[X86] Update BTVER2 sched numbers for some AVX instructions (xmm version)

[avt77]

Summary

The current sched model does not have proper numbers for many128-bit instructions. This patch is the first one which tries to fix this issue.

[avt77]

Ping.

[RKSimon]

We're adding a lot of instregex exntries just to get around the poor quality of the sched classes in X86Schedule.td - better to fix them?

lib/Target/X86/X86ScheduleBtVer2.td

578 ↗

(On Diff #122990)

Shouldn't stores (mr) have a JSAGU dependency not a JLAGU

583 ↗

(On Diff #122990)

Shouldn't stores (mr) have a JSAGU dependency not a JLAGU

588 ↗

(On Diff #122990)

Shouldn't stores (mr) have a JSAGU dependency not a JLAGU?

596 ↗

(On Diff #122990)

Remove ResourceCyles? Looks like the default.

598 ↗

(On Diff #122990)

I wonder if we'd be better off introducing WriteVecLoad/WriteVecStore/WriteVecMove sched rw classes like the FIXME comment at line 140 suggests?

602 ↗

(On Diff #122990)

Shouldn't stores (mr) have a JSAGU dependency?

604 ↗

(On Diff #122990)

Should JLAGU be first for a load?

606 ↗

(On Diff #122990)

Remove ResourceCyles? Looks like the default.

624 ↗

(On Diff #122990)

Would we be better off introducing a WriteFCmp sched rw class to cover max/min and cmp?

674 ↗

(On Diff #122990)

Don't include commented out lines.

680 ↗

(On Diff #122990)

Don't include commented out lines.

[avt77]

I fixed all Simon's notes/comments/requirements.

[avt77]

In D40067#935009, @RKSimon wrote:

We're adding a lot of instregex exntries just to get around the poor quality of the sched classes in X86Schedule.td - better to fix them?

It seems I missed this comment: what do you mean here? I suppose it is what you wrote in PR17367. Am I right? If yes I could try to make the first steps in this direction but obviously the patch will increase in this case because:

• we're going to add the common classes;
• as result we'll be forced to fix several models simulteniously.

Is it OK?

[RKSimon]

lib/Target/X86/X86ScheduleBtVer2.td

582 ↗

(On Diff #124360)

The MULPS versions don't seem to be necessary - they match the WriteFMul defaults - its just the MULPD versions we need.

610 ↗

(On Diff #124360)

Please can you add the SSE4A MOVNTSDmr/MOVNTSSmr cases here? Note they don't have a V prefix.

619 ↗

(On Diff #124360)

Are the moves doing anything at all? I didn't see any changes to the schedule tests. If not just remove this

[avt77]

It seems all requirements from Simon were satisfied.

[avt77]

I removed all changes related to Atom.

[avt77]

I removed everything related to WriteVecMove and WriteVecStore.

[avt77]

I changed lattency for MOVSS/MOVSD and MOVAPS/MOVAPD and MOVDQA/MOVDQU.

[avt77]

RKSimon, do you like to see here anything more?

[avt77]

I removed WriteVecLoad as RKSimon required: it did not have any influence on cost numbers.

[RKSimon]

LGTM - thanks

llvm/trunk/lib/Target/X86/X86ScheduleBtVer2.td

 def : InstRW<[WriteVMULYPS], (instregex "VMULPSYrr", "VRCPPSYr", "VRSQRTPSYr")>;
 
 def WriteVMULYPSLd: SchedWriteRes<[JLAGU, JFPU1]> {
   let Latency = 7;
   let ResourceCycles = [1, 2];
 }
 def : InstRW<[WriteVMULYPSLd, ReadAfterLd], (instregex "VMULPSYrm", "VRCPPSYm", "VRSQRTPSYm")>;
 
+def WriteVMULPD: SchedWriteRes<[JFPU1]> {
+  let Latency = 4;
+  let ResourceCycles = [2];
+}
+def : InstRW<[WriteVMULPD], (instregex "VMULPDrr", "VMULSDrr")>;
+
+def WriteVMULPDLd: SchedWriteRes<[JLAGU, JFPU1]> {
+  let Latency = 9;
+  let ResourceCycles = [1, 2];
+}
+def : InstRW<[WriteVMULPDLd], (instregex "VMULPDrm", "VMULSDrm")>;
+
 def WriteVCVTY: SchedWriteRes<[JSTC]> {
   let Latency = 3;
   let ResourceCycles = [2];
 }
 def : InstRW<[WriteVCVTY], (instregex "VCVTDQ2P(S|D)Yrr")>;
 def : InstRW<[WriteVCVTY], (instregex "VROUNDYP(S|D)r")>;
 def : InstRW<[WriteVCVTY], (instregex "VCVTPS2DQYrr")>;
 def : InstRW<[WriteVCVTY], (instregex "VCVTTPS2DQYrr")>;
 
 def WriteVCVTYLd: SchedWriteRes<[JLAGU, JSTC]> {
   let Latency = 8;
   let ResourceCycles = [1, 2];
 }
 def : InstRW<[WriteVCVTYLd, ReadAfterLd], (instregex "VCVTDQ2P(S|D)Yrm")>;
 def : InstRW<[WriteVCVTYLd, ReadAfterLd], (instregex "VROUNDYP(S|D)m")>;
 def : InstRW<[WriteVCVTYLd, ReadAfterLd], (instregex "VCVTPS2DQYrm")>;
 def : InstRW<[WriteVCVTYLd, ReadAfterLd], (instregex "VCVTTPS2DQYrm")>;
 
-def WriteVMONTPSt: SchedWriteRes<[JSTC, JLAGU]> {
+def WriteVMOVTDQSt: SchedWriteRes<[JSTC, JSAGU]> {
+  let Latency = 2;
+}
+def : InstRW<[WriteVMOVTDQSt], (instregex "VMOVNTDQmr")>;
+
+def WriteMOVNTSt: SchedWriteRes<[JSTC, JSAGU]> {
+  let Latency = 3;
+}
+def : InstRW<[WriteMOVNTSt], (instregex "VMOVNTP(S|D)mr")>;
+def : InstRW<[WriteMOVNTSt], (instregex "MOVNTS(S|D)")>;
+
+def WriteVMONTPYSt: SchedWriteRes<[JSTC, JSAGU]> {
   let Latency = 3;
   let ResourceCycles = [2,1];
 }
-def : InstRW<[WriteVMONTPSt], (instregex "VMOVNTP(S|D)Ymr")>;
-def : InstRW<[WriteVMONTPSt], (instregex "VMOVNTDQYmr")>;
+def : InstRW<[WriteVMONTPYSt], (instregex "VMOVNTP(S|D)Ymr")>;
+def : InstRW<[WriteVMONTPYSt], (instregex "VMOVNTDQYmr")>;
+
+def WriteFCmp: SchedWriteRes<[JFPU0]> {
+  let Latency = 2;
+}
+
+def : InstRW<[WriteFCmp], (instregex "VMAXP(D|S)rr", "VMAXS(D|S)rr")>;
+def : InstRW<[WriteFCmp], (instregex "VMINP(D|S)rr", "VMINS(D|S)rr")>;
+def : InstRW<[WriteFCmp], (instregex "VCMPP(S|D)rri", "VCMPS(S|D)rri")>;
+
+def WriteFCmpLd: SchedWriteRes<[JLAGU, JFPU0]> {
+  let Latency = 7;
+}
+
+def : InstRW<[WriteFCmpLd], (instregex "VMAXP(D|S)rm", "VMAXS(D|S)rm")>;
+def : InstRW<[WriteFCmpLd], (instregex "VMINP(D|S)rm", "VMINS(D|S)rm")>;
+def : InstRW<[WriteFCmpLd], (instregex "VCMPP(S|D)rmi", "VCMPS(S|D)rmi")>;
 
 def WriteVCVTPDY: SchedWriteRes<[JSTC, JFPU01]> {
   let Latency = 6;
   let ResourceCycles = [2, 4];
 }
 def : InstRW<[WriteVCVTPDY], (instregex "VCVTPD2(DQ|PS)Yrr")>;
 def : InstRW<[WriteVCVTPDY], (instregex "VCVTTPD2DQYrr")>;
 

llvm/trunk/test/CodeGen/X86/sse-schedule.ll

 ; SKX-NEXT:    vcmpeqps %xmm1, %xmm0, %k0 # sched: [3:1.00]
 ; SKX-NEXT:    vcmpeqps (%rdi), %xmm0, %k1 # sched: [9:1.00]
 ; SKX-NEXT:    korw %k1, %k0, %k0 # sched: [1:1.00]
 ; SKX-NEXT:    vpmovm2d %k0, %xmm0 # sched: [1:0.25]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_cmpps:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [3:1.00]
-; BTVER2-NEXT:    vcmpeqps (%rdi), %xmm0, %xmm0 # sched: [8:1.00]
+; BTVER2-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [2:1.00]
+; BTVER2-NEXT:    vcmpeqps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; BTVER2-NEXT:    vorps %xmm0, %xmm1, %xmm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_cmpps:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vcmpeqps (%rdi), %xmm0, %xmm0 # sched: [10:1.00]
 ; ZNVER1-NEXT:    vorps %xmm0, %xmm1, %xmm0 # sched: [1:0.25]

 ; SKX-LABEL: test_maxps:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]
 ; SKX-NEXT:    vmaxps (%rdi), %xmm0, %xmm0 # sched: [10:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_maxps:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; BTVER2-NEXT:    vmaxps (%rdi), %xmm0, %xmm0 # sched: [8:1.00]
+; BTVER2-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; BTVER2-NEXT:    vmaxps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_maxps:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vmaxps (%rdi), %xmm0, %xmm0 # sched: [10:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <4 x float> @llvm.x86.sse.max.ps(<4 x float> %a0, <4 x float> %a1)

 ; SKX-LABEL: test_maxss:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vmaxss %xmm1, %xmm0, %xmm0 # sched: [4:0.33]
 ; SKX-NEXT:    vmaxss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_maxss:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    vmaxss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; BTVER2-NEXT:    vmaxss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]
+; BTVER2-NEXT:    vmaxss %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; BTVER2-NEXT:    vmaxss (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_maxss:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vmaxss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vmaxss (%rdi), %xmm0, %xmm0 # sched: [10:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <4 x float> @llvm.x86.sse.max.ss(<4 x float> %a0, <4 x float> %a1)

 ; SKX-LABEL: test_minps:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vminps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]
 ; SKX-NEXT:    vminps (%rdi), %xmm0, %xmm0 # sched: [10:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_minps:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    vminps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; BTVER2-NEXT:    vminps (%rdi), %xmm0, %xmm0 # sched: [8:1.00]
+; BTVER2-NEXT:    vminps %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; BTVER2-NEXT:    vminps (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_minps:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vminps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vminps (%rdi), %xmm0, %xmm0 # sched: [10:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <4 x float> @llvm.x86.sse.min.ps(<4 x float> %a0, <4 x float> %a1)

 ; SKX-LABEL: test_minss:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vminss %xmm1, %xmm0, %xmm0 # sched: [4:0.33]
 ; SKX-NEXT:    vminss (%rdi), %xmm0, %xmm0 # sched: [9:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_minss:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    vminss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; BTVER2-NEXT:    vminss (%rdi), %xmm0, %xmm0 # sched: [8:1.00]
+; BTVER2-NEXT:    vminss %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; BTVER2-NEXT:    vminss (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_minss:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vminss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vminss (%rdi), %xmm0, %xmm0 # sched: [10:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <4 x float> @llvm.x86.sse.min.ss(<4 x float> %a0, <4 x float> %a1)

 ;
 ; SKX-LABEL: test_movntps:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vmovntps %xmm0, (%rdi) # sched: [1:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_movntps:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    vmovntps %xmm0, (%rdi) # sched: [1:1.00]
+; BTVER2-NEXT:    vmovntps %xmm0, (%rdi) # sched: [3:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_movntps:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vmovntps %xmm0, (%rdi) # sched: [1:0.50]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   store <4 x float> %a0, <4 x float> *%a1, align 16, !nontemporal !0
   ret void

llvm/trunk/test/CodeGen/X86/sse2-schedule.ll

 ; SKX-NEXT:    vcmpeqpd %xmm1, %xmm0, %k0 # sched: [3:1.00]
 ; SKX-NEXT:    vcmpeqpd (%rdi), %xmm0, %k1 # sched: [9:1.00]
 ; SKX-NEXT:    korw %k1, %k0, %k0 # sched: [1:1.00]
 ; SKX-NEXT:    vpmovm2q %k0, %xmm0 # sched: [1:0.25]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_cmppd:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm1 # sched: [3:1.00]
-; BTVER2-NEXT:    vcmpeqpd (%rdi), %xmm0, %xmm0 # sched: [8:1.00]
+; BTVER2-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm1 # sched: [2:1.00]
+; BTVER2-NEXT:    vcmpeqpd (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; BTVER2-NEXT:    vorpd %xmm0, %xmm1, %xmm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_cmppd:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm1 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vcmpeqpd (%rdi), %xmm0, %xmm0 # sched: [10:1.00]
 ; ZNVER1-NEXT:    vorpd %xmm0, %xmm1, %xmm0 # sched: [1:0.25]

 ; SKX-LABEL: test_maxpd:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vmaxpd %xmm1, %xmm0, %xmm0 # sched: [4:0.33]
 ; SKX-NEXT:    vmaxpd (%rdi), %xmm0, %xmm0 # sched: [10:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_maxpd:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    vmaxpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; BTVER2-NEXT:    vmaxpd (%rdi), %xmm0, %xmm0 # sched: [8:1.00]
+; BTVER2-NEXT:    vmaxpd %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; BTVER2-NEXT:    vmaxpd (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_maxpd:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vmaxpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vmaxpd (%rdi), %xmm0, %xmm0 # sched: [10:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <2 x double> @llvm.x86.sse2.max.pd(<2 x double> %a0, <2 x double> %a1)

 ; SKX-LABEL: test_maxsd:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vmaxsd %xmm1, %xmm0, %xmm0 # sched: [4:0.33]
 ; SKX-NEXT:    vmaxsd (%rdi), %xmm0, %xmm0 # sched: [9:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_maxsd:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    vmaxsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; BTVER2-NEXT:    vmaxsd (%rdi), %xmm0, %xmm0 # sched: [8:1.00]
+; BTVER2-NEXT:    vmaxsd %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; BTVER2-NEXT:    vmaxsd (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_maxsd:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vmaxsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vmaxsd (%rdi), %xmm0, %xmm0 # sched: [10:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <2 x double> @llvm.x86.sse2.max.sd(<2 x double> %a0, <2 x double> %a1)

 ; SKX-LABEL: test_minpd:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vminpd %xmm1, %xmm0, %xmm0 # sched: [4:0.33]
 ; SKX-NEXT:    vminpd (%rdi), %xmm0, %xmm0 # sched: [10:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_minpd:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    vminpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; BTVER2-NEXT:    vminpd (%rdi), %xmm0, %xmm0 # sched: [8:1.00]
+; BTVER2-NEXT:    vminpd %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; BTVER2-NEXT:    vminpd (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_minpd:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vminpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vminpd (%rdi), %xmm0, %xmm0 # sched: [10:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <2 x double> @llvm.x86.sse2.min.pd(<2 x double> %a0, <2 x double> %a1)

 ; SKX-LABEL: test_minsd:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vminsd %xmm1, %xmm0, %xmm0 # sched: [4:0.33]
 ; SKX-NEXT:    vminsd (%rdi), %xmm0, %xmm0 # sched: [9:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_minsd:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    vminsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; BTVER2-NEXT:    vminsd (%rdi), %xmm0, %xmm0 # sched: [8:1.00]
+; BTVER2-NEXT:    vminsd %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; BTVER2-NEXT:    vminsd (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_minsd:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vminsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vminsd (%rdi), %xmm0, %xmm0 # sched: [10:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <2 x double> @llvm.x86.sse2.min.sd(<2 x double> %a0, <2 x double> %a1)

 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vpaddq %xmm0, %xmm0, %xmm0 # sched: [1:0.33]
 ; SKX-NEXT:    vmovntdq %xmm0, (%rdi) # sched: [1:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_movntdqa:
 ; BTVER2:       # %bb.0:
 ; BTVER2-NEXT:    vpaddq %xmm0, %xmm0, %xmm0 # sched: [1:0.50]
-; BTVER2-NEXT:    vmovntdq %xmm0, (%rdi) # sched: [1:1.00]
+; BTVER2-NEXT:    vmovntdq %xmm0, (%rdi) # sched: [2:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_movntdqa:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vpaddq %xmm0, %xmm0, %xmm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    vmovntdq %xmm0, (%rdi) # sched: [1:0.50]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = add <2 x i64> %a0, %a0

 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vaddpd %xmm0, %xmm0, %xmm0 # sched: [4:0.33]
 ; SKX-NEXT:    vmovntpd %xmm0, (%rdi) # sched: [1:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_movntpd:
 ; BTVER2:       # %bb.0:
 ; BTVER2-NEXT:    vaddpd %xmm0, %xmm0, %xmm0 # sched: [3:1.00]
-; BTVER2-NEXT:    vmovntpd %xmm0, (%rdi) # sched: [1:1.00]
+; BTVER2-NEXT:    vmovntpd %xmm0, (%rdi) # sched: [3:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_movntpd:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vaddpd %xmm0, %xmm0, %xmm0 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vmovntpd %xmm0, (%rdi) # sched: [1:0.50]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = fadd <2 x double> %a0, %a0

 ; SKX-LABEL: test_mulpd:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vmulpd %xmm1, %xmm0, %xmm0 # sched: [4:0.33]
 ; SKX-NEXT:    vmulpd (%rdi), %xmm0, %xmm0 # sched: [10:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_mulpd:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    vmulpd %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
-; BTVER2-NEXT:    vmulpd (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
+; BTVER2-NEXT:    vmulpd %xmm1, %xmm0, %xmm0 # sched: [4:2.00]
+; BTVER2-NEXT:    vmulpd (%rdi), %xmm0, %xmm0 # sched: [9:2.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_mulpd:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vmulpd %xmm1, %xmm0, %xmm0 # sched: [3:0.50]
 ; ZNVER1-NEXT:    vmulpd (%rdi), %xmm0, %xmm0 # sched: [10:0.50]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = fmul <2 x double> %a0, %a1

 ; SKX-LABEL: test_mulsd:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vmulsd %xmm1, %xmm0, %xmm0 # sched: [4:0.33]
 ; SKX-NEXT:    vmulsd (%rdi), %xmm0, %xmm0 # sched: [9:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_mulsd:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    vmulsd %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
-; BTVER2-NEXT:    vmulsd (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
+; BTVER2-NEXT:    vmulsd %xmm1, %xmm0, %xmm0 # sched: [4:2.00]
+; BTVER2-NEXT:    vmulsd (%rdi), %xmm0, %xmm0 # sched: [9:2.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_mulsd:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vmulsd %xmm1, %xmm0, %xmm0 # sched: [3:0.50]
 ; ZNVER1-NEXT:    vmulsd (%rdi), %xmm0, %xmm0 # sched: [10:0.50]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = fmul double %a0, %a1

llvm/trunk/test/CodeGen/X86/sse4a-schedule.ll

 define void @test_movntsd(i8* %p, <2 x double> %a) {
 ; GENERIC-LABEL: test_movntsd:
 ; GENERIC:       # %bb.0:
 ; GENERIC-NEXT:    movntsd %xmm0, (%rdi) # sched: [1:1.00]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; BTVER2-LABEL: test_movntsd:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    movntsd %xmm0, (%rdi) # sched: [1:1.00]
+; BTVER2-NEXT:    movntsd %xmm0, (%rdi) # sched: [3:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_movntsd:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    movntsd %xmm0, (%rdi) # sched: [8:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   tail call void @llvm.x86.sse4a.movnt.sd(i8* %p, <2 x double> %a)
   ret void
 }
 declare void @llvm.x86.sse4a.movnt.sd(i8*, <2 x double>)
 
 define void @test_movntss(i8* %p, <4 x float> %a) {
 ; GENERIC-LABEL: test_movntss:
 ; GENERIC:       # %bb.0:
 ; GENERIC-NEXT:    movntss %xmm0, (%rdi) # sched: [1:1.00]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; BTVER2-LABEL: test_movntss:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    movntss %xmm0, (%rdi) # sched: [1:1.00]
+; BTVER2-NEXT:    movntss %xmm0, (%rdi) # sched: [3:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_movntss:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    movntss %xmm0, (%rdi) # sched: [8:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   tail call void @llvm.x86.sse4a.movnt.ss(i8* %p, <4 x float> %a)
   ret void
 }
 declare void @llvm.x86.sse4a.movnt.ss(i8*, <4 x float>)