>;
 
 defm : X86WriteRes<WriteBSWAP32, [Zn2ALU], 1, [4], 1>;
 defm : X86WriteRes<WriteBSWAP64, [Zn2ALU], 1, [4], 1>;
-defm : X86WriteRes<WriteCMPXCHG, [Zn2ALU], 1, [1], 1>;
+defm : X86WriteRes<WriteCMPXCHG, [Zn2ALU], 3, [1], 1>;
 defm : X86WriteRes<WriteCMPXCHGRMW,[Zn2ALU,Zn2AGU], 8, [1,1], 5>;
 defm : X86WriteRes<WriteXCHG, [Zn2ALU], 1, [2], 2>;
 
 defm : Zn2WriteResPair<WriteShift, [Zn2ALU], 1>;
 defm : Zn2WriteResPair<WriteShiftCL,  [Zn2ALU], 1>;
 defm : Zn2WriteResPair<WriteRotate,   [Zn2ALU], 1>;
 defm : Zn2WriteResPair<WriteRotateCL, [Zn2ALU], 1>;
 

 
 defm : X86WriteRes<WriteBitTest,         [Zn2ALU], 1, [1], 1>;
 defm : X86WriteRes<WriteBitTestImmLd,    [Zn2ALU,Zn2AGU], 5, [1,1], 2>;
 defm : X86WriteRes<WriteBitTestRegLd,    [Zn2ALU,Zn2AGU], 5, [1,1], 2>;
 defm : X86WriteRes<WriteBitTestSet,      [Zn2ALU], 2, [1], 2>;
 
 // Bit counts.
 defm : Zn2WriteResPair<WriteBSF, [Zn2ALU], 3>;
-defm : Zn2WriteResPair<WriteBSR, [Zn2ALU], 3>;
+defm : Zn2WriteResPair<WriteBSR, [Zn2ALU], 4>;
 defm : Zn2WriteResPair<WriteLZCNT,          [Zn2ALU], 1>;
 defm : Zn2WriteResPair<WriteTZCNT,          [Zn2ALU], 2>;
 defm : Zn2WriteResPair<WritePOPCNT,         [Zn2ALU], 1>;
 
 // Treat misc copies as a move.
 def : InstRW<[WriteMove], (instrs COPY)>;
 
 // BMI1 BEXTR, BMI2 BZHI

 defm : Zn2WriteResFpuPair<WriteFAdd,      [Zn2FPU0],  3>;
 defm : Zn2WriteResFpuPair<WriteFAddX,     [Zn2FPU0],  3>;
 defm : Zn2WriteResFpuPair<WriteFAddY,     [Zn2FPU0],  3>;
 defm : X86WriteResPairUnsupported<WriteFAddZ>;
 defm : Zn2WriteResFpuPair<WriteFAdd64,    [Zn2FPU0],  3>;
 defm : Zn2WriteResFpuPair<WriteFAdd64X,   [Zn2FPU0],  3>;
 defm : Zn2WriteResFpuPair<WriteFAdd64Y,   [Zn2FPU0],  3>;
 defm : X86WriteResPairUnsupported<WriteFAdd64Z>;
-defm : Zn2WriteResFpuPair<WriteFCmp,      [Zn2FPU0],  3>;
-defm : Zn2WriteResFpuPair<WriteFCmpX,     [Zn2FPU0],  3>;
-defm : Zn2WriteResFpuPair<WriteFCmpY,     [Zn2FPU0],  3>;
+defm : Zn2WriteResFpuPair<WriteFCmp,      [Zn2FPU0],  1>;
+defm : Zn2WriteResFpuPair<WriteFCmpX,     [Zn2FPU0],  1>;
+defm : Zn2WriteResFpuPair<WriteFCmpY,     [Zn2FPU0],  1>;
 defm : X86WriteResPairUnsupported<WriteFCmpZ>;
-defm : Zn2WriteResFpuPair<WriteFCmp64,    [Zn2FPU0],  3>;
-defm : Zn2WriteResFpuPair<WriteFCmp64X,   [Zn2FPU0],  3>;
-defm : Zn2WriteResFpuPair<WriteFCmp64Y,   [Zn2FPU0],  3>;
+defm : Zn2WriteResFpuPair<WriteFCmp64,    [Zn2FPU0],  1>;
+defm : Zn2WriteResFpuPair<WriteFCmp64X,   [Zn2FPU0],  1>;
+defm : Zn2WriteResFpuPair<WriteFCmp64Y,   [Zn2FPU0],  1>;
 defm : X86WriteResPairUnsupported<WriteFCmp64Z>;
 defm : Zn2WriteResFpuPair<WriteFCom,      [Zn2FPU0],  3>;
 defm : Zn2WriteResFpuPair<WriteFComX,     [Zn2FPU0],  3>;
 defm : Zn2WriteResFpuPair<WriteFBlend,    [Zn2FPU01], 1>;
 defm : Zn2WriteResFpuPair<WriteFBlendY,   [Zn2FPU01], 1>;
 defm : X86WriteResPairUnsupported<WriteFBlendZ>;
 defm : Zn2WriteResFpuPair<WriteFVarBlend, [Zn2FPU01], 1>;
 defm : Zn2WriteResFpuPair<WriteFVarBlendY,[Zn2FPU01], 1>;

 defm : X86WriteResPairUnsupported<WriteCvtI2PDZ>;
 defm : Zn2WriteResFpuPair<WriteFDiv,      [Zn2FPU3], 15>;
 defm : Zn2WriteResFpuPair<WriteFDivX,     [Zn2FPU3], 15>;
 defm : X86WriteResPairUnsupported<WriteFDivZ>;
 defm : Zn2WriteResFpuPair<WriteFDiv64,    [Zn2FPU3], 15>;
 defm : Zn2WriteResFpuPair<WriteFDiv64X,   [Zn2FPU3], 15>;
 defm : X86WriteResPairUnsupported<WriteFDiv64Z>;
 defm : Zn2WriteResFpuPair<WriteFSign,     [Zn2FPU3],  2>;
-defm : Zn2WriteResFpuPair<WriteFRnd,      [Zn2FPU3],  4, [1], 1, 7, 0>;
-defm : Zn2WriteResFpuPair<WriteFRndY,     [Zn2FPU3],  4, [1], 1, 7, 0>;
+defm : Zn2WriteResFpuPair<WriteFRnd,      [Zn2FPU3],  3, [1], 1, 7, 0>;
+defm : Zn2WriteResFpuPair<WriteFRndY,     [Zn2FPU3],  3, [1], 1, 7, 0>;
 defm : X86WriteResPairUnsupported<WriteFRndZ>;
 defm : Zn2WriteResFpuPair<WriteFLogic,    [Zn2FPU],   1>;
 defm : Zn2WriteResFpuPair<WriteFLogicY,   [Zn2FPU],   1>;
 defm : X86WriteResPairUnsupported<WriteFLogicZ>;
 defm : Zn2WriteResFpuPair<WriteFTest,     [Zn2FPU],   1>;
 defm : Zn2WriteResFpuPair<WriteFTestY,    [Zn2FPU],   1>;
 defm : X86WriteResPairUnsupported<WriteFTestZ>;
 defm : Zn2WriteResFpuPair<WriteFShuffle,  [Zn2FPU12], 1>;
 defm : Zn2WriteResFpuPair<WriteFShuffleY, [Zn2FPU12], 1>;
 defm : X86WriteResPairUnsupported<WriteFShuffleZ>;
-defm : Zn2WriteResFpuPair<WriteFVarShuffle, [Zn2FPU12], 1>;
-defm : Zn2WriteResFpuPair<WriteFVarShuffleY,[Zn2FPU12], 1>;
+defm : Zn2WriteResFpuPair<WriteFVarShuffle, [Zn2FPU12], 3>;
+defm : Zn2WriteResFpuPair<WriteFVarShuffleY,[Zn2FPU12], 3>;
 defm : X86WriteResPairUnsupported<WriteFVarShuffleZ>;
 defm : Zn2WriteResFpuPair<WriteFMul,      [Zn2FPU01], 3, [1], 1, 7, 1>;
 defm : Zn2WriteResFpuPair<WriteFMulX,     [Zn2FPU01], 3, [1], 1, 7, 1>;
-defm : Zn2WriteResFpuPair<WriteFMulY,     [Zn2FPU01], 4, [1], 1, 7, 1>;
+defm : Zn2WriteResFpuPair<WriteFMulY,     [Zn2FPU01], 3, [1], 1, 7, 1>;
 defm : X86WriteResPairUnsupported<WriteFMulZ>;
 defm : Zn2WriteResFpuPair<WriteFMul64,    [Zn2FPU01], 3, [1], 1, 7, 1>;
 defm : Zn2WriteResFpuPair<WriteFMul64X,   [Zn2FPU01], 3, [1], 1, 7, 1>;
-defm : Zn2WriteResFpuPair<WriteFMul64Y,   [Zn2FPU01], 4, [1], 1, 7, 1>;
+defm : Zn2WriteResFpuPair<WriteFMul64Y,   [Zn2FPU01], 3, [1], 1, 7, 1>;
 defm : X86WriteResPairUnsupported<WriteFMul64Z>;
 defm : Zn2WriteResFpuPair<WriteFMA,       [Zn2FPU03], 5>;
 defm : Zn2WriteResFpuPair<WriteFMAX,      [Zn2FPU03], 5>;
 defm : Zn2WriteResFpuPair<WriteFMAY,      [Zn2FPU03], 5>;
 defm : X86WriteResPairUnsupported<WriteFMAZ>;
 defm : Zn2WriteResFpuPair<WriteFRcp,      [Zn2FPU01], 5>;
 defm : Zn2WriteResFpuPair<WriteFRcpX,     [Zn2FPU01], 5>;
 defm : Zn2WriteResFpuPair<WriteFRcpY,     [Zn2FPU01], 5, [1], 1, 7, 2>;

 defm : X86WriteRes<WriteVecMoveX,        [Zn2FPU], 1, [1], 1>;
 defm : X86WriteRes<WriteVecMoveY,        [Zn2FPU], 2, [1], 2>;
 defm : X86WriteRes<WriteVecMoveToGpr,    [Zn2FPU2], 2, [1], 1>;
 defm : X86WriteRes<WriteVecMoveFromGpr,  [Zn2FPU2], 3, [1], 1>;
 defm : X86WriteRes<WriteEMMS,            [Zn2FPU], 2, [1], 1>;
 
 defm : Zn2WriteResFpuPair<WriteVecShift,   [Zn2FPU],   1>;
 defm : Zn2WriteResFpuPair<WriteVecShiftX,  [Zn2FPU2],  1>;
-defm : Zn2WriteResFpuPair<WriteVecShiftY,  [Zn2FPU2],  2>;
+defm : Zn2WriteResFpuPair<WriteVecShiftY,  [Zn2FPU2],  1>;
 defm : X86WriteResPairUnsupported<WriteVecShiftZ>;
 defm : Zn2WriteResFpuPair<WriteVecShiftImm,  [Zn2FPU], 1>;
 defm : Zn2WriteResFpuPair<WriteVecShiftImmX, [Zn2FPU], 1>;
 defm : Zn2WriteResFpuPair<WriteVecShiftImmY, [Zn2FPU], 1>;
 defm : X86WriteResPairUnsupported<WriteVecShiftImmZ>;
 defm : Zn2WriteResFpuPair<WriteVecLogic,   [Zn2FPU],   1>;
 defm : Zn2WriteResFpuPair<WriteVecLogicX,  [Zn2FPU],   1>;
 defm : Zn2WriteResFpuPair<WriteVecLogicY,  [Zn2FPU],   1>;
 defm : X86WriteResPairUnsupported<WriteVecLogicZ>;
 defm : Zn2WriteResFpuPair<WriteVecTest,    [Zn2FPU12], 1, [2], 1, 7, 1>;
 defm : Zn2WriteResFpuPair<WriteVecTestY,   [Zn2FPU12], 1, [2], 1, 7, 1>;
 defm : X86WriteResPairUnsupported<WriteVecTestZ>;
 defm : Zn2WriteResFpuPair<WriteVecALU,     [Zn2FPU],   1>;
 defm : Zn2WriteResFpuPair<WriteVecALUX,    [Zn2FPU],   1>;
 defm : Zn2WriteResFpuPair<WriteVecALUY,    [Zn2FPU],   1>;
 defm : X86WriteResPairUnsupported<WriteVecALUZ>;
 defm : Zn2WriteResFpuPair<WriteVecIMul,    [Zn2FPU0],  4>;
 defm : Zn2WriteResFpuPair<WriteVecIMulX,   [Zn2FPU0],  4>;
 defm : Zn2WriteResFpuPair<WriteVecIMulY,   [Zn2FPU0],  4>;
 defm : X86WriteResPairUnsupported<WriteVecIMulZ>;
 defm : Zn2WriteResFpuPair<WritePMULLD,     [Zn2FPU0],  4, [1], 1, 7, 1>;
-defm : Zn2WriteResFpuPair<WritePMULLDY,    [Zn2FPU0],  3, [1], 1, 7, 1>;
+defm : Zn2WriteResFpuPair<WritePMULLDY,    [Zn2FPU0],  4, [1], 1, 7, 1>;
 defm : X86WriteResPairUnsupported<WritePMULLDZ>;
 defm : Zn2WriteResFpuPair<WriteShuffle,    [Zn2FPU],   1>;
 defm : Zn2WriteResFpuPair<WriteShuffleX,   [Zn2FPU],   1>;
 defm : Zn2WriteResFpuPair<WriteShuffleY,   [Zn2FPU],   1>;
 defm : X86WriteResPairUnsupported<WriteShuffleZ>;
 defm : Zn2WriteResFpuPair<WriteVarShuffle, [Zn2FPU],   1>;
 defm : Zn2WriteResFpuPair<WriteVarShuffleX,[Zn2FPU],   1>;
 defm : Zn2WriteResFpuPair<WriteVarShuffleY,[Zn2FPU],   1>;
 defm : X86WriteResPairUnsupported<WriteVarShuffleZ>;
 defm : Zn2WriteResFpuPair<WriteBlend,      [Zn2FPU01], 1>;
 defm : Zn2WriteResFpuPair<WriteBlendY,     [Zn2FPU01], 1>;
 defm : X86WriteResPairUnsupported<WriteBlendZ>;
 defm : Zn2WriteResFpuPair<WriteShuffle256, [Zn2FPU],   2>;
 defm : Zn2WriteResFpuPair<WriteVarShuffle256, [Zn2FPU],   2>;
 defm : Zn2WriteResFpuPair<WritePSADBW,     [Zn2FPU0],  3>;
 defm : Zn2WriteResFpuPair<WritePSADBWX,    [Zn2FPU0],  3>;
 defm : Zn2WriteResFpuPair<WritePSADBWY,    [Zn2FPU0],  3>;
 defm : X86WriteResPairUnsupported<WritePSADBWZ>;
 defm : Zn2WriteResFpuPair<WritePHMINPOS,   [Zn2FPU0],  4>;
 
 // Vector Shift Operations
-defm : Zn2WriteResFpuPair<WriteVarVecShift,  [Zn2FPU12], 1>;
-defm : Zn2WriteResFpuPair<WriteVarVecShiftY, [Zn2FPU12], 1>;
+defm : Zn2WriteResFpuPair<WriteVarVecShift,  [Zn2FPU12], 3>;
+defm : Zn2WriteResFpuPair<WriteVarVecShiftY, [Zn2FPU12], 3>;
 defm : X86WriteResPairUnsupported<WriteVarVecShiftZ>;
 
 // Vector insert/extract operations.
 defm : Zn2WriteResFpuPair<WriteVecInsert,   [Zn2FPU],   1>;
 
 def : WriteRes<WriteVecExtract, [Zn2FPU12, Zn2FPU2]> {
   let Latency = 2;
   let ResourceCycles = [1, 2];

 // We set long latency so as to block the entire pipeline.
 defm : Zn2WriteResFpuPair<WriteFShuffle256, [Zn2FPU], 100>;
 defm : Zn2WriteResFpuPair<WriteFVarShuffle256, [Zn2FPU], 100>;
 
 // Microcoded Instructions
 def Zn2WriteMicrocoded : SchedWriteRes<[]> {
   let Latency = 100;
 }
+defm : Zn2WriteResPair<WriteDPPS, [], 15>;
+defm : Zn2WriteResPair<WriteFHAdd, [], 7>;
+defm : Zn2WriteResPair<WriteFHAddY, [], 7>;
+defm : Zn2WriteResPair<WritePHAdd, [], 3>;
+defm : Zn2WriteResPair<WritePHAddX, [], 3>;
+defm : Zn2WriteResPair<WritePHAddY, [], 3>;
 
 def : SchedAlias<WriteMicrocoded, Zn2WriteMicrocoded>;
 def : SchedAlias<WriteFCMOV, Zn2WriteMicrocoded>;
 def : SchedAlias<WriteSystem, Zn2WriteMicrocoded>;
 def : SchedAlias<WriteMPSAD, Zn2WriteMicrocoded>;
 def : SchedAlias<WriteMPSADY, Zn2WriteMicrocoded>;
 def : SchedAlias<WriteMPSADLd, Zn2WriteMicrocoded>;
 def : SchedAlias<WriteMPSADYLd, Zn2WriteMicrocoded>;

 def : InstRW<[WriteLoad], (instregex "MOV(S|Z)X32rm(8|16)")>;
 
 // XCHG.
 // r,r.
 def Zn2WriteXCHG : SchedWriteRes<[Zn2ALU]> {
   let NumMicroOps = 2;
 }
 
-def : InstRW<[Zn2WriteXCHG], (instregex "XCHG(8|16|32|64)rr", "XCHG(16|32|64)ar")>;
+def : InstRW<[Zn2WriteXCHG], (instregex "^XCHG(8|16|32|64)rr", "^XCHG(16|32|64)ar")>;
 
 // r,m.
 def Zn2WriteXCHGrm : SchedWriteRes<[Zn2AGU, Zn2ALU]> {
   let Latency = 5;
   let NumMicroOps = 2;
 }
-def : InstRW<[Zn2WriteXCHGrm, ReadAfterLd], (instregex "XCHG(8|16|32|64)rm")>;
+def : InstRW<[Zn2WriteXCHGrm, ReadAfterLd], (instregex "^XCHG(8|16|32|64)rm")>;
 
 def : InstRW<[WriteMicrocoded], (instrs XLAT)>;
 
 // POP16.
 // r.
 def Zn2WritePop16r : SchedWriteRes<[Zn2AGU]>{
   let Latency = 5;
   let NumMicroOps = 2;

 def : InstRW<[WriteALULd],
              (instregex "(INC|DEC|NOT|NEG)(8|16|32|64)m")>;
 
 // MUL IMUL.
 // r16.
 def Zn2WriteMul16 : SchedWriteRes<[Zn2ALU1, Zn2Multiplier]> {
   let Latency = 3;
 }
+def Zn2WriteMul16Imm : SchedWriteRes<[Zn2ALU1, Zn2Multiplier]> {
+  let Latency = 4;
+}
 def : SchedAlias<WriteIMul16, Zn2WriteMul16>;
-def : SchedAlias<WriteIMul16Imm, Zn2WriteMul16>;
+def : SchedAlias<WriteIMul16Imm, Zn2WriteMul16Imm>;
 def : SchedAlias<WriteIMul16Reg, Zn2WriteMul16>;
 
 // m16.
 def Zn2WriteMul16Ld : SchedWriteRes<[Zn2AGU, Zn2ALU1, Zn2Multiplier]> {
   let Latency = 7;
 }
 def : SchedAlias<WriteIMul16Ld, Zn2WriteMul16Ld>;
 def : SchedAlias<WriteIMul16ImmLd, Zn2WriteMul16Ld>;

 def : InstRW<[WriteMicrocoded], (instrs FNINIT)>;
 
 //=== Integer MMX and XMM Instructions ===//
 
 // PACKSSWB/DW.
 // mm <- mm.
 def Zn2WriteFPU12 : SchedWriteRes<[Zn2FPU12]> ;
 def Zn2WriteFPU12Y : SchedWriteRes<[Zn2FPU12]> {
+  let Latency = 4;
   let NumMicroOps = 2;
 }
 def Zn2WriteFPU12m : SchedWriteRes<[Zn2AGU, Zn2FPU12]> ;
 def Zn2WriteFPU12Ym : SchedWriteRes<[Zn2AGU, Zn2FPU12]> {
   let Latency = 8;
   let NumMicroOps = 2;
 }
 

 def : InstRW<[Zn2WriteVPBROADCAST256Ld],
                                      (instregex "VPBROADCAST(B|W)Yrm")>;
 
 // VPGATHER.
 def : InstRW<[WriteMicrocoded], (instregex "VPGATHER(Q|D)(Q|D)(Y?)rm")>;
 
 //-- Arithmetic instructions --//
 
-// HADD, HSUB PS/PD
-// PHADD|PHSUB (S) W/D.
-def : SchedAlias<WritePHAdd,    Zn2WriteMicrocoded>;
-def : SchedAlias<WritePHAddLd,  Zn2WriteMicrocoded>;
-def : SchedAlias<WritePHAddX,   Zn2WriteMicrocoded>;
-def : SchedAlias<WritePHAddXLd, Zn2WriteMicrocoded>;
-def : SchedAlias<WritePHAddY,   Zn2WriteMicrocoded>;
-def : SchedAlias<WritePHAddYLd, Zn2WriteMicrocoded>;
-
 // PCMPGTQ.
 def Zn2WritePCMPGTQr : SchedWriteRes<[Zn2FPU03]>;
 def : InstRW<[Zn2WritePCMPGTQr], (instregex "(V?)PCMPGTQ(Y?)rr")>;
 
 // x <- x,m.
 def Zn2WritePCMPGTQm : SchedWriteRes<[Zn2AGU, Zn2FPU03]> {
   let Latency = 8;
 }
 // ymm.
 def Zn2WritePCMPGTQYm : SchedWriteRes<[Zn2AGU, Zn2FPU03]> {
   let Latency = 8;
 }
 def : InstRW<[Zn2WritePCMPGTQm], (instregex "(V?)PCMPGTQrm")>;
 def : InstRW<[Zn2WritePCMPGTQYm], (instrs VPCMPGTQYrm)>;
 
 //-- Logic instructions --//
 
 // PSLL,PSRL,PSRA W/D/Q.
 // x,x / v,v,x.
-def Zn2WritePShift  : SchedWriteRes<[Zn2FPU2]> ;
-def Zn2WritePShiftY : SchedWriteRes<[Zn2FPU2]> ;
+def Zn2WritePShift  : SchedWriteRes<[Zn2FPU2]> {
+  let Latency = 3;
+}
+def Zn2WritePShiftY : SchedWriteRes<[Zn2FPU2]> {
+  let Latency = 3;
+}
 
 // PSLL,PSRL DQ.
 def : InstRW<[Zn2WritePShift], (instregex "(V?)PS(R|L)LDQri")>;
 def : InstRW<[Zn2WritePShiftY], (instregex "(V?)PS(R|L)LDQYri")>;
 
 //=== Floating Point XMM and YMM Instructions ===//
 //-- Move instructions --//
 

 }
 def : SchedAlias<WriteCvtSS2SDLd, Zn2WriteCVTSS2SDLd>;
 
 def Zn2WriteCVTDQ2PDr: SchedWriteRes<[Zn2FPU12,Zn2FPU3]> {
   let Latency = 3;
 }
 // CVTDQ2PD.
 // x,x.
-def : InstRW<[Zn2WriteCVTDQ2PDr], (instregex "(V)?CVTDQ2PDrr")>;
+def : InstRW<[Zn2WriteCVTDQ2PDr], (instregex "(V)?CVTDQ2P(D|S)rr")>;
 
 // Same as xmm
 // y,x.
 def : InstRW<[Zn2WriteCVTDQ2PDr], (instrs VCVTDQ2PDYrr)>;
 def : InstRW<[Zn2WriteCVTDQ2PDr], (instrs VCVTDQ2PSYrr)>;
 
 def Zn2WriteCVTPD2DQr: SchedWriteRes<[Zn2FPU12, Zn2FPU3]> {
   let Latency = 3;
 }
-// CVT(T)PD2DQ.
+// CVT(T)P(D|S)2DQ.
 // x,x.
-def : InstRW<[Zn2WriteCVTPD2DQr], (instregex "(V?)CVT(T?)PD2DQrr")>;
+def : InstRW<[Zn2WriteCVTPD2DQr], (instregex "(V?)CVT(T?)P(D|S)2DQrr")>;
 
 def Zn2WriteCVTPD2DQLd: SchedWriteRes<[Zn2AGU,Zn2FPU12,Zn2FPU3]> {
   let Latency = 10;
   let NumMicroOps = 2;
 }
 // x,m128.
 def : InstRW<[Zn2WriteCVTPD2DQLd], (instregex "(V?)CVT(T?)PD2DQrm")>;
 // same as xmm handling

 // x,mm.
 def : InstRW<[Zn2WriteCVTPS2PDr], (instrs MMX_CVTPI2PDirr)>;
 
 // CVT(T)PD2PI.
 // mm,x.
 def : InstRW<[Zn2WriteCVTPS2PIr], (instregex "MMX_CVT(T?)PD2PIirr")>;
 
 def Zn2WriteCVSTSI2SSr: SchedWriteRes<[Zn2FPU3]> {
-  let Latency = 4;
+  let Latency = 3;
 }
 
 // same as CVTPD2DQr
 // CVT(T)SS2SI.
 // r32,x.
 def : InstRW<[Zn2WriteCVTPD2DQr], (instregex "(V?)CVT(T?)SS2SI(64)?rr")>;
 // same as CVTPD2DQm
 // r32,m32.
 def : InstRW<[Zn2WriteCVTPD2DQLd], (instregex "(V?)CVT(T?)SS2SI(64)?rm")>;
 
 def Zn2WriteCVSTSI2SDr: SchedWriteRes<[Zn2FPU013, Zn2FPU3]> {
-  let Latency = 4;
+  let Latency = 3;
 }
 // CVTSI2SD.
 // x,r32/64.
 def : InstRW<[Zn2WriteCVSTSI2SDr], (instregex "(V?)CVTSI(64)?2SDrr")>;
 
 
 def Zn2WriteCVSTSI2SIr: SchedWriteRes<[Zn2FPU3, Zn2FPU2]> {
   let Latency = 4;

 // v,m.
 def : SchedAlias<WriteCvtPH2PSLd,  Zn2WriteMicrocoded>;
 def : SchedAlias<WriteCvtPH2PSYLd, Zn2WriteMicrocoded>;
 defm : X86WriteResUnsupported<WriteCvtPH2PSZLd>;
 
 //-- SSE4A instructions --//
 // EXTRQ
 def Zn2WriteEXTRQ: SchedWriteRes<[Zn2FPU12, Zn2FPU2]> {
-  let Latency = 2;
+  let Latency = 3;
 }
 def : InstRW<[Zn2WriteEXTRQ], (instregex "EXTRQ")>;
 
 // INSERTQ
 def Zn2WriteINSERTQ: SchedWriteRes<[Zn2FPU03,Zn2FPU1]> {
   let Latency = 4;
 }
 def : InstRW<[Zn2WriteINSERTQ], (instregex "INSERTQ")>;

 // x,m.
 def Zn2WriteSHA256RNDS2Ld : SchedWriteRes<[Zn2AGU, Zn2FPU1]> {
   let Latency = 11;
 }
 def : InstRW<[Zn2WriteSHA256RNDS2Ld], (instregex "SHA256RNDS2rm")>;
 
 //-- Arithmetic instructions --//
 
-// HADD, HSUB PS/PD
-def : SchedAlias<WriteFHAdd,    Zn2WriteMicrocoded>;
-def : SchedAlias<WriteFHAddLd,  Zn2WriteMicrocoded>;
-def : SchedAlias<WriteFHAddY,   Zn2WriteMicrocoded>;
-def : SchedAlias<WriteFHAddYLd, Zn2WriteMicrocoded>;
-
 // VDIVPS.
 // TODO - convert to Zn2WriteResFpuPair
 // y,y,y.
 def Zn2WriteVDIVPSYr : SchedWriteRes<[Zn2FPU3]> {
   let Latency = 10;
   let ResourceCycles = [10];
 }
 def : SchedAlias<WriteFDivY,   Zn2WriteVDIVPSYr>;
def Zn2WriteVDIVPDYLd : SchedWriteRes<[Zn2AGU, Zn2FPU3]> {
   let Latency = 20;
   let NumMicroOps = 2;
   let ResourceCycles = [1,20];
 }
 def : SchedAlias<WriteFDiv64YLd, Zn2WriteVDIVPDYLd>;
 
 // DPPS.
 // x,x,i / v,v,v,i.
-def : SchedAlias<WriteDPPS,   Zn2WriteMicrocoded>;
 def : SchedAlias<WriteDPPSY,  Zn2WriteMicrocoded>;
 
 // x,m,i / v,v,m,i.
-def : SchedAlias<WriteDPPSLd, Zn2WriteMicrocoded>;
 def : SchedAlias<WriteDPPSYLd,Zn2WriteMicrocoded>;
 
 // DPPD.
 // x,x,i.
 def : SchedAlias<WriteDPPD,   Zn2WriteMicrocoded>;
 
 // x,m,i.
 def : SchedAlias<WriteDPPDLd, Zn2WriteMicrocoded>;

llvm/test/tools/llvm-mca/X86/Znver2/resources-avx1.s

 # CHECK-NEXT:  1      1     0.50                        vblendvps %xmm3, %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     0.50    *                   vblendvps %xmm3, (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.50                        vblendvps %ymm3, %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.50    *                   vblendvps %ymm3, (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  2      8     0.50    *                   vbroadcastf128  (%rax), %ymm2
 # CHECK-NEXT:  1      8     0.50    *                   vbroadcastsd  (%rax), %ymm2
 # CHECK-NEXT:  1      8     0.50    *                   vbroadcastss  (%rax), %xmm2
 # CHECK-NEXT:  1      8     0.50    *                   vbroadcastss  (%rax), %ymm2
-# CHECK-NEXT:  1      3     1.00                        vcmpeqpd  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   vcmpeqpd  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      3     1.00                        vcmpeqpd  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      10    1.00    *                   vcmpeqpd  (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      3     1.00                        vcmpeqps  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   vcmpeqps  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      3     1.00                        vcmpeqps  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      10    1.00    *                   vcmpeqps  (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      3     1.00                        vcmpeqsd  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   vcmpeqsd  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      3     1.00                        vcmpeqss  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   vcmpeqss  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00                        vcmpeqpd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   vcmpeqpd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00                        vcmpeqpd  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vcmpeqpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vcmpeqps  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   vcmpeqps  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00                        vcmpeqps  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vcmpeqps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vcmpeqsd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   vcmpeqsd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00                        vcmpeqss  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   vcmpeqss  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      3     1.00                        vcomisd %xmm0, %xmm1
 # CHECK-NEXT:  1      10    1.00    *                   vcomisd (%rax), %xmm1
 # CHECK-NEXT:  1      3     1.00                        vcomiss %xmm0, %xmm1
 # CHECK-NEXT:  1      10    1.00    *                   vcomiss (%rax), %xmm1
 # CHECK-NEXT:  1      3     1.00                        vcvtdq2pd %xmm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   vcvtdq2pd (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        vcvtdq2pd %xmm0, %ymm2
 # CHECK-NEXT:  1      12    1.00    *                   vcvtdq2pd (%rax), %ymm2
-# CHECK-NEXT:  1      5     1.00                        vcvtdq2ps %xmm0, %xmm2
+# CHECK-NEXT:  1      3     1.00                        vcvtdq2ps %xmm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   vcvtdq2ps (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        vcvtdq2ps %ymm0, %ymm2
 # CHECK-NEXT:  1      12    1.00    *                   vcvtdq2ps (%rax), %ymm2
 # CHECK-NEXT:  1      3     1.00                        vcvtpd2dq %xmm0, %xmm2
 # CHECK-NEXT:  2      10    1.00    *                   vcvtpd2dqx  (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        vcvtpd2dq %ymm0, %xmm2
 # CHECK-NEXT:  2      10    1.00    *                   vcvtpd2dqy  (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        vcvtpd2ps %xmm0, %xmm2
 # CHECK-NEXT:  2      10    0.50    *                   vcvtpd2psx  (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        vcvtpd2ps %ymm0, %xmm2
 # CHECK-NEXT:  1      10    1.00    *                   vcvtpd2psy  (%rax), %xmm2
-# CHECK-NEXT:  1      5     1.00                        vcvtps2dq %xmm0, %xmm2
+# CHECK-NEXT:  1      3     1.00                        vcvtps2dq %xmm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   vcvtps2dq (%rax), %xmm2
 # CHECK-NEXT:  1      5     1.00                        vcvtps2dq %ymm0, %ymm2
 # CHECK-NEXT:  1      12    1.00    *                   vcvtps2dq (%rax), %ymm2
 # CHECK-NEXT:  1      3     1.00                        vcvtps2pd %xmm0, %xmm2
 # CHECK-NEXT:  2      10    1.00    *                   vcvtps2pd (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        vcvtps2pd %xmm0, %ymm2
 # CHECK-NEXT:  2      10    1.00    *                   vcvtps2pd (%rax), %ymm2
 # CHECK-NEXT:  1      4     1.00                        vcvtsd2si %xmm0, %ecx
 # CHECK-NEXT:  1      4     1.00                        vcvtsd2si %xmm0, %rcx
 # CHECK-NEXT:  1      11    1.00    *                   vcvtsd2si (%rax), %ecx
 # CHECK-NEXT:  1      11    1.00    *                   vcvtsd2si (%rax), %rcx
 # CHECK-NEXT:  1      3     1.00                        vcvtsd2ss %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  2      10    0.50    *                   vcvtsd2ss (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      4     1.00                        vcvtsi2sd %ecx, %xmm0, %xmm2
-# CHECK-NEXT:  1      4     1.00                        vcvtsi2sd %rcx, %xmm0, %xmm2
+# CHECK-NEXT:  1      3     1.00                        vcvtsi2sd %ecx, %xmm0, %xmm2
+# CHECK-NEXT:  1      3     1.00                        vcvtsi2sd %rcx, %xmm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   vcvtsi2sdl  (%rax), %xmm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   vcvtsi2sdq  (%rax), %xmm0, %xmm2
 # CHECK-NEXT:  1      5     1.00                        vcvtsi2ss %ecx, %xmm0, %xmm2
 # CHECK-NEXT:  1      5     1.00                        vcvtsi2ss %rcx, %xmm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   vcvtsi2ssl  (%rax), %xmm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   vcvtsi2ssq  (%rax), %xmm0, %xmm2
 # CHECK-NEXT:  1      3     1.00                        vcvtss2sd %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  2      10    2.00    *                   vcvtss2sd (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      3     1.00                        vcvtss2si %xmm0, %ecx
 # CHECK-NEXT:  1      3     1.00                        vcvtss2si %xmm0, %rcx
 # CHECK-NEXT:  2      10    1.00    *                   vcvtss2si (%rax), %ecx
 # CHECK-NEXT:  2      10    1.00    *                   vcvtss2si (%rax), %rcx
 # CHECK-NEXT:  1      3     1.00                        vcvttpd2dq  %xmm0, %xmm2
 # CHECK-NEXT:  2      10    1.00    *                   vcvttpd2dqx (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        vcvttpd2dq  %ymm0, %xmm2
 # CHECK-NEXT:  2      10    1.00    *                   vcvttpd2dqy (%rax), %xmm2
-# CHECK-NEXT:  1      5     1.00                        vcvttps2dq  %xmm0, %xmm2
+# CHECK-NEXT:  1      3     1.00                        vcvttps2dq  %xmm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   vcvttps2dq  (%rax), %xmm2
 # CHECK-NEXT:  1      5     1.00                        vcvttps2dq  %ymm0, %ymm2
 # CHECK-NEXT:  1      12    1.00    *                   vcvttps2dq  (%rax), %ymm2
 # CHECK-NEXT:  1      4     1.00                        vcvttsd2si  %xmm0, %ecx
 # CHECK-NEXT:  1      4     1.00                        vcvttsd2si  %xmm0, %rcx
 # CHECK-NEXT:  1      11    1.00    *                   vcvttsd2si  (%rax), %ecx
 # CHECK-NEXT:  1      11    1.00    *                   vcvttsd2si  (%rax), %rcx
 # CHECK-NEXT:  1      3     1.00                        vcvttss2si  %xmm0, %ecx

 # CHECK-NEXT:  1      10    10.00                       vdivps  %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  2      17    17.00   *                   vdivps  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      15    1.00                        vdivsd  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      22    1.00    *                   vdivsd  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      15    1.00                        vdivss  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      22    1.00    *                   vdivss  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      100   0.25                        vdppd $22, %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      100   0.25    *                   vdppd $22, (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25                        vdpps $22, %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   vdpps $22, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      15    0.25                        vdpps $22, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  2      19    0.33    *                   vdpps $22, (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      100   0.25                        vdpps $22, %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      100   0.25    *                   vdpps $22, (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.33                        vextractf128  $1, %ymm0, %xmm2
 # CHECK-NEXT:  2      8     0.33           *            vextractf128  $1, %ymm0, (%rax)
 # CHECK-NEXT:  1      2     2.00                        vextractps  $1, %xmm0, %ecx
 # CHECK-NEXT:  2      5     2.00           *            vextractps  $1, %xmm0, (%rax)
-# CHECK-NEXT:  1      100   0.25                        vhaddpd %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   vhaddpd (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25                        vhaddpd %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25    *                   vhaddpd (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25                        vhaddps %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   vhaddps (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25                        vhaddps %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25    *                   vhaddps (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25                        vhsubpd %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   vhsubpd (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25                        vhsubpd %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25    *                   vhsubpd (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25                        vhsubps %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   vhsubps (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25                        vhsubps %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25    *                   vhsubps (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      7     0.25                        vhaddpd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  2      11    0.33    *                   vhaddpd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      7     0.25                        vhaddpd %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      11    0.33    *                   vhaddpd (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      7     0.25                        vhaddps %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  2      11    0.33    *                   vhaddps (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      7     0.25                        vhaddps %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      11    0.33    *                   vhaddps (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      7     0.25                        vhsubpd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  2      11    0.33    *                   vhsubpd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      7     0.25                        vhsubpd %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      11    0.33    *                   vhsubpd (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      7     0.25                        vhsubps %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  2      11    0.33    *                   vhsubps (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      7     0.25                        vhsubps %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      11    0.33    *                   vhsubps (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      2     0.33                        vinsertf128 $1, %xmm0, %ymm1, %ymm2
 # CHECK-NEXT:  2      9     0.33    *                   vinsertf128 $1, (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.50                        vinsertps $1, %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     0.50    *                   vinsertps $1, (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   vlddqu  (%rax), %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   vlddqu  (%rax), %ymm2
 # CHECK-NEXT:  1      100   0.25    *             U     vldmxcsr  (%rax)
 # CHECK-NEXT:  1      100   0.25    *      *      U     vmaskmovdqu %xmm0, %xmm1
 # CHECK-NEXT:  1      8     0.50    *                   vmaskmovpd  (%rax), %xmm0, %xmm2
 # CHECK-NEXT:  2      8     0.50    *                   vmaskmovpd  (%rax), %ymm0, %ymm2
 # CHECK-NEXT:  1      4     0.50    *      *            vmaskmovpd  %xmm0, %xmm1, (%rax)
 # CHECK-NEXT:  2      5     1.00    *      *            vmaskmovpd  %ymm0, %ymm1, (%rax)
 # CHECK-NEXT:  1      8     0.50    *                   vmaskmovps  (%rax), %xmm0, %xmm2
 # CHECK-NEXT:  2      8     0.50    *                   vmaskmovps  (%rax), %ymm0, %ymm2
 # CHECK-NEXT:  1      4     0.50    *      *            vmaskmovps  %xmm0, %xmm1, (%rax)
 # CHECK-NEXT:  2      5     1.00    *      *            vmaskmovps  %ymm0, %ymm1, (%rax)
-# CHECK-NEXT:  1      3     1.00                        vmaxpd  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   vmaxpd  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      3     1.00                        vmaxpd  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      10    1.00    *                   vmaxpd  (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      3     1.00                        vmaxps  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   vmaxps  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      3     1.00                        vmaxps  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      10    1.00    *                   vmaxps  (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      3     1.00                        vmaxsd  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   vmaxsd  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      3     1.00                        vmaxss  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   vmaxss  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      3     1.00                        vminpd  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   vminpd  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      3     1.00                        vminpd  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      10    1.00    *                   vminpd  (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      3     1.00                        vminps  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   vminps  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      3     1.00                        vminps  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      10    1.00    *                   vminps  (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      3     1.00                        vminsd  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   vminsd  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      3     1.00                        vminss  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   vminss  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00                        vmaxpd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   vmaxpd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00                        vmaxpd  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vmaxpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vmaxps  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   vmaxps  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00                        vmaxps  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vmaxps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vmaxsd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   vmaxsd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00                        vmaxss  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   vmaxss  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00                        vminpd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   vminpd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00                        vminpd  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vminpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vminps  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   vminps  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00                        vminps  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vminps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vminsd  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   vminsd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      1     1.00                        vminss  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   vminss  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vmovapd %xmm0, %xmm2
 # CHECK-NEXT:  1      1     0.33           *            vmovapd %xmm0, (%rax)
 # CHECK-NEXT:  1      8     0.33    *                   vmovapd (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        vmovapd %ymm0, %ymm2
 # CHECK-NEXT:  1      1     0.33           *            vmovapd %ymm0, (%rax)
 # CHECK-NEXT:  1      8     0.33    *                   vmovapd (%rax), %ymm2
 # CHECK-NEXT:  1      1     0.25                        vmovaps %xmm0, %xmm2
 # CHECK-NEXT:  1      1     0.33           *            vmovaps %xmm0, (%rax)

 # CHECK-NEXT:  1      8     0.33    *                   vmovups (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        vmovups %ymm0, %ymm2
 # CHECK-NEXT:  1      1     0.33           *            vmovups %ymm0, (%rax)
 # CHECK-NEXT:  1      8     0.33    *                   vmovups (%rax), %ymm2
 # CHECK-NEXT:  1      100   0.25                        vmpsadbw  $1, %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      100   0.25    *                   vmpsadbw  $1, (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      3     0.50                        vmulpd  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  2      10    0.50    *                   vmulpd  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      4     0.50                        vmulpd  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  2      11    0.50    *                   vmulpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.50                        vmulpd  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      10    0.50    *                   vmulpd  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      3     0.50                        vmulps  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  2      10    0.50    *                   vmulps  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      4     0.50                        vmulps  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  2      11    0.50    *                   vmulps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.50                        vmulps  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      10    0.50    *                   vmulps  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      3     0.50                        vmulsd  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  2      10    0.50    *                   vmulsd  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      3     0.50                        vmulss  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  2      10    0.50    *                   vmulss  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vorpd %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   vorpd (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vorpd %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vorpd (%rax), %ymm1, %ymm2

 # CHECK-NEXT:  1      100   0.25                        vpcmpistri  $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      100   0.25    *                   vpcmpistri  $1, (%rax), %xmm2
 # CHECK-NEXT:  1      100   0.25                        vpcmpistrm  $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      100   0.25    *                   vpcmpistrm  $1, (%rax), %xmm2
 # CHECK-NEXT:  1      100   0.25                        vperm2f128  $1, %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      100   0.25    *                   vperm2f128  $1, (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.50                        vpermilpd $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.50    *                   vpermilpd $1, (%rax), %xmm2
-# CHECK-NEXT:  1      1     0.50                        vpermilpd %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      8     0.50    *                   vpermilpd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     0.50                        vpermilpd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      10    0.50    *                   vpermilpd (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.50                        vpermilpd $1, %ymm0, %ymm2
 # CHECK-NEXT:  1      8     0.50    *                   vpermilpd $1, (%rax), %ymm2
-# CHECK-NEXT:  1      1     0.50                        vpermilpd %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      8     0.50    *                   vpermilpd (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.50                        vpermilpd %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      10    0.50    *                   vpermilpd (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.50                        vpermilps $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.50    *                   vpermilps $1, (%rax), %xmm2
-# CHECK-NEXT:  1      1     0.50                        vpermilps %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      8     0.50    *                   vpermilps (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     0.50                        vpermilps %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      10    0.50    *                   vpermilps (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.50                        vpermilps $1, %ymm0, %ymm2
 # CHECK-NEXT:  1      8     0.50    *                   vpermilps $1, (%rax), %ymm2
-# CHECK-NEXT:  1      1     0.50                        vpermilps %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      8     0.50    *                   vpermilps (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.50                        vpermilps %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      10    0.50    *                   vpermilps (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      2     2.00                        vpextrb $1, %xmm0, %ecx
 # CHECK-NEXT:  2      5     3.00           *            vpextrb $1, %xmm0, (%rax)
 # CHECK-NEXT:  1      2     2.00                        vpextrd $1, %xmm0, %ecx
 # CHECK-NEXT:  2      5     3.00           *            vpextrd $1, %xmm0, (%rax)
 # CHECK-NEXT:  1      2     2.00                        vpextrq $1, %xmm0, %rcx
 # CHECK-NEXT:  2      5     3.00           *            vpextrq $1, %xmm0, (%rax)
 # CHECK-NEXT:  1      2     2.00                        vpextrw $1, %xmm0, %ecx
 # CHECK-NEXT:  2      5     3.00           *            vpextrw $1, %xmm0, (%rax)
-# CHECK-NEXT:  1      100   0.25                        vphaddd %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   vphaddd (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25                        vphaddsw  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   vphaddsw  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25                        vphaddw %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   vphaddw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     0.25                        vphaddd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  2      7     0.33    *                   vphaddd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     0.25                        vphaddsw  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  2      7     0.33    *                   vphaddsw  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     0.25                        vphaddw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  2      7     0.33    *                   vphaddw (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      4     1.00                        vphminposuw %xmm0, %xmm2
 # CHECK-NEXT:  1      11    1.00    *                   vphminposuw (%rax), %xmm2
-# CHECK-NEXT:  1      100   0.25                        vphsubd %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   vphsubd (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25                        vphsubsw  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   vphsubsw  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25                        vphsubw %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   vphsubw (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     0.25                        vphsubd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  2      7     0.33    *                   vphsubd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     0.25                        vphsubsw  %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  2      7     0.33    *                   vphsubsw  (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     0.25                        vphsubw %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  2      7     0.33    *                   vphsubw (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpinsrb $1, %eax, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   vpinsrb $1, (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpinsrd $1, %eax, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   vpinsrd $1, (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpinsrq $1, %rax, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   vpinsrq $1, (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpinsrw $1, %eax, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   vpinsrw $1, (%rax), %xmm1, %xmm2

 # CHECK-NEXT:  1      8     0.33    *                   vpsignb (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpsignd %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   vpsignd (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpsignw %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   vpsignw (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpslld  $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      1     1.00                        vpslld  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   vpslld  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      1     1.00                        vpslldq $1, %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                        vpslldq $1, %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpsllq  $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      1     1.00                        vpsllq  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   vpsllq  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpsllw  $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      1     1.00                        vpsllw  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   vpsllw  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpsrad  $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      1     1.00                        vpsrad  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   vpsrad  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpsraw  $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      1     1.00                        vpsraw  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   vpsraw  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpsrld  $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      1     1.00                        vpsrld  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   vpsrld  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      1     1.00                        vpsrldq $1, %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                        vpsrldq $1, %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpsrlq  $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      1     1.00                        vpsrlq  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   vpsrlq  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpsrlw  $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      1     1.00                        vpsrlw  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   vpsrlw  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        vpsubb  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   vpsubb  (%rax), %xmm1, %xmm2

 # CHECK-NEXT:  1      1     0.25                        vpxor %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   vpxor (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      5     0.50                        vrcpps  %xmm0, %xmm2
 # CHECK-NEXT:  1      12    0.50    *                   vrcpps  (%rax), %xmm2
 # CHECK-NEXT:  1      5     0.50                        vrcpps  %ymm0, %ymm2
 # CHECK-NEXT:  3      12    0.50    *                   vrcpps  (%rax), %ymm2
 # CHECK-NEXT:  1      5     0.50                        vrcpss  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      12    0.50    *                   vrcpss  (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      4     1.00                        vroundpd  $1, %xmm0, %xmm2
-# CHECK-NEXT:  1      11    1.00    *                   vroundpd  $1, (%rax), %xmm2
-# CHECK-NEXT:  1      4     1.00                        vroundpd  $1, %ymm0, %ymm2
-# CHECK-NEXT:  1      11    1.00    *                   vroundpd  $1, (%rax), %ymm2
-# CHECK-NEXT:  1      4     1.00                        vroundps  $1, %xmm0, %xmm2
-# CHECK-NEXT:  1      11    1.00    *                   vroundps  $1, (%rax), %xmm2
-# CHECK-NEXT:  1      4     1.00                        vroundps  $1, %ymm0, %ymm2
-# CHECK-NEXT:  1      11    1.00    *                   vroundps  $1, (%rax), %ymm2
-# CHECK-NEXT:  1      4     1.00                        vroundsd  $1, %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      11    1.00    *                   vroundsd  $1, (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      4     1.00                        vroundss  $1, %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      11    1.00    *                   vroundss  $1, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                        vroundpd  $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      10    1.00    *                   vroundpd  $1, (%rax), %xmm2
+# CHECK-NEXT:  1      3     1.00                        vroundpd  $1, %ymm0, %ymm2
+# CHECK-NEXT:  1      10    1.00    *                   vroundpd  $1, (%rax), %ymm2
+# CHECK-NEXT:  1      3     1.00                        vroundps  $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      10    1.00    *                   vroundps  $1, (%rax), %xmm2
+# CHECK-NEXT:  1      3     1.00                        vroundps  $1, %ymm0, %ymm2
+# CHECK-NEXT:  1      10    1.00    *                   vroundps  $1, (%rax), %ymm2
+# CHECK-NEXT:  1      3     1.00                        vroundsd  $1, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      10    1.00    *                   vroundsd  $1, (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     1.00                        vroundss  $1, %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      10    1.00    *                   vroundss  $1, (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      5     0.50                        vrsqrtps  %xmm0, %xmm2
 # CHECK-NEXT:  2      12    0.50    *                   vrsqrtps  (%rax), %xmm2
 # CHECK-NEXT:  2      5     1.00                        vrsqrtps  %ymm0, %ymm2
 # CHECK-NEXT:  2      12    0.50    *                   vrsqrtps  (%rax), %ymm2
 # CHECK-NEXT:  1      5     0.50                        vrsqrtss  %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  2      12    1.00    *                   vrsqrtss  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  1      1     0.50                        vshufpd $1, %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  1      8     0.50    *                   vshufpd $1, (%rax), %xmm1, %xmm2

 # CHECK-NEXT: [8]   - Zn2FPU0
 # CHECK-NEXT: [9]   - Zn2FPU1
 # CHECK-NEXT: [10]  - Zn2FPU2
 # CHECK-NEXT: [11]  - Zn2FPU3
 # CHECK-NEXT: [12]  - Zn2Multiplier
 
 # CHECK:      Resource pressure per iteration:
 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]
-# CHECK-NEXT: 112.00 112.00 112.00 0.25   0.25   0.25   0.25    -     191.92 141.92 168.75 455.42  -
+# CHECK-NEXT: 117.00 117.00 117.00 0.25   0.25   0.25   0.25    -     191.92 143.42 170.25 455.42  -
 
 # CHECK:      Resource pressure by instruction:
 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]   Instructions:
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     vaddpd %xmm0, %xmm1, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     vaddpd (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     vaddpd %ymm0, %ymm1, %ymm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     vaddpd (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     vaddps %xmm0, %xmm1, %xmm2

 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     vcomisd  %xmm0, %xmm1
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     vcomisd  (%rax), %xmm1
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     vcomiss  %xmm0, %xmm1
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     vcomiss  (%rax), %xmm1
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     vcvtdq2pd  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     vcvtdq2pd  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     vcvtdq2pd  %xmm0, %ymm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     vcvtdq2pd  (%rax), %ymm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     vcvtdq2ps  %xmm0, %xmm2
+# CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     vcvtdq2ps  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     vcvtdq2ps  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     vcvtdq2ps  %ymm0, %ymm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     vcvtdq2ps  (%rax), %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     vcvtpd2dq  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     0.50   0.50   1.00    -     vcvtpd2dqx (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     vcvtpd2dq  %ymm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     0.50   0.50   1.00    -     vcvtpd2dqy (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     vcvtpd2ps  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.50    -      -     0.50    -     vcvtpd2psx (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     vcvtpd2ps  %ymm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     vcvtpd2psy (%rax), %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     vcvtps2dq  %xmm0, %xmm2
+# CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     vcvtps2dq  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     vcvtps2dq  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     vcvtps2dq  %ymm0, %ymm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     vcvtps2dq  (%rax), %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     vcvtps2pd  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     vcvtps2pd  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     vcvtps2pd  %xmm0, %ymm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     vcvtps2pd  (%rax), %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -     1.00   1.00    -     vcvtsd2si  %xmm0, %ecx

 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     vcvtss2si  %xmm0, %ecx
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     vcvtss2si  %xmm0, %rcx
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     0.50   0.50   1.00    -     vcvtss2si  (%rax), %ecx
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     0.50   0.50   1.00    -     vcvtss2si  (%rax), %rcx
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     vcvttpd2dq %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     0.50   0.50   1.00    -     vcvttpd2dqx  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     vcvttpd2dq %ymm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     0.50   0.50   1.00    -     vcvttpd2dqy  (%rax), %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     vcvttps2dq %xmm0, %xmm2
+# CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     vcvttps2dq %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     vcvttps2dq (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     vcvttps2dq %ymm0, %ymm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     vcvttps2dq (%rax), %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -     1.00   1.00    -     vcvttsd2si %xmm0, %ecx
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -     1.00   1.00    -     vcvttsd2si %xmm0, %rcx
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -     1.00   1.00    -     vcvttsd2si (%rax), %ecx
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -     1.00   1.00    -     vcvttsd2si (%rax), %rcx
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     vcvttss2si %xmm0, %ecx

 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     17.00   -     vdivps (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     vdivsd %xmm0, %xmm1, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     vdivsd (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     vdivss %xmm0, %xmm1, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     vdivss (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vdppd  $22, %xmm0, %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vdppd  $22, (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vdpps  $22, %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vdpps  $22, (%rax), %xmm1, %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vdpps  $22, (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vdpps  $22, %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vdpps  $22, (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.33   0.33    -     0.33    -     vextractf128 $1, %ymm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.33   0.33    -     0.33    -     vextractf128 $1, %ymm0, (%rax)
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   2.50    -      -     vextractps $1, %xmm0, %ecx
 # CHECK-NEXT: 1.67   1.67   1.67    -      -      -      -      -      -     0.50   2.50    -      -     vextractps $1, %xmm0, (%rax)
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhaddpd  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhaddpd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vhaddpd  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhaddpd  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhaddpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vhaddpd  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhaddps  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhaddps  (%rax), %xmm1, %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vhaddps  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhaddps  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhaddps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vhaddps  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhsubpd  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhsubpd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vhsubpd  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhsubpd  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhsubpd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vhsubpd  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhsubps  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhsubps  (%rax), %xmm1, %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vhsubps  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhsubps  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vhsubps  (%rax), %ymm1, %ymm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vhsubps  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.33   0.33    -     0.33    -     vinsertf128  $1, %xmm0, %ymm1, %ymm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.33   0.33    -     0.33    -     vinsertf128  $1, (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50    -      -     vinsertps  $1, %xmm0, %xmm1, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     0.50   0.50    -      -     vinsertps  $1, (%rax), %xmm1, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vlddqu (%rax), %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vlddqu (%rax), %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vldmxcsr (%rax)
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vmaskmovdqu  %xmm0, %xmm1

 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     1.00   4.00    -      -     vpextrb  $1, %xmm0, (%rax)
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   2.50    -      -     vpextrd  $1, %xmm0, %ecx
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     1.00   4.00    -      -     vpextrd  $1, %xmm0, (%rax)
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   2.50    -      -     vpextrq  $1, %xmm0, %rcx
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     1.00   4.00    -      -     vpextrq  $1, %xmm0, (%rax)
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   2.50    -      -     vpextrw  $1, %xmm0, %ecx
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     1.00   4.00    -      -     vpextrw  $1, %xmm0, (%rax)
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphaddd  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphaddd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vphaddd  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphaddsw %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphaddsw (%rax), %xmm1, %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vphaddsw (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphaddw  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphaddw  (%rax), %xmm1, %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vphaddw  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     vphminposuw  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     vphminposuw  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphsubd  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphsubd  (%rax), %xmm1, %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vphsubd  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphsubsw %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphsubsw (%rax), %xmm1, %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vphsubsw (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphsubw  %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphsubw  (%rax), %xmm1, %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vphsubw  (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     vpinsrb  $1, %eax, %xmm1, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     vpinsrb  $1, (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     vpinsrd  $1, %eax, %xmm1, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     vpinsrd  $1, (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     vpinsrq  $1, %rax, %xmm1, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     vpinsrq  $1, (%rax), %xmm1, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     vpinsrw  $1, %eax, %xmm1, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     vpinsrw  $1, (%rax), %xmm1, %xmm2

llvm/test/tools/llvm-mca/X86/Znver2/resources-avx2.s

 # CHECK-NEXT:  1      100   0.25    *                   vpgatherdd  %xmm0, (%rax,%xmm1,2), %xmm2
 # CHECK-NEXT:  1      100   0.25    *                   vpgatherdd  %ymm0, (%rax,%ymm1,2), %ymm2
 # CHECK-NEXT:  1      100   0.25    *                   vpgatherdq  %xmm0, (%rax,%xmm1,2), %xmm2
 # CHECK-NEXT:  1      100   0.25    *                   vpgatherdq  %ymm0, (%rax,%xmm1,2), %ymm2
 # CHECK-NEXT:  1      100   0.25    *                   vpgatherqd  %xmm0, (%rax,%xmm1,2), %xmm2
 # CHECK-NEXT:  1      100   0.25    *                   vpgatherqd  %xmm0, (%rax,%ymm1,2), %xmm2
 # CHECK-NEXT:  1      100   0.25    *                   vpgatherqq  %xmm0, (%rax,%xmm1,2), %xmm2
 # CHECK-NEXT:  1      100   0.25    *                   vpgatherqq  %ymm0, (%rax,%ymm1,2), %ymm2
-# CHECK-NEXT:  1      100   0.25                        vphaddd %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25    *                   vphaddd (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25                        vphaddsw  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25    *                   vphaddsw  (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25                        vphaddw %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25    *                   vphaddw (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25                        vphsubd %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25    *                   vphsubd (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25                        vphsubsw  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25    *                   vphsubsw  (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25                        vphsubw %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25    *                   vphsubw (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.25                        vphaddd %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      7     0.33    *                   vphaddd (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.25                        vphaddsw  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      7     0.33    *                   vphaddsw  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.25                        vphaddw %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      7     0.33    *                   vphaddw (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.25                        vphsubd %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      7     0.33    *                   vphsubd (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.25                        vphsubsw  %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      7     0.33    *                   vphsubsw  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.25                        vphsubw %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      7     0.33    *                   vphsubw (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      4     1.00                        vpmaddubsw  %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      11    1.00    *                   vpmaddubsw  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      4     1.00                        vpmaddwd  %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      11    1.00    *                   vpmaddwd  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      100   0.25    *                   vpmaskmovd  (%rax), %xmm0, %xmm2
 # CHECK-NEXT:  1      100   0.25    *                   vpmaskmovd  (%rax), %ymm0, %ymm2
 # CHECK-NEXT:  1      100   0.25    *      *            vpmaskmovd  %xmm0, %xmm1, (%rax)
 # CHECK-NEXT:  1      100   0.25    *      *            vpmaskmovd  %ymm0, %ymm1, (%rax)

 # CHECK-NEXT:  1      8     0.33    *                   vpminsw (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpminub %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpminub (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpminud %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpminud (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpminuw %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpminuw (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  2      2     2.00                        vpmovmskb %ymm0, %ecx
-# CHECK-NEXT:  2      1     0.50                        vpmovsxbd %xmm0, %ymm2
+# CHECK-NEXT:  2      4     0.50                        vpmovsxbd %xmm0, %ymm2
 # CHECK-NEXT:  2      8     0.50    *                   vpmovsxbd (%rax), %ymm2
-# CHECK-NEXT:  2      1     0.50                        vpmovsxbq %xmm0, %ymm2
+# CHECK-NEXT:  2      4     0.50                        vpmovsxbq %xmm0, %ymm2
 # CHECK-NEXT:  2      8     0.50    *                   vpmovsxbq (%rax), %ymm2
-# CHECK-NEXT:  2      1     0.50                        vpmovsxbw %xmm0, %ymm2
+# CHECK-NEXT:  2      4     0.50                        vpmovsxbw %xmm0, %ymm2
 # CHECK-NEXT:  2      8     0.50    *                   vpmovsxbw (%rax), %ymm2
-# CHECK-NEXT:  2      1     0.50                        vpmovsxdq %xmm0, %ymm2
+# CHECK-NEXT:  2      4     0.50                        vpmovsxdq %xmm0, %ymm2
 # CHECK-NEXT:  2      8     0.50    *                   vpmovsxdq (%rax), %ymm2
-# CHECK-NEXT:  2      1     0.50                        vpmovsxwd %xmm0, %ymm2
+# CHECK-NEXT:  2      4     0.50                        vpmovsxwd %xmm0, %ymm2
 # CHECK-NEXT:  2      8     0.50    *                   vpmovsxwd (%rax), %ymm2
-# CHECK-NEXT:  2      1     0.50                        vpmovsxwq %xmm0, %ymm2
+# CHECK-NEXT:  2      4     0.50                        vpmovsxwq %xmm0, %ymm2
 # CHECK-NEXT:  2      8     0.50    *                   vpmovsxwq (%rax), %ymm2
-# CHECK-NEXT:  2      1     0.50                        vpmovzxbd %xmm0, %ymm2
+# CHECK-NEXT:  2      4     0.50                        vpmovzxbd %xmm0, %ymm2
 # CHECK-NEXT:  2      8     0.50    *                   vpmovzxbd (%rax), %ymm2
-# CHECK-NEXT:  2      1     0.50                        vpmovzxbq %xmm0, %ymm2
+# CHECK-NEXT:  2      4     0.50                        vpmovzxbq %xmm0, %ymm2
 # CHECK-NEXT:  2      8     0.50    *                   vpmovzxbq (%rax), %ymm2
-# CHECK-NEXT:  2      1     0.50                        vpmovzxbw %xmm0, %ymm2
+# CHECK-NEXT:  2      4     0.50                        vpmovzxbw %xmm0, %ymm2
 # CHECK-NEXT:  2      8     0.50    *                   vpmovzxbw (%rax), %ymm2
-# CHECK-NEXT:  2      1     0.50                        vpmovzxdq %xmm0, %ymm2
+# CHECK-NEXT:  2      4     0.50                        vpmovzxdq %xmm0, %ymm2
 # CHECK-NEXT:  2      8     0.50    *                   vpmovzxdq (%rax), %ymm2
-# CHECK-NEXT:  2      1     0.50                        vpmovzxwd %xmm0, %ymm2
+# CHECK-NEXT:  2      4     0.50                        vpmovzxwd %xmm0, %ymm2
 # CHECK-NEXT:  2      8     0.50    *                   vpmovzxwd (%rax), %ymm2
-# CHECK-NEXT:  2      1     0.50                        vpmovzxwq %xmm0, %ymm2
+# CHECK-NEXT:  2      4     0.50                        vpmovzxwq %xmm0, %ymm2
 # CHECK-NEXT:  2      8     0.50    *                   vpmovzxwq (%rax), %ymm2
 # CHECK-NEXT:  1      4     1.00                        vpmuldq %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      11    1.00    *                   vpmuldq (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      4     1.00                        vpmulhrsw %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      11    1.00    *                   vpmulhrsw (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      4     1.00                        vpmulhuw  %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      11    1.00    *                   vpmulhuw  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      4     1.00                        vpmulhw %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      11    1.00    *                   vpmulhw (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      3     1.00                        vpmulld %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  2      10    1.00    *                   vpmulld (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      4     1.00                        vpmulld %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  2      11    1.00    *                   vpmulld (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      4     1.00                        vpmullw %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      11    1.00    *                   vpmullw (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      4     1.00                        vpmuludq  %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      11    1.00    *                   vpmuludq  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpor  %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpor  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      3     1.00                        vpsadbw %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      10    1.00    *                   vpsadbw (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpshufb %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpshufb (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpshufd $1, %ymm0, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpshufd $1, (%rax), %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpshufhw  $1, %ymm0, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpshufhw  $1, (%rax), %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpshuflw  $1, %ymm0, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpshuflw  $1, (%rax), %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsignb %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpsignb (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsignd %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpsignd (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsignw %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpsignw (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpslld  $1, %ymm0, %ymm2
-# CHECK-NEXT:  1      2     1.00                        vpslld  %xmm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      9     1.00    *                   vpslld  (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      1     1.00                        vpslldq $1, %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vpslld  %xmm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vpslld  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     1.00                        vpslldq $1, %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsllq  $1, %ymm0, %ymm2
-# CHECK-NEXT:  1      2     1.00                        vpsllq  %xmm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      9     1.00    *                   vpsllq  (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      1     0.50                        vpsllvd %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      8     0.50    *                   vpsllvd (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      1     0.50                        vpsllvd %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      8     0.50    *                   vpsllvd (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      1     0.50                        vpsllvq %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      8     0.50    *                   vpsllvq (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      1     0.50                        vpsllvq %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      8     0.50    *                   vpsllvq (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vpsllq  %xmm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vpsllq  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.50                        vpsllvd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      10    0.50    *                   vpsllvd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     0.50                        vpsllvd %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      10    0.50    *                   vpsllvd (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.50                        vpsllvq %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      10    0.50    *                   vpsllvq (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     0.50                        vpsllvq %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      10    0.50    *                   vpsllvq (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsllw  $1, %ymm0, %ymm2
-# CHECK-NEXT:  1      2     1.00                        vpsllw  %xmm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      9     1.00    *                   vpsllw  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vpsllw  %xmm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vpsllw  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsrad  $1, %ymm0, %ymm2
-# CHECK-NEXT:  1      2     1.00                        vpsrad  %xmm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      9     1.00    *                   vpsrad  (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      1     0.50                        vpsravd %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      8     0.50    *                   vpsravd (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      1     0.50                        vpsravd %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      8     0.50    *                   vpsravd (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vpsrad  %xmm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vpsrad  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.50                        vpsravd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      10    0.50    *                   vpsravd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     0.50                        vpsravd %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      10    0.50    *                   vpsravd (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsraw  $1, %ymm0, %ymm2
-# CHECK-NEXT:  1      2     1.00                        vpsraw  %xmm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      9     1.00    *                   vpsraw  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vpsraw  %xmm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vpsraw  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsrld  $1, %ymm0, %ymm2
-# CHECK-NEXT:  1      2     1.00                        vpsrld  %xmm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      9     1.00    *                   vpsrld  (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      1     1.00                        vpsrldq $1, %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vpsrld  %xmm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vpsrld  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     1.00                        vpsrldq $1, %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsrlq  $1, %ymm0, %ymm2
-# CHECK-NEXT:  1      2     1.00                        vpsrlq  %xmm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      9     1.00    *                   vpsrlq  (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      1     0.50                        vpsrlvd %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      8     0.50    *                   vpsrlvd (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      1     0.50                        vpsrlvd %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      8     0.50    *                   vpsrlvd (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      1     0.50                        vpsrlvq %xmm0, %xmm1, %xmm2
-# CHECK-NEXT:  1      8     0.50    *                   vpsrlvq (%rax), %xmm1, %xmm2
-# CHECK-NEXT:  1      1     0.50                        vpsrlvq %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      8     0.50    *                   vpsrlvq (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vpsrlq  %xmm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vpsrlq  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.50                        vpsrlvd %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      10    0.50    *                   vpsrlvd (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     0.50                        vpsrlvd %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      10    0.50    *                   vpsrlvd (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      3     0.50                        vpsrlvq %xmm0, %xmm1, %xmm2
+# CHECK-NEXT:  1      10    0.50    *                   vpsrlvq (%rax), %xmm1, %xmm2
+# CHECK-NEXT:  1      3     0.50                        vpsrlvq %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      10    0.50    *                   vpsrlvq (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsrlw  $1, %ymm0, %ymm2
-# CHECK-NEXT:  1      2     1.00                        vpsrlw  %xmm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      9     1.00    *                   vpsrlw  (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      1     1.00                        vpsrlw  %xmm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      8     1.00    *                   vpsrlw  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsubb  %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpsubb  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsubd  %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpsubd  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsubq  %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpsubq  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      1     0.25                        vpsubsb %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      8     0.33    *                   vpsubsb (%rax), %ymm1, %ymm2

 # CHECK-NEXT: [8]   - Zn2FPU0
 # CHECK-NEXT: [9]   - Zn2FPU1
 # CHECK-NEXT: [10]  - Zn2FPU2
 # CHECK-NEXT: [11]  - Zn2FPU3
 # CHECK-NEXT: [12]  - Zn2Multiplier
 
 # CHECK:      Resource pressure per iteration:
 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]
-# CHECK-NEXT: 42.67  42.67  42.67   -      -      -      -      -     70.17  75.17  85.00  42.67   -
+# CHECK-NEXT: 44.67  44.67  44.67   -      -      -      -      -     70.17  75.17  85.00  42.67   -
 
 # CHECK:      Resource pressure by instruction:
 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]   Instructions:
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     vbroadcasti128 (%rax), %ymm0
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     vbroadcastsd %xmm0, %ymm0
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     vbroadcastss %xmm0, %ymm0
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     vextracti128 $1, %ymm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vextracti128 $1, %ymm0, (%rax)

 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vpgatherdd %ymm0, (%rax,%ymm1,2), %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vpgatherdq %xmm0, (%rax,%xmm1,2), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vpgatherdq %ymm0, (%rax,%xmm1,2), %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vpgatherqd %xmm0, (%rax,%xmm1,2), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vpgatherqd %xmm0, (%rax,%ymm1,2), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vpgatherqq %xmm0, (%rax,%xmm1,2), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vpgatherqq %ymm0, (%rax,%ymm1,2), %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphaddd  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphaddd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vphaddd  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphaddsw %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphaddsw (%rax), %ymm1, %ymm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vphaddsw (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphaddw  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphaddw  (%rax), %ymm1, %ymm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vphaddw  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphsubd  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphsubd  (%rax), %ymm1, %ymm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vphsubd  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphsubsw %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphsubsw (%rax), %ymm1, %ymm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vphsubsw (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphsubw  %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vphsubw  (%rax), %ymm1, %ymm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     vphsubw  (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     vpmaddubsw %ymm0, %ymm1, %ymm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     vpmaddubsw (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     vpmaddwd %ymm0, %ymm1, %ymm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     vpmaddwd (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vpmaskmovd (%rax), %xmm0, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vpmaskmovd (%rax), %ymm0, %ymm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vpmaskmovd %xmm0, %xmm1, (%rax)
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     vpmaskmovd %ymm0, %ymm1, (%rax)

llvm/test/tools/llvm-mca/X86/Znver2/resources-sse1.s

 # CHECK-NEXT:  1      3     1.00                        addps %xmm0, %xmm2
 # CHECK-NEXT:  1      10    1.00    *                   addps (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        addss %xmm0, %xmm2
 # CHECK-NEXT:  1      10    1.00    *                   addss (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        andnps  %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   andnps  (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        andps %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   andps (%rax), %xmm2
-# CHECK-NEXT:  1      3     1.00                        cmpeqps %xmm0, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   cmpeqps (%rax), %xmm2
-# CHECK-NEXT:  1      3     1.00                        cmpeqss %xmm0, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   cmpeqss (%rax), %xmm2
+# CHECK-NEXT:  1      1     1.00                        cmpeqps %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   cmpeqps (%rax), %xmm2
+# CHECK-NEXT:  1      1     1.00                        cmpeqss %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   cmpeqss (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        comiss  %xmm0, %xmm1
 # CHECK-NEXT:  1      10    1.00    *                   comiss  (%rax), %xmm1
 # CHECK-NEXT:  1      5     1.00                        cvtpi2ps  %mm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   cvtpi2ps  (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        cvtps2pi  %xmm0, %mm2
 # CHECK-NEXT:  1      12    1.00    *                   cvtps2pi  (%rax), %mm2
 # CHECK-NEXT:  1      5     1.00                        cvtsi2ss  %ecx, %xmm2
 # CHECK-NEXT:  1      5     1.00                        cvtsi2ss  %rcx, %xmm2

 # CHECK-NEXT:  2      10    1.00    *                   cvttss2si (%rax), %ecx
 # CHECK-NEXT:  2      10    1.00    *                   cvttss2si (%rax), %rcx
 # CHECK-NEXT:  1      15    1.00                        divps %xmm0, %xmm2
 # CHECK-NEXT:  1      22    1.00    *                   divps (%rax), %xmm2
 # CHECK-NEXT:  1      15    1.00                        divss %xmm0, %xmm2
 # CHECK-NEXT:  1      22    1.00    *                   divss (%rax), %xmm2
 # CHECK-NEXT:  1      100   0.25    *             U     ldmxcsr (%rax)
 # CHECK-NEXT:  1      100   0.25    *      *      U     maskmovq  %mm0, %mm1
-# CHECK-NEXT:  1      3     1.00                        maxps %xmm0, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   maxps (%rax), %xmm2
-# CHECK-NEXT:  1      3     1.00                        maxss %xmm0, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   maxss (%rax), %xmm2
-# CHECK-NEXT:  1      3     1.00                        minps %xmm0, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   minps (%rax), %xmm2
-# CHECK-NEXT:  1      3     1.00                        minss %xmm0, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   minss (%rax), %xmm2
+# CHECK-NEXT:  1      1     1.00                        maxps %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   maxps (%rax), %xmm2
+# CHECK-NEXT:  1      1     1.00                        maxss %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   maxss (%rax), %xmm2
+# CHECK-NEXT:  1      1     1.00                        minps %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   minps (%rax), %xmm2
+# CHECK-NEXT:  1      1     1.00                        minss %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   minss (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        movaps  %xmm0, %xmm2
 # CHECK-NEXT:  1      1     0.33           *            movaps  %xmm0, (%rax)
 # CHECK-NEXT:  1      8     0.33    *                   movaps  (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.50                        movhlps %xmm0, %xmm2
 # CHECK-NEXT:  1      1     0.50                        movlhps %xmm0, %xmm2
 # CHECK-NEXT:  1      1     0.33           *            movhps  %xmm0, (%rax)
 # CHECK-NEXT:  1      8     0.50    *                   movhps  (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.33           *            movlps  %xmm0, (%rax)

llvm/test/tools/llvm-mca/X86/Znver2/resources-sse2.s

 # CHECK-NEXT:  1      10    1.00    *                   addpd (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        addsd %xmm0, %xmm2
 # CHECK-NEXT:  1      10    1.00    *                   addsd (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        andnpd  %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   andnpd  (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        andpd %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   andpd (%rax), %xmm2
 # CHECK-NEXT:  1      8     0.33    *      *      U     clflush (%rax)
-# CHECK-NEXT:  1      3     1.00                        cmpeqpd %xmm0, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   cmpeqpd (%rax), %xmm2
-# CHECK-NEXT:  1      3     1.00                        cmpeqsd %xmm0, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   cmpeqsd (%rax), %xmm2
+# CHECK-NEXT:  1      1     1.00                        cmpeqpd %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   cmpeqpd (%rax), %xmm2
+# CHECK-NEXT:  1      1     1.00                        cmpeqsd %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   cmpeqsd (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        comisd  %xmm0, %xmm1
 # CHECK-NEXT:  1      10    1.00    *                   comisd  (%rax), %xmm1
 # CHECK-NEXT:  1      3     1.00                        cvtdq2pd  %xmm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   cvtdq2pd  (%rax), %xmm2
-# CHECK-NEXT:  1      5     1.00                        cvtdq2ps  %xmm0, %xmm2
+# CHECK-NEXT:  1      3     1.00                        cvtdq2ps  %xmm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   cvtdq2ps  (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        cvtpd2dq  %xmm0, %xmm2
 # CHECK-NEXT:  2      10    1.00    *                   cvtpd2dq  (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        cvtpd2pi  %xmm0, %mm2
 # CHECK-NEXT:  1      12    1.00    *                   cvtpd2pi  (%rax), %mm2
 # CHECK-NEXT:  1      3     1.00                        cvtpd2ps  %xmm0, %xmm2
 # CHECK-NEXT:  2      10    0.50    *                   cvtpd2ps  (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        cvtpi2pd  %mm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   cvtpi2pd  (%rax), %xmm2
-# CHECK-NEXT:  1      5     1.00                        cvtps2dq  %xmm0, %xmm2
+# CHECK-NEXT:  1      3     1.00                        cvtps2dq  %xmm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   cvtps2dq  (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        cvtps2pd  %xmm0, %xmm2
 # CHECK-NEXT:  2      10    1.00    *                   cvtps2pd  (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        cvtsd2si  %xmm0, %ecx
 # CHECK-NEXT:  1      4     1.00                        cvtsd2si  %xmm0, %rcx
 # CHECK-NEXT:  1      11    1.00    *                   cvtsd2si  (%rax), %ecx
 # CHECK-NEXT:  1      11    1.00    *                   cvtsd2si  (%rax), %rcx
 # CHECK-NEXT:  1      3     1.00                        cvtsd2ss  %xmm0, %xmm2
 # CHECK-NEXT:  2      10    0.50    *                   cvtsd2ss  (%rax), %xmm2
-# CHECK-NEXT:  1      4     1.00                        cvtsi2sd  %ecx, %xmm2
-# CHECK-NEXT:  1      4     1.00                        cvtsi2sd  %rcx, %xmm2
+# CHECK-NEXT:  1      3     1.00                        cvtsi2sd  %ecx, %xmm2
+# CHECK-NEXT:  1      3     1.00                        cvtsi2sd  %rcx, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   cvtsi2sdl (%rax), %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   cvtsi2sdl (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        cvtss2sd  %xmm0, %xmm2
 # CHECK-NEXT:  2      10    2.00    *                   cvtss2sd  (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        cvttpd2dq %xmm0, %xmm2
 # CHECK-NEXT:  2      10    1.00    *                   cvttpd2dq (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        cvttpd2pi %xmm0, %mm2
 # CHECK-NEXT:  1      12    1.00    *                   cvttpd2pi (%rax), %mm2
-# CHECK-NEXT:  1      5     1.00                        cvttps2dq %xmm0, %xmm2
+# CHECK-NEXT:  1      3     1.00                        cvttps2dq %xmm0, %xmm2
 # CHECK-NEXT:  1      12    1.00    *                   cvttps2dq (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        cvttsd2si %xmm0, %ecx
 # CHECK-NEXT:  1      4     1.00                        cvttsd2si %xmm0, %rcx
 # CHECK-NEXT:  1      11    1.00    *                   cvttsd2si (%rax), %ecx
 # CHECK-NEXT:  1      11    1.00    *                   cvttsd2si (%rax), %rcx
 # CHECK-NEXT:  1      15    1.00                        divpd %xmm0, %xmm2
 # CHECK-NEXT:  1      22    1.00    *                   divpd (%rax), %xmm2
 # CHECK-NEXT:  1      15    1.00                        divsd %xmm0, %xmm2
 # CHECK-NEXT:  1      22    1.00    *                   divsd (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.33    *      *      U     lfence
 # CHECK-NEXT:  1      100   0.25    *      *      U     maskmovdqu  %xmm0, %xmm1
-# CHECK-NEXT:  1      3     1.00                        maxpd %xmm0, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   maxpd (%rax), %xmm2
-# CHECK-NEXT:  1      3     1.00                        maxsd %xmm0, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   maxsd (%rax), %xmm2
+# CHECK-NEXT:  1      1     1.00                        maxpd %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   maxpd (%rax), %xmm2
+# CHECK-NEXT:  1      1     1.00                        maxsd %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   maxsd (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.33    *      *      U     mfence
-# CHECK-NEXT:  1      3     1.00                        minpd %xmm0, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   minpd (%rax), %xmm2
-# CHECK-NEXT:  1      3     1.00                        minsd %xmm0, %xmm2
-# CHECK-NEXT:  1      10    1.00    *                   minsd (%rax), %xmm2
+# CHECK-NEXT:  1      1     1.00                        minpd %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   minpd (%rax), %xmm2
+# CHECK-NEXT:  1      1     1.00                        minsd %xmm0, %xmm2
+# CHECK-NEXT:  1      8     1.00    *                   minsd (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        movapd  %xmm0, %xmm2
 # CHECK-NEXT:  1      1     0.33           *            movapd  %xmm0, (%rax)
 # CHECK-NEXT:  1      8     0.33    *                   movapd  (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        movd  %eax, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   movd  (%rax), %xmm2
 # CHECK-NEXT:  1      2     1.00                        movd  %xmm0, %ecx
 # CHECK-NEXT:  1      1     0.33           *            movd  %xmm0, (%rax)
 # CHECK-NEXT:  1      1     0.25                        movdqa  %xmm0, %xmm2

 # CHECK-NEXT:  1      8     0.33    *                   pshufd  $1, (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        pshufhw $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   pshufhw $1, (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        pshuflw $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   pshuflw $1, (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        pslld $1, %xmm2
 # CHECK-NEXT:  1      1     1.00                        pslld %xmm0, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   pslld (%rax), %xmm2
-# CHECK-NEXT:  1      1     1.00                        pslldq  $1, %xmm2
+# CHECK-NEXT:  1      3     1.00                        pslldq  $1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        psllq $1, %xmm2
 # CHECK-NEXT:  1      1     1.00                        psllq %xmm0, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   psllq (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        psllw $1, %xmm2
 # CHECK-NEXT:  1      1     1.00                        psllw %xmm0, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   psllw (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        psrad $1, %xmm2
 # CHECK-NEXT:  1      1     1.00                        psrad %xmm0, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   psrad (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        psraw $1, %xmm2
 # CHECK-NEXT:  1      1     1.00                        psraw %xmm0, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   psraw (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        psrld $1, %xmm2
 # CHECK-NEXT:  1      1     1.00                        psrld %xmm0, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   psrld (%rax), %xmm2
-# CHECK-NEXT:  1      1     1.00                        psrldq  $1, %xmm2
+# CHECK-NEXT:  1      3     1.00                        psrldq  $1, %xmm2
 # CHECK-NEXT:  1      1     0.25                        psrlq $1, %xmm2
 # CHECK-NEXT:  1      1     1.00                        psrlq %xmm0, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   psrlq (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        psrlw $1, %xmm2
 # CHECK-NEXT:  1      1     1.00                        psrlw %xmm0, %xmm2
 # CHECK-NEXT:  1      8     1.00    *                   psrlw (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        psubb %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   psubb (%rax), %xmm2

 # CHECK-NEXT: [8]   - Zn2FPU0
 # CHECK-NEXT: [9]   - Zn2FPU1
 # CHECK-NEXT: [10]  - Zn2FPU2
 # CHECK-NEXT: [11]  - Zn2FPU3
 # CHECK-NEXT: [12]  - Zn2Multiplier
 
 # CHECK:      Resource pressure per iteration:
 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]
-# CHECK-NEXT: 44.33  44.33  44.33   -      -      -      -      -     71.92  40.42  71.75  152.92  -
+# CHECK-NEXT: 44.33  44.33  44.33   -      -      -      -      -     71.92  41.92  73.25  152.92  -
 
 # CHECK:      Resource pressure by instruction:
 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]   Instructions:
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     addpd  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     addpd  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     addsd  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     addsd  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     andnpd %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     andnpd (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     andpd  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     andpd  (%rax), %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     clflush  (%rax)
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     cmpeqpd  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     cmpeqpd  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     cmpeqsd  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     cmpeqsd  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     comisd %xmm0, %xmm1
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     comisd (%rax), %xmm1
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     cvtdq2pd %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     cvtdq2pd (%rax), %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     cvtdq2ps %xmm0, %xmm2
+# CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     cvtdq2ps %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     cvtdq2ps (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     cvtpd2dq %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     0.50   0.50   1.00    -     cvtpd2dq (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     cvtpd2pi %xmm0, %mm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     cvtpd2pi (%rax), %mm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     cvtpd2ps %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.50    -      -     0.50    -     cvtpd2ps (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     cvtpi2pd %mm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     cvtpi2pd (%rax), %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     cvtps2dq %xmm0, %xmm2
+# CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     cvtps2dq %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     cvtps2dq (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     cvtps2pd %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     cvtps2pd (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -     1.00   1.00    -     cvtsd2si %xmm0, %ecx
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -     1.00   1.00    -     cvtsd2si %xmm0, %rcx
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -     1.00   1.00    -     cvtsd2si (%rax), %ecx
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -     1.00   1.00    -     cvtsd2si (%rax), %rcx
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     cvtsd2ss %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.50    -      -     0.50    -     cvtsd2ss (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.33   0.33    -     1.33    -     cvtsi2sd %ecx, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.33   0.33    -     1.33    -     cvtsi2sd %rcx, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     cvtsi2sdl  (%rax), %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     cvtsi2sdl  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     cvtss2sd %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     2.00    -     cvtss2sd (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     cvttpd2dq  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     0.50   0.50   1.00    -     cvttpd2dq  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     cvttpd2pi  %xmm0, %mm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     cvttpd2pi  (%rax), %mm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     cvttps2dq  %xmm0, %xmm2
+# CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50   1.00    -     cvttps2dq  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     cvttps2dq  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -     1.00   1.00    -     cvttsd2si  %xmm0, %ecx
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -     1.00   1.00    -     cvttsd2si  %xmm0, %rcx
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -     1.00   1.00    -     cvttsd2si  (%rax), %ecx
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -     1.00   1.00    -     cvttsd2si  (%rax), %rcx
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     divpd  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -     1.00    -     divpd  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -     1.00    -     divsd  %xmm0, %xmm2

llvm/test/tools/llvm-mca/X86/Znver2/resources-sse3.s

 # CHECK-NEXT: [5]: MayStore
 # CHECK-NEXT: [6]: HasSideEffects (U)
 
 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 # CHECK-NEXT:  1      3     1.00                        addsubpd  %xmm0, %xmm2
 # CHECK-NEXT:  1      10    1.00    *                   addsubpd  (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00                        addsubps  %xmm0, %xmm2
 # CHECK-NEXT:  1      10    1.00    *                   addsubps  (%rax), %xmm2
-# CHECK-NEXT:  1      100   0.25                        haddpd  %xmm0, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   haddpd  (%rax), %xmm2
-# CHECK-NEXT:  1      100   0.25                        haddps  %xmm0, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   haddps  (%rax), %xmm2
-# CHECK-NEXT:  1      100   0.25                        hsubpd  %xmm0, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   hsubpd  (%rax), %xmm2
-# CHECK-NEXT:  1      100   0.25                        hsubps  %xmm0, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   hsubps  (%rax), %xmm2
+# CHECK-NEXT:  1      7     0.25                        haddpd  %xmm0, %xmm2
+# CHECK-NEXT:  2      11    0.33    *                   haddpd  (%rax), %xmm2
+# CHECK-NEXT:  1      7     0.25                        haddps  %xmm0, %xmm2
+# CHECK-NEXT:  2      11    0.33    *                   haddps  (%rax), %xmm2
+# CHECK-NEXT:  1      7     0.25                        hsubpd  %xmm0, %xmm2
+# CHECK-NEXT:  2      11    0.33    *                   hsubpd  (%rax), %xmm2
+# CHECK-NEXT:  1      7     0.25                        hsubps  %xmm0, %xmm2
+# CHECK-NEXT:  2      11    0.33    *                   hsubps  (%rax), %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   lddqu (%rax), %xmm2
 # CHECK-NEXT:  1      100   0.25                  U     monitor
 # CHECK-NEXT:  1      1     0.50                        movddup %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.50    *                   movddup (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.50                        movshdup  %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.50    *                   movshdup  (%rax), %xmm2
 # CHECK-NEXT:  1      100   0.25                        movsldup  %xmm0, %xmm2
 # CHECK-NEXT:  1      100   0.25    *                   movsldup  (%rax), %xmm2

 # CHECK-NEXT: [8]   - Zn2FPU0
 # CHECK-NEXT: [9]   - Zn2FPU1
 # CHECK-NEXT: [10]  - Zn2FPU2
 # CHECK-NEXT: [11]  - Zn2FPU3
 # CHECK-NEXT: [12]  - Zn2Multiplier
 
 # CHECK:      Resource pressure per iteration:
 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]
-# CHECK-NEXT: 1.67   1.67   1.67    -      -      -      -      -     4.00   2.00   2.00    -      -
+# CHECK-NEXT: 3.00   3.00   3.00    -      -      -      -      -     4.00   2.00   2.00    -      -
 
 # CHECK:      Resource pressure by instruction:
 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]   Instructions:
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     addsubpd %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     addsubpd (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     addsubps %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     addsubps (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     haddpd %xmm0, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     haddpd (%rax), %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     haddpd (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     haddps %xmm0, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     haddps (%rax), %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     haddps (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     hsubpd %xmm0, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     hsubpd (%rax), %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     hsubpd (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     hsubps %xmm0, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     hsubps (%rax), %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     hsubps (%rax), %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     lddqu  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     monitor
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50    -      -     movddup  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     0.50   0.50    -      -     movddup  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50    -      -     movshdup %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     0.50   0.50    -      -     movshdup (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     movsldup %xmm0, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     movsldup (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     mwait

llvm/test/tools/llvm-mca/X86/Znver2/resources-sse41.s

 # CHECK-NEXT:  1      1     0.50                        blendps $11, %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.50    *                   blendps $11, (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.50                        blendvpd  %xmm0, %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.50    *                   blendvpd  %xmm0, (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.50                        blendvps  %xmm0, %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.50    *                   blendvps  %xmm0, (%rax), %xmm2
 # CHECK-NEXT:  1      100   0.25                        dppd  $22, %xmm0, %xmm2
 # CHECK-NEXT:  1      100   0.25    *                   dppd  $22, (%rax), %xmm2
-# CHECK-NEXT:  1      100   0.25                        dpps  $22, %xmm0, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   dpps  $22, (%rax), %xmm2
+# CHECK-NEXT:  1      15    0.25                        dpps  $22, %xmm0, %xmm2
+# CHECK-NEXT:  2      19    0.33    *                   dpps  $22, (%rax), %xmm2
 # CHECK-NEXT:  1      2     2.00                        extractps $1, %xmm0, %ecx
 # CHECK-NEXT:  2      5     2.00           *            extractps $1, %xmm0, (%rax)
 # CHECK-NEXT:  1      1     0.50                        insertps  $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.50    *                   insertps  $1, (%rax), %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   movntdqa  (%rax), %xmm2
 # CHECK-NEXT:  1      100   0.25                        mpsadbw $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      100   0.25    *                   mpsadbw $1, (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        packusdw  %xmm0, %xmm2

 # CHECK-NEXT:  1      1     0.25                        pmovzxwq  %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   pmovzxwq  (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        pmuldq  %xmm0, %xmm2
 # CHECK-NEXT:  1      11    1.00    *                   pmuldq  (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        pmulld  %xmm0, %xmm2
 # CHECK-NEXT:  2      11    1.00    *                   pmulld  (%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        ptest %xmm0, %xmm1
 # CHECK-NEXT:  2      8     1.00    *                   ptest (%rax), %xmm1
-# CHECK-NEXT:  1      4     1.00                        roundpd $1, %xmm0, %xmm2
-# CHECK-NEXT:  1      11    1.00    *                   roundpd $1, (%rax), %xmm2
-# CHECK-NEXT:  1      4     1.00                        roundps $1, %xmm0, %xmm2
-# CHECK-NEXT:  1      11    1.00    *                   roundps $1, (%rax), %xmm2
-# CHECK-NEXT:  1      4     1.00                        roundsd $1, %xmm0, %xmm2
-# CHECK-NEXT:  1      11    1.00    *                   roundsd $1, (%rax), %xmm2
-# CHECK-NEXT:  1      4     1.00                        roundss $1, %xmm0, %xmm2
-# CHECK-NEXT:  1      11    1.00    *                   roundss $1, (%rax), %xmm2
+# CHECK-NEXT:  1      3     1.00                        roundpd $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      10    1.00    *                   roundpd $1, (%rax), %xmm2
+# CHECK-NEXT:  1      3     1.00                        roundps $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      10    1.00    *                   roundps $1, (%rax), %xmm2
+# CHECK-NEXT:  1      3     1.00                        roundsd $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      10    1.00    *                   roundsd $1, (%rax), %xmm2
+# CHECK-NEXT:  1      3     1.00                        roundss $1, %xmm0, %xmm2
+# CHECK-NEXT:  1      10    1.00    *                   roundss $1, (%rax), %xmm2
 
 # CHECK:      Resources:
 # CHECK-NEXT: [0]   - Zn2AGU0
 # CHECK-NEXT: [1]   - Zn2AGU1
 # CHECK-NEXT: [2]   - Zn2AGU2
 # CHECK-NEXT: [3]   - Zn2ALU0
 # CHECK-NEXT: [4]   - Zn2ALU1
 # CHECK-NEXT: [5]   - Zn2ALU2
 # CHECK-NEXT: [6]   - Zn2ALU3
 # CHECK-NEXT: [7]   - Zn2Divider
 # CHECK-NEXT: [8]   - Zn2FPU0
 # CHECK-NEXT: [9]   - Zn2FPU1
 # CHECK-NEXT: [10]  - Zn2FPU2
 # CHECK-NEXT: [11]  - Zn2FPU3
 # CHECK-NEXT: [12]  - Zn2Multiplier
 
 # CHECK:      Resource pressure per iteration:
 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]
-# CHECK-NEXT: 16.67  16.67  16.67   -      -      -      -      -     25.17  26.67  44.00  21.17   -
+# CHECK-NEXT: 17.00  17.00  17.00   -      -      -      -      -     25.17  26.67  44.00  21.17   -
 
 # CHECK:      Resource pressure by instruction:
 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]   Instructions:
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.50   0.50    -      -      -     blendpd  $11, %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.50   0.50    -      -      -     blendpd  $11, (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.50   0.50    -      -      -     blendps  $11, %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.50   0.50    -      -      -     blendps  $11, (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.50   0.50    -      -      -     blendvpd %xmm0, %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.50   0.50    -      -      -     blendvpd %xmm0, (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.50   0.50    -      -      -     blendvps %xmm0, %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.50   0.50    -      -      -     blendvps %xmm0, (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     dppd $22, %xmm0, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     dppd $22, (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     dpps $22, %xmm0, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     dpps $22, (%rax), %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     dpps $22, (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   2.50    -      -     extractps  $1, %xmm0, %ecx
 # CHECK-NEXT: 1.67   1.67   1.67    -      -      -      -      -      -     0.50   2.50    -      -     extractps  $1, %xmm0, (%rax)
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     0.50   0.50    -      -     insertps $1, %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -     0.50   0.50    -      -     insertps $1, (%rax), %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     movntdqa (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     mpsadbw  $1, %xmm0, %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     mpsadbw  $1, (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     packusdw %xmm0, %xmm2

llvm/test/tools/llvm-mca/X86/Znver2/resources-sse4a.s

 # CHECK-NEXT: [1]: #uOps
 # CHECK-NEXT: [2]: Latency
 # CHECK-NEXT: [3]: RThroughput
 # CHECK-NEXT: [4]: MayLoad
 # CHECK-NEXT: [5]: MayStore
 # CHECK-NEXT: [6]: HasSideEffects (U)
 
 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
-# CHECK-NEXT:  1      2     1.00                        extrq %xmm0, %xmm2
-# CHECK-NEXT:  1      2     1.00                        extrq $22, $2, %xmm2
+# CHECK-NEXT:  1      3     1.00                        extrq %xmm0, %xmm2
+# CHECK-NEXT:  1      3     1.00                        extrq $22, $2, %xmm2
 # CHECK-NEXT:  1      4     1.00                        insertq %xmm0, %xmm2
 # CHECK-NEXT:  1      4     1.00                        insertq $22, $22, %xmm0, %xmm2
 # CHECK-NEXT:  1      8     1.00           *            movntsd %xmm0, (%rax)
 # CHECK-NEXT:  1      8     1.00           *            movntss %xmm0, (%rax)
 
 # CHECK:      Resources:
 # CHECK-NEXT: [0]   - Zn2AGU0
 # CHECK-NEXT: [1]   - Zn2AGU1

llvm/test/tools/llvm-mca/X86/Znver2/resources-ssse3.s

 # CHECK-NEXT:  1      1     0.25                        pabsw %mm0, %mm2
 # CHECK-NEXT:  1      8     0.33    *                   pabsw (%rax), %mm2
 # CHECK-NEXT:  1      1     0.25                        pabsw %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   pabsw (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.25                        palignr $1, %mm0, %mm2
 # CHECK-NEXT:  1      8     0.33    *                   palignr $1, (%rax), %mm2
 # CHECK-NEXT:  1      1     0.25                        palignr $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      8     0.33    *                   palignr $1, (%rax), %xmm2
-# CHECK-NEXT:  1      100   0.25                        phaddd  %mm0, %mm2
-# CHECK-NEXT:  1      100   0.25    *                   phaddd  (%rax), %mm2
-# CHECK-NEXT:  1      100   0.25                        phaddd  %xmm0, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   phaddd  (%rax), %xmm2
-# CHECK-NEXT:  1      100   0.25                        phaddsw %mm0, %mm2
-# CHECK-NEXT:  1      100   0.25    *                   phaddsw (%rax), %mm2
-# CHECK-NEXT:  1      100   0.25                        phaddsw %xmm0, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   phaddsw (%rax), %xmm2
-# CHECK-NEXT:  1      100   0.25                        phaddw  %mm0, %mm2
-# CHECK-NEXT:  1      100   0.25    *                   phaddw  (%rax), %mm2
-# CHECK-NEXT:  1      100   0.25                        phaddw  %xmm0, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   phaddw  (%rax), %xmm2
-# CHECK-NEXT:  1      100   0.25                        phsubd  %mm0, %mm2
-# CHECK-NEXT:  1      100   0.25    *                   phsubd  (%rax), %mm2
-# CHECK-NEXT:  1      100   0.25                        phsubd  %xmm0, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   phsubd  (%rax), %xmm2
-# CHECK-NEXT:  1      100   0.25                        phsubsw %mm0, %mm2
-# CHECK-NEXT:  1      100   0.25    *                   phsubsw (%rax), %mm2
-# CHECK-NEXT:  1      100   0.25                        phsubsw %xmm0, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   phsubsw (%rax), %xmm2
-# CHECK-NEXT:  1      100   0.25                        phsubw  %mm0, %mm2
-# CHECK-NEXT:  1      100   0.25    *                   phsubw  (%rax), %mm2
-# CHECK-NEXT:  1      100   0.25                        phsubw  %xmm0, %xmm2
-# CHECK-NEXT:  1      100   0.25    *                   phsubw  (%rax), %xmm2
+# CHECK-NEXT:  1      3     0.25                        phaddd  %mm0, %mm2
+# CHECK-NEXT:  2      7     0.33    *                   phaddd  (%rax), %mm2
+# CHECK-NEXT:  1      3     0.25                        phaddd  %xmm0, %xmm2
+# CHECK-NEXT:  2      7     0.33    *                   phaddd  (%rax), %xmm2
+# CHECK-NEXT:  1      3     0.25                        phaddsw %mm0, %mm2
+# CHECK-NEXT:  2      7     0.33    *                   phaddsw (%rax), %mm2
+# CHECK-NEXT:  1      3     0.25                        phaddsw %xmm0, %xmm2
+# CHECK-NEXT:  2      7     0.33    *                   phaddsw (%rax), %xmm2
+# CHECK-NEXT:  1      3     0.25                        phaddw  %mm0, %mm2
+# CHECK-NEXT:  2      7     0.33    *                   phaddw  (%rax), %mm2
+# CHECK-NEXT:  1      3     0.25                        phaddw  %xmm0, %xmm2
+# CHECK-NEXT:  2      7     0.33    *                   phaddw  (%rax), %xmm2
+# CHECK-NEXT:  1      3     0.25                        phsubd  %mm0, %mm2
+# CHECK-NEXT:  2      7     0.33    *                   phsubd  (%rax), %mm2
+# CHECK-NEXT:  1      3     0.25                        phsubd  %xmm0, %xmm2
+# CHECK-NEXT:  2      7     0.33    *                   phsubd  (%rax), %xmm2
+# CHECK-NEXT:  1      3     0.25                        phsubsw %mm0, %mm2
+# CHECK-NEXT:  2      7     0.33    *                   phsubsw (%rax), %mm2
+# CHECK-NEXT:  1      3     0.25                        phsubsw %xmm0, %xmm2
+# CHECK-NEXT:  2      7     0.33    *                   phsubsw (%rax), %xmm2
+# CHECK-NEXT:  1      3     0.25                        phsubw  %mm0, %mm2
+# CHECK-NEXT:  2      7     0.33    *                   phsubw  (%rax), %mm2
+# CHECK-NEXT:  1      3     0.25                        phsubw  %xmm0, %xmm2
+# CHECK-NEXT:  2      7     0.33    *                   phsubw  (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        pmaddubsw %mm0, %mm2
 # CHECK-NEXT:  1      11    1.00    *                   pmaddubsw (%rax), %mm2
 # CHECK-NEXT:  1      4     1.00                        pmaddubsw %xmm0, %xmm2
 # CHECK-NEXT:  1      11    1.00    *                   pmaddubsw (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        pmulhrsw  %mm0, %mm2
 # CHECK-NEXT:  1      11    1.00    *                   pmulhrsw  (%rax), %mm2
 # CHECK-NEXT:  1      4     1.00                        pmulhrsw  %xmm0, %xmm2
 # CHECK-NEXT:  1      11    1.00    *                   pmulhrsw  (%rax), %xmm2

 # CHECK-NEXT: [8]   - Zn2FPU0
 # CHECK-NEXT: [9]   - Zn2FPU1
 # CHECK-NEXT: [10]  - Zn2FPU2
 # CHECK-NEXT: [11]  - Zn2FPU3
 # CHECK-NEXT: [12]  - Zn2Multiplier
 
 # CHECK:      Resource pressure per iteration:
 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]
-# CHECK-NEXT: 6.67   6.67   6.67    -      -      -      -      -     16.00  8.00   8.00   8.00    -
+# CHECK-NEXT: 10.67  10.67  10.67   -      -      -      -      -     16.00  8.00   8.00   8.00    -
 
 # CHECK:      Resource pressure by instruction:
 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]   Instructions:
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     pabsb  %mm0, %mm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     pabsb  (%rax), %mm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     pabsb  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     pabsb  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     pabsd  %mm0, %mm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     pabsd  (%rax), %mm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     pabsd  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     pabsd  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     pabsw  %mm0, %mm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     pabsw  (%rax), %mm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     pabsw  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     pabsw  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     palignr  $1, %mm0, %mm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     palignr  $1, (%rax), %mm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.25   0.25   0.25   0.25    -     palignr  $1, %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     0.25   0.25   0.25   0.25    -     palignr  $1, (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phaddd %mm0, %mm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phaddd (%rax), %mm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     phaddd (%rax), %mm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phaddd %xmm0, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phaddd (%rax), %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     phaddd (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phaddsw  %mm0, %mm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phaddsw  (%rax), %mm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     phaddsw  (%rax), %mm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phaddsw  %xmm0, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phaddsw  (%rax), %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     phaddsw  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phaddw %mm0, %mm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phaddw (%rax), %mm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     phaddw (%rax), %mm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phaddw %xmm0, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phaddw (%rax), %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     phaddw (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phsubd %mm0, %mm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phsubd (%rax), %mm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     phsubd (%rax), %mm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phsubd %xmm0, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phsubd (%rax), %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     phsubd (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phsubsw  %mm0, %mm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phsubsw  (%rax), %mm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     phsubsw  (%rax), %mm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phsubsw  %xmm0, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phsubsw  (%rax), %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     phsubsw  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phsubw %mm0, %mm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phsubw (%rax), %mm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     phsubw (%rax), %mm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phsubw %xmm0, %xmm2
-# CHECK-NEXT:  -      -      -      -      -      -      -      -      -      -      -      -      -     phsubw (%rax), %xmm2
+# CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -      -      -      -      -      -     phsubw (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     pmaddubsw  %mm0, %mm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     pmaddubsw  (%rax), %mm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     pmaddubsw  %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     pmaddubsw  (%rax), %xmm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     pmulhrsw %mm0, %mm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     pmulhrsw (%rax), %mm2
 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00    -      -      -      -     pmulhrsw %xmm0, %xmm2
 # CHECK-NEXT: 0.33   0.33   0.33    -      -      -      -      -     1.00    -      -      -      -     pmulhrsw (%rax), %xmm2

llvm/test/tools/llvm-mca/X86/Znver2/resources-x86_64.s

 # CHECK-NEXT:  1      1     0.25                        andq  $665536, %rdi
 # CHECK-NEXT:  2      5     0.33    *      *            andq  $665536, (%rax)
 # CHECK-NEXT:  1      1     0.25                        andq  $7, %rdi
 # CHECK-NEXT:  2      5     0.33    *      *            andq  $7, (%rax)
 # CHECK-NEXT:  1      1     0.25                        andq  %rsi, %rdi
 # CHECK-NEXT:  2      5     0.33    *      *            andq  %rsi, (%rax)
 # CHECK-NEXT:  2      5     0.33    *                   andq  (%rax), %rdi
 # CHECK-NEXT:  1      3     0.25                        bsfw  %si, %di
-# CHECK-NEXT:  1      3     0.25                        bsrw  %si, %di
+# CHECK-NEXT:  1      4     0.25                        bsrw  %si, %di
 # CHECK-NEXT:  2      7     0.33    *                   bsfw  (%rax), %di
-# CHECK-NEXT:  2      7     0.33    *                   bsrw  (%rax), %di
+# CHECK-NEXT:  2      8     0.33    *                   bsrw  (%rax), %di
 # CHECK-NEXT:  1      3     0.25                        bsfl  %esi, %edi
-# CHECK-NEXT:  1      3     0.25                        bsrl  %esi, %edi
+# CHECK-NEXT:  1      4     0.25                        bsrl  %esi, %edi
 # CHECK-NEXT:  2      7     0.33    *                   bsfl  (%rax), %edi
-# CHECK-NEXT:  2      7     0.33    *                   bsrl  (%rax), %edi
+# CHECK-NEXT:  2      8     0.33    *                   bsrl  (%rax), %edi
 # CHECK-NEXT:  1      3     0.25                        bsfq  %rsi, %rdi
-# CHECK-NEXT:  1      3     0.25                        bsrq  %rsi, %rdi
+# CHECK-NEXT:  1      4     0.25                        bsrq  %rsi, %rdi
 # CHECK-NEXT:  2      7     0.33    *                   bsfq  (%rax), %rdi
-# CHECK-NEXT:  2      7     0.33    *                   bsrq  (%rax), %rdi
+# CHECK-NEXT:  2      8     0.33    *                   bsrq  (%rax), %rdi
 # CHECK-NEXT:  1      1     1.00                        bswapl  %eax
 # CHECK-NEXT:  1      1     1.00                        bswapq  %rax
 # CHECK-NEXT:  1      1     0.25                        btw %si, %di
 # CHECK-NEXT:  2      2     0.25                        btcw  %si, %di
 # CHECK-NEXT:  2      2     0.25                        btrw  %si, %di
 # CHECK-NEXT:  2      2     0.25                        btsw  %si, %di
 # CHECK-NEXT:  2      5     0.33    *                   btw %si, (%rax)
 # CHECK-NEXT:  2      6     0.33    *      *            btcw  %si, (%rax)

 # CHECK-NEXT:  2      5     0.33    *                   cmpq  $7, (%rax)
 # CHECK-NEXT:  1      1     0.25                        cmpq  %rsi, %rdi
 # CHECK-NEXT:  2      5     0.33    *                   cmpq  %rsi, (%rax)
 # CHECK-NEXT:  2      5     0.33    *                   cmpq  (%rax), %rdi
 # CHECK-NEXT:  1      100   0.25                  U     cmpsb %es:(%rdi), (%rsi)
 # CHECK-NEXT:  1      100   0.25                  U     cmpsw %es:(%rdi), (%rsi)
 # CHECK-NEXT:  1      100   0.25                  U     cmpsl %es:(%rdi), (%rsi)
 # CHECK-NEXT:  1      100   0.25                  U     cmpsq %es:(%rdi), (%rsi)
-# CHECK-NEXT:  1      1     0.25                        cmpxchgb  %cl, %bl
+# CHECK-NEXT:  1      3     0.25                        cmpxchgb  %cl, %bl
 # CHECK-NEXT:  5      8     0.33    *      *            cmpxchgb  %cl, (%rbx)
-# CHECK-NEXT:  1      1     0.25                        cmpxchgw  %cx, %bx
+# CHECK-NEXT:  1      3     0.25                        cmpxchgw  %cx, %bx
 # CHECK-NEXT:  5      8     0.33    *      *            cmpxchgw  %cx, (%rbx)
-# CHECK-NEXT:  1      1     0.25                        cmpxchgl  %ecx, %ebx
+# CHECK-NEXT:  1      3     0.25                        cmpxchgl  %ecx, %ebx
 # CHECK-NEXT:  5      8     0.33    *      *            cmpxchgl  %ecx, (%rbx)
-# CHECK-NEXT:  1      1     0.25                        cmpxchgq  %rcx, %rbx
+# CHECK-NEXT:  1      3     0.25                        cmpxchgq  %rcx, %rbx
 # CHECK-NEXT:  5      8     0.33    *      *            cmpxchgq  %rcx, (%rbx)
 # CHECK-NEXT:  1      100   0.25                  U     cpuid
 # CHECK-NEXT:  1      1     0.25                        decb  %dil
 # CHECK-NEXT:  2      5     0.33    *      *            decb  (%rax)
 # CHECK-NEXT:  1      1     0.25                        decw  %di
 # CHECK-NEXT:  2      5     0.33    *      *            decw  (%rax)
 # CHECK-NEXT:  1      1     0.25                        decl  %edi
 # CHECK-NEXT:  2      5     0.33    *      *            decl  (%rax)

 # CHECK-NEXT:  2      41    41.00                 U     idivq %rcx
 # CHECK-NEXT:  3      45    41.00   *             U     idivq (%rax)
 # CHECK-NEXT:  1      4     1.00                        imulb %dil
 # CHECK-NEXT:  2      8     1.00    *                   imulb (%rax)
 # CHECK-NEXT:  1      3     1.00                        imulw %di
 # CHECK-NEXT:  1      7     1.00    *                   imulw (%rax)
 # CHECK-NEXT:  1      3     1.00                        imulw %si, %di
 # CHECK-NEXT:  1      7     1.00    *                   imulw (%rax), %di
-# CHECK-NEXT:  1      3     1.00                        imulw $511, %si, %di
+# CHECK-NEXT:  1      4     1.00                        imulw $511, %si, %di
 # CHECK-NEXT:  1      7     1.00    *                   imulw $511, (%rax), %di
-# CHECK-NEXT:  1      3     1.00                        imulw $7, %si, %di
+# CHECK-NEXT:  1      4     1.00                        imulw $7, %si, %di
 # CHECK-NEXT:  1      7     1.00    *                   imulw $7, (%rax), %di
 # CHECK-NEXT:  1      3     1.00                        imull %edi
 # CHECK-NEXT:  1      7     1.00    *                   imull (%rax)
 # CHECK-NEXT:  1      3     1.00                        imull %esi, %edi
 # CHECK-NEXT:  1      7     1.00    *                   imull (%rax), %edi
 # CHECK-NEXT:  1      3     1.00                        imull $665536, %esi, %edi
 # CHECK-NEXT:  1      7     1.00    *                   imull $665536, (%rax), %edi
 # CHECK-NEXT:  1      3     1.00                        imull $7, %esi, %edi