[X86][AVX] Update VCVTx, VMOVNTPx and VROUNDYPx instructions scheduling on btver2

[avt77]

Summary

This patch updates/adds some schedule numbers to AVX instructions on btver2 CPU.

[avt77]

Postpone review of this patch until D39046 is committed: I should re-base it after that time.

[avt77]

I rebased the sources and made changes shorter: now we're dealing with 3 changed instrs only.

[RKSimon]

lib/Target/X86/X86ScheduleBtVer2.td

542

Split off MOVNT* - they are stores so require a JAGU stage

[avt77]

I fixed MOVNT* issue raised by Simon.

[RKSimon]

lib/Target/X86/X86ScheduleBtVer2.td

544

Please rename these WriteVCVTY / WriteVCVTYLd (or something similar). I think you can put the VCVTPS2DQ instructions in here as well?

555

Can you add VMOVNTDQYmr as well? Even though we don't test for it properly due to domain issues.....

576

This should be 8? Anyway, the VCVTPS2DQ can be merged with VCVTDQ2PD/VROUND above

[avt77]

All issues raised by Simon were fixed.

[RKSimon]

lib/Target/X86/X86ScheduleBtVer2.td

543

We need VCVTPS2DQYrr and VCVTTPS2DQYrr

551

We need VCVTPS2DQYrm and VCVTTPS2DQYrm

564

We need VCVTTPD2DQYrr as well

570

We need VCVTTPD2DQYrm as well

[avt77]

New required instructions were added.

[RKSimon]

Please can you rebase? I add the missing AVX cvtpd2dq/cvtps2dq tests at rL316926

[avt77]

I re-based the required test.

[RKSimon]

test/CodeGen/X86/avx-schedule.ll

1221–1222

Load latency looks wrong

1347

This should be [8:2.00]

[avt77]

I fixed the bug with vcvttp*.

[RKSimon]

LGTM - thanks

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 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 "VCVTTPD2DQYrm")>;
+
+def WriteVMONTPSt: SchedWriteRes<[JSTC, JLAGU]> {
+  let Latency = 3;
+  let ResourceCycles = [2,1];
+}
+def : InstRW<[WriteVMONTPSt], (instregex "VMOVNTP(S|D)Ymr")>;
+def : InstRW<[WriteVMONTPSt], (instregex "VMOVNTDQYmr")>;
+
+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")>;
+
+def WriteVCVTPDYLd: SchedWriteRes<[JLAGU, JSTC, JFPU01]> {
+  let Latency = 11;
+  let ResourceCycles = [1, 2, 4];
+}
+def : InstRW<[WriteVCVTPDYLd, ReadAfterLd], (instregex "VCVTPD2(DQ|PS)Yrm")>;
+def : InstRW<[WriteVCVTPDYLd, ReadAfterLd], (instregex "VCVTTPD2DQYrm")>;
+
 def WriteVSQRTYPD: SchedWriteRes<[JFPU1]> {
   let Latency = 54;
   let ResourceCycles = [54];
 }
 def : InstRW<[WriteVSQRTYPD], (instregex "VSQRTPDYr")>;
 
 def WriteVSQRTYPDLd: SchedWriteRes<[JLAGU, JFPU1]> {
   let Latency = 59;
   let ResourceCycles = [1, 54];

test/CodeGen/X86/avx-schedule.ll

 ; SKX:       # BB#0:
 ; SKX-NEXT:    vcvtdq2pd %xmm0, %ymm0 # sched: [7:1.00]
 ; SKX-NEXT:    vcvtdq2pd (%rdi), %ymm1 # sched: [13:1.00]
 ; SKX-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [4:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_cvtdq2pd:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vcvtdq2pd (%rdi), %ymm1 # sched: [8:1.00]
-; BTVER2-NEXT:    vcvtdq2pd %xmm0, %ymm0 # sched: [3:1.00]
+; BTVER2-NEXT:    vcvtdq2pd (%rdi), %ymm1 # sched: [8:2.00]
+; BTVER2-NEXT:    vcvtdq2pd %xmm0, %ymm0 # sched: [3:2.00]
 ; BTVER2-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_cvtdq2pd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vcvtdq2pd (%rdi), %ymm1 # sched: [12:1.00]
 ; ZNVER1-NEXT:    vcvtdq2pd %xmm0, %ymm0 # sched: [5:1.00]
 ; ZNVER1-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]

 ; SKX:       # BB#0:
 ; SKX-NEXT:    vcvtdq2ps %ymm0, %ymm0 # sched: [4:0.33]
 ; SKX-NEXT:    vcvtdq2ps (%rdi), %ymm1 # sched: [11:0.50]
 ; SKX-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [4:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_cvtdq2ps:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vcvtdq2ps (%rdi), %ymm1 # sched: [8:1.00]
-; BTVER2-NEXT:    vcvtdq2ps %ymm0, %ymm0 # sched: [3:1.00]
+; BTVER2-NEXT:    vcvtdq2ps (%rdi), %ymm1 # sched: [8:2.00]
+; BTVER2-NEXT:    vcvtdq2ps %ymm0, %ymm0 # sched: [3:2.00]
 ; BTVER2-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_cvtdq2ps:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vcvtdq2ps (%rdi), %ymm1 # sched: [12:1.00]
 ; ZNVER1-NEXT:    vcvtdq2ps %ymm0, %ymm0 # sched: [5:1.00]
 ; ZNVER1-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]

 ; SKX:       # BB#0:
 ; SKX-NEXT:    vcvttpd2dq %ymm0, %xmm0 # sched: [7:1.00]
 ; SKX-NEXT:    vcvttpd2dqy (%rdi), %xmm1 # sched: [8:1.00]
 ; SKX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_cvtpd2dq:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vcvttpd2dqy (%rdi), %xmm1 # sched: [8:1.00]
-; BTVER2-NEXT:    vcvttpd2dq %ymm0, %xmm0 # sched: [3:1.00]
+; BTVER2-NEXT:    vcvttpd2dqy (%rdi), %xmm1 # sched: [8:2.00]
+; BTVER2-NEXT:    vcvttpd2dq %ymm0, %xmm0 # sched: [6:2.00]
 ; BTVER2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_cvtpd2dq:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vcvttpd2dqy (%rdi), %xmm1 # sched: [12:1.00]
 ; ZNVER1-NEXT:    vcvttpd2dq %ymm0, %xmm0 # sched: [5:1.00]
 ; ZNVER1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 # sched: [2:0.67]

 ; SKX:       # BB#0:
 ; SKX-NEXT:    vcvtpd2ps %ymm0, %xmm0 # sched: [7:1.00]
 ; SKX-NEXT:    vcvtpd2psy (%rdi), %xmm1 # sched: [8:1.00]
 ; SKX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_cvtpd2ps:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vcvtpd2psy (%rdi), %xmm1 # sched: [8:1.00]
-; BTVER2-NEXT:    vcvtpd2ps %ymm0, %xmm0 # sched: [3:1.00]
+; BTVER2-NEXT:    vcvtpd2psy (%rdi), %xmm1 # sched: [11:2.00]
+; BTVER2-NEXT:    vcvtpd2ps %ymm0, %xmm0 # sched: [6:2.00]
 ; BTVER2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_cvtpd2ps:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vcvtpd2psy (%rdi), %xmm1 # sched: [11:1.00]
 ; ZNVER1-NEXT:    vcvtpd2ps %ymm0, %xmm0 # sched: [5:1.00]
 ; ZNVER1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 # sched: [2:0.67]

 ; SKX:       # BB#0:
 ; SKX-NEXT:    vcvttps2dq %ymm0, %ymm0 # sched: [4:0.33]
 ; SKX-NEXT:    vcvttps2dq (%rdi), %ymm1 # sched: [11:0.50]
 ; SKX-NEXT:    vorps %ymm1, %ymm0, %ymm0 # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_cvtps2dq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvttps2dq (%rdi), %ymm1 # sched: [8:1.00]
-; BTVER2-NEXT:    vcvttps2dq %ymm0, %ymm0 # sched: [3:1.00]
+; BTVER2-NEXT:    vcvttps2dq %ymm0, %ymm0 # sched: [3:2.00]
 ; BTVER2-NEXT:    vorps %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_cvtps2dq:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vcvttps2dq (%rdi), %ymm1 # sched: [12:1.00]
 ; ZNVER1-NEXT:    vcvttps2dq %ymm0, %ymm0 # sched: [5:1.00]
 ; ZNVER1-NEXT:    vorps %ymm1, %ymm0, %ymm0 # sched: [1:0.25]

 ; SKX:       # BB#0:
 ; SKX-NEXT:    vaddpd %ymm0, %ymm0, %ymm0 # sched: [4:0.33]
 ; SKX-NEXT:    vmovntpd %ymm0, (%rdi) # sched: [1:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_movntpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddpd %ymm0, %ymm0, %ymm0 # sched: [3:2.00]
-; BTVER2-NEXT:    vmovntpd %ymm0, (%rdi) # sched: [1:1.00]
+; BTVER2-NEXT:    vmovntpd %ymm0, (%rdi) # sched: [3:2.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_movntpd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vaddpd %ymm0, %ymm0, %ymm0 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vmovntpd %ymm0, (%rdi) # sched: [1:0.50]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = fadd <4 x double> %a0, %a0

 ; SKX:       # BB#0:
 ; SKX-NEXT:    vaddps %ymm0, %ymm0, %ymm0 # sched: [4:0.33]
 ; SKX-NEXT:    vmovntps %ymm0, (%rdi) # sched: [1:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_movntps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddps %ymm0, %ymm0, %ymm0 # sched: [3:2.00]
-; BTVER2-NEXT:    vmovntps %ymm0, (%rdi) # sched: [1:1.00]
+; BTVER2-NEXT:    vmovntps %ymm0, (%rdi) # sched: [3:2.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_movntps:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vaddps %ymm0, %ymm0, %ymm0 # sched: [3:1.00]
 ; ZNVER1-NEXT:    vmovntps %ymm0, (%rdi) # sched: [1:0.50]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = fadd <8 x float> %a0, %a0

 ; SKX:       # BB#0:
 ; SKX-NEXT:    vroundpd $7, %ymm0, %ymm0 # sched: [8:0.67]
 ; SKX-NEXT:    vroundpd $7, (%rdi), %ymm1 # sched: [15:0.67]
 ; SKX-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [4:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_roundpd:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vroundpd $7, (%rdi), %ymm1 # sched: [8:1.00]
-; BTVER2-NEXT:    vroundpd $7, %ymm0, %ymm0 # sched: [3:1.00]
+; BTVER2-NEXT:    vroundpd $7, (%rdi), %ymm1 # sched: [8:2.00]
+; BTVER2-NEXT:    vroundpd $7, %ymm0, %ymm0 # sched: [3:2.00]
 ; BTVER2-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_roundpd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vroundpd $7, (%rdi), %ymm1 # sched: [11:1.00]
 ; ZNVER1-NEXT:    vroundpd $7, %ymm0, %ymm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]

 ; SKX:       # BB#0:
 ; SKX-NEXT:    vroundps $7, %ymm0, %ymm0 # sched: [8:0.67]
 ; SKX-NEXT:    vroundps $7, (%rdi), %ymm1 # sched: [15:0.67]
 ; SKX-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [4:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_roundps:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vroundps $7, (%rdi), %ymm1 # sched: [8:1.00]
-; BTVER2-NEXT:    vroundps $7, %ymm0, %ymm0 # sched: [3:1.00]
+; BTVER2-NEXT:    vroundps $7, (%rdi), %ymm1 # sched: [8:2.00]
+; BTVER2-NEXT:    vroundps $7, %ymm0, %ymm0 # sched: [3:2.00]
 ; BTVER2-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_roundps:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vroundps $7, (%rdi), %ymm1 # sched: [11:1.00]
 ; ZNVER1-NEXT:    vroundps $7, %ymm0, %ymm0 # sched: [4:1.00]
 ; ZNVER1-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]