Sched model improving on btver2: JFPU01 resource, vtestp* for xmm.

[avt77]

Summary

I fixed some issues covered in the title.

[andreadb]

lib/Target/X86/X86ScheduleBtVer2.td

202 ↗

(On Diff #122093)

My understanding is that ResourceCycles has to be 4 because you want to be able to compute a reciprocal throughput of 2. According to the amd documents, the pipes used by a float variable blend are "FPA|FPM".

I wonder whether we could have [JFPU0, JFPU1] instead of [JFPU01], and then change ResourceCycles to [2, 1].

A variable blend is 3 uOps, and internally, it is likely to be implemented as the sequence {VAND,VANDN,VOR}, where VAND and VANDN can execute in parallel.

It may be worthy to run some experiments to see which approach is better. That being said, your approach is not wrong, and I don't have a strong opinion on this.

202–206 ↗

(On Diff #122093)

let NumMicroOps = 3;

218–225 ↗

(On Diff #122093)

Variable blend instructions are 3 macro ops.
You should add let NumMicroOps = 3;.

[avt77]

lib/Target/X86/X86ScheduleBtVer2.td

202 ↗

(On Diff #122093)

I have AMD laptop to make experiments but I don't have any perf test using a variable blend. Could anyone to help me with such a test?

[avt77]

I made updates required by andreadb except the question about JFPU01: if we need perf experiments please help me with perf test(s). If we agree to use the current implementation then I could commit the patch if you get me LGTM.

[andreadb]

In D39802#920299, @avt77 wrote:

I made updates required by andreadb except the question about JFPU01: if we need perf experiments please help me with perf test(s). If we agree to use the current implementation then I could commit the patch if you get me LGTM.

I don't think it is worthy to run experiments just for the blendv case.
I just wanted to point out that there may be an alternative solution for that particular case. Having a resource cycle of 4 for the blendv is a bit odd (at least to me)..
But, as I wrote said, I really don't have a strong opinion on what numbers should go there (neither solution is ideal).
If you fix the remaining NumMicroOps issue I pointed out, then the patch looks good to me.

lib/Target/X86/X86ScheduleBtVer2.td

656–659 ↗

(On Diff #122218)

let NumMicroOps = 3;

663–666 ↗

(On Diff #122218)

Same.

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

 
 // FIXME: integer pipes
 defm : JWriteResFpuPair<WriteCvtF2I,    JFPU1,  3>; // Float -> Integer.
 defm : JWriteResFpuPair<WriteCvtI2F,    JFPU1,  3>; // Integer -> Float.
 defm : JWriteResFpuPair<WriteCvtF2F,    JFPU1,  3>; // Float -> Float size conversion.
 
 def : WriteRes<WriteFVarBlend, [JFPU01]> {
   let Latency = 2;
-  let ResourceCycles = [2];
+  let ResourceCycles = [4];
+  let NumMicroOps = 3;
 }
 def : WriteRes<WriteFVarBlendLd, [JLAGU, JFPU01]> {
   let Latency = 7;
-  let ResourceCycles = [1, 2];
+  let ResourceCycles = [1, 4];
+  let NumMicroOps = 3;
 }
 
 // Vector integer operations.
 defm : JWriteResFpuPair<WriteVecALU,   JFPU01,  1>;
 defm : JWriteResFpuPair<WriteVecShift, JFPU01,  1>;
 defm : JWriteResFpuPair<WriteVecIMul,  JFPU0,   2>;
 defm : JWriteResFpuPair<WriteShuffle,  JFPU01,  1>;
 defm : JWriteResFpuPair<WriteBlend,    JFPU01,  1>;
 defm : JWriteResFpuPair<WriteVecLogic, JFPU01,  1>;
 defm : JWriteResFpuPair<WriteShuffle256, JFPU01, 1>;
 
 def : WriteRes<WriteVarBlend, [JFPU01]> {
   let Latency = 2;
-  let ResourceCycles = [2];
+  let ResourceCycles = [4];
+  let NumMicroOps = 3;
 }
 def : WriteRes<WriteVarBlendLd, [JLAGU, JFPU01]> {
   let Latency = 7;
-  let ResourceCycles = [1, 2];
+  let ResourceCycles = [1, 4];
+  let NumMicroOps = 3;
 }
 
 // FIXME: why do we need to define AVX2 resource on CPU that doesn't have AVX2?
-def : WriteRes<WriteVarVecShift, [JFPU01]> {
-  let Latency = 1;
-  let ResourceCycles = [1];
-}
+def : WriteRes<WriteVarVecShift, [JFPU01]> {}
 def : WriteRes<WriteVarVecShiftLd, [JLAGU, JFPU01]> {
   let Latency = 6;
-  let ResourceCycles = [1, 1];
+  let ResourceCycles = [1, 2];
 }
 
 def : WriteRes<WriteMPSAD, [JFPU0]> {
   let Latency = 3;
   let ResourceCycles = [2];
 }
 def : WriteRes<WriteMPSADLd, [JLAGU, JFPU0]> {
   let Latency = 8;

 def : InstRW<[WriteVMOVMSK], (instregex "VMOVMSKP(D|S)(Y)?rr")>;
 
 // TODO: In fact we have latency '3+i'. The +i represents an additional 1 cycle transfer
 // operation which moves the floating point result to the integer unit. During this
 // additional cycle the floating point unit execution resources are not occupied
 // and ALU0 in the integer unit is occupied instead.
 def WriteVTESTY: SchedWriteRes<[JFPU01, JFPU0]> {
   let Latency = 4;
-  let ResourceCycles = [4, 2];
+  let ResourceCycles = [2, 2];
+  let NumMicroOps = 3;
 }
 def : InstRW<[WriteVTESTY], (instregex "VTESTP(S|D)Yrr")>;
 def : InstRW<[WriteVTESTY], (instregex "VPTESTYrr")>;
 
 def WriteVTESTYLd: SchedWriteRes<[JLAGU, JFPU01, JFPU0]> {
   let Latency = 9;
-  let ResourceCycles = [1, 4, 2];
+  let ResourceCycles = [1, 2, 2];
+  let NumMicroOps = 3;
 }
 def : InstRW<[WriteVTESTYLd], (instregex "VTESTP(S|D)Yrm")>;
 def : InstRW<[WriteVTESTYLd], (instregex "VPTESTYrm")>;
 
+def WriteVTEST: SchedWriteRes<[JFPU0]> {
+  let Latency = 3;
+}
+def : InstRW<[WriteVTEST], (instregex "VTESTP(S|D)rr")>;
+def : InstRW<[WriteVTEST], (instregex "VPTESTrr")>;
+
+def WriteVTESTLd: SchedWriteRes<[JLAGU, JFPU0]> {
+  let Latency = 8;
+}
+def : InstRW<[WriteVTESTLd], (instregex "VTESTP(S|D)rm")>;
+def : InstRW<[WriteVTESTLd], (instregex "VPTESTrm")>;
+
 def WriteVSQRTYPD: SchedWriteRes<[JFPU1]> {
   let Latency = 54;
   let ResourceCycles = [54];
 }
 def : InstRW<[WriteVSQRTYPD], (instregex "VSQRTPDYr")>;
 
 def WriteVSQRTYPDLd: SchedWriteRes<[JLAGU, JFPU1]> {
   let Latency = 59;

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

 ; SKX-NEXT:    setb %al # sched: [1:0.50]
 ; SKX-NEXT:    vtestpd (%rdi), %xmm0 # sched: [8:1.00]
 ; SKX-NEXT:    adcl $0, %eax # sched: [1:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_testpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    xorl %eax, %eax # sched: [1:0.50]
-; BTVER2-NEXT:    vtestpd %xmm1, %xmm0 # sched: [1:0.50]
+; BTVER2-NEXT:    vtestpd %xmm1, %xmm0 # sched: [3:1.00]
 ; BTVER2-NEXT:    setb %al # sched: [1:0.50]
-; BTVER2-NEXT:    vtestpd (%rdi), %xmm0 # sched: [6:1.00]
+; BTVER2-NEXT:    vtestpd (%rdi), %xmm0 # sched: [8:1.00]
 ; BTVER2-NEXT:    adcl $0, %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_testpd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    xorl %eax, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    vtestpd %xmm1, %xmm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    setb %al # sched: [1:0.25]

 ; SKX-NEXT:    vtestpd (%rdi), %ymm0 # sched: [9:1.00]
 ; SKX-NEXT:    adcl $0, %eax # sched: [1:0.50]
 ; SKX-NEXT:    vzeroupper # sched: [4:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_testpd_ymm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    xorl %eax, %eax # sched: [1:0.50]
-; BTVER2-NEXT:    vtestpd %ymm1, %ymm0 # sched: [4:3.00]
+; BTVER2-NEXT:    vtestpd %ymm1, %ymm0 # sched: [4:2.00]
 ; BTVER2-NEXT:    setb %al # sched: [1:0.50]
-; BTVER2-NEXT:    vtestpd (%rdi), %ymm0 # sched: [9:3.00]
+; BTVER2-NEXT:    vtestpd (%rdi), %ymm0 # sched: [9:2.00]
 ; BTVER2-NEXT:    adcl $0, %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_testpd_ymm:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    xorl %eax, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    vtestpd %ymm1, %ymm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    setb %al # sched: [1:0.25]

 ; SKX-NEXT:    setb %al # sched: [1:0.50]
 ; SKX-NEXT:    vtestps (%rdi), %xmm0 # sched: [8:1.00]
 ; SKX-NEXT:    adcl $0, %eax # sched: [1:0.50]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_testps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    xorl %eax, %eax # sched: [1:0.50]
-; BTVER2-NEXT:    vtestps %xmm1, %xmm0 # sched: [1:0.50]
+; BTVER2-NEXT:    vtestps %xmm1, %xmm0 # sched: [3:1.00]
 ; BTVER2-NEXT:    setb %al # sched: [1:0.50]
-; BTVER2-NEXT:    vtestps (%rdi), %xmm0 # sched: [6:1.00]
+; BTVER2-NEXT:    vtestps (%rdi), %xmm0 # sched: [8:1.00]
 ; BTVER2-NEXT:    adcl $0, %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_testps:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    xorl %eax, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    vtestps %xmm1, %xmm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    setb %al # sched: [1:0.25]

 ; SKX-NEXT:    vtestps (%rdi), %ymm0 # sched: [9:1.00]
 ; SKX-NEXT:    adcl $0, %eax # sched: [1:0.50]
 ; SKX-NEXT:    vzeroupper # sched: [4:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_testps_ymm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    xorl %eax, %eax # sched: [1:0.50]
-; BTVER2-NEXT:    vtestps %ymm1, %ymm0 # sched: [4:3.00]
+; BTVER2-NEXT:    vtestps %ymm1, %ymm0 # sched: [4:2.00]
 ; BTVER2-NEXT:    setb %al # sched: [1:0.50]
-; BTVER2-NEXT:    vtestps (%rdi), %ymm0 # sched: [9:3.00]
+; BTVER2-NEXT:    vtestps (%rdi), %ymm0 # sched: [9:2.00]
 ; BTVER2-NEXT:    adcl $0, %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_testps_ymm:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    xorl %eax, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    vtestps %ymm1, %ymm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    setb %al # sched: [1:0.25]

llvm/trunk/test/CodeGen/X86/sse41-schedule.ll

 ; SKX-LABEL: test_blendvpd:
 ; SKX:       # BB#0:
 ; SKX-NEXT:    vblendvpd %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:0.67]
 ; SKX-NEXT:    vblendvpd %xmm2, (%rdi), %xmm0, %xmm0 # sched: [8:0.67]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_blendvpd:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vblendvpd %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
-; BTVER2-NEXT:    vblendvpd %xmm2, (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
+; BTVER2-NEXT:    vblendvpd %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:2.00]
+; BTVER2-NEXT:    vblendvpd %xmm2, (%rdi), %xmm0, %xmm0 # sched: [7:2.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_blendvpd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vblendvpd %xmm2, %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; ZNVER1-NEXT:    vblendvpd %xmm2, (%rdi), %xmm0, %xmm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <2 x double> @llvm.x86.sse41.blendvpd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2)

 ; SKX-LABEL: test_blendvps:
 ; SKX:       # BB#0:
 ; SKX-NEXT:    vblendvps %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:0.67]
 ; SKX-NEXT:    vblendvps %xmm2, (%rdi), %xmm0, %xmm0 # sched: [8:0.67]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_blendvps:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vblendvps %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
-; BTVER2-NEXT:    vblendvps %xmm2, (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
+; BTVER2-NEXT:    vblendvps %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:2.00]
+; BTVER2-NEXT:    vblendvps %xmm2, (%rdi), %xmm0, %xmm0 # sched: [7:2.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_blendvps:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vblendvps %xmm2, %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; ZNVER1-NEXT:    vblendvps %xmm2, (%rdi), %xmm0, %xmm0 # sched: [8:0.50]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <4 x float> @llvm.x86.sse41.blendvps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2)

 ; SKX-LABEL: test_pblendvb:
 ; SKX:       # BB#0:
 ; SKX-NEXT:    vpblendvb %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:0.67]
 ; SKX-NEXT:    vpblendvb %xmm2, (%rdi), %xmm0, %xmm0 # sched: [8:0.67]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pblendvb:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vpblendvb %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
-; BTVER2-NEXT:    vpblendvb %xmm2, (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
+; BTVER2-NEXT:    vpblendvb %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:2.00]
+; BTVER2-NEXT:    vpblendvb %xmm2, (%rdi), %xmm0, %xmm0 # sched: [7:2.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pblendvb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpblendvb %xmm2, %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
 ; ZNVER1-NEXT:    vpblendvb %xmm2, (%rdi), %xmm0, %xmm0 # sched: [8:1.00]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <16 x i8> @llvm.x86.sse41.pblendvb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> %a2)

 ; SKX-NEXT:    vptest (%rdi), %xmm0 # sched: [9:1.00]
 ; SKX-NEXT:    setb %cl # sched: [1:0.50]
 ; SKX-NEXT:    andb %al, %cl # sched: [1:0.25]
 ; SKX-NEXT:    movzbl %cl, %eax # sched: [1:0.25]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_ptest:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vptest %xmm1, %xmm0 # sched: [1:0.50]
+; BTVER2-NEXT:    vptest %xmm1, %xmm0 # sched: [3:1.00]
 ; BTVER2-NEXT:    setb %al # sched: [1:0.50]
-; BTVER2-NEXT:    vptest (%rdi), %xmm0 # sched: [6:1.00]
+; BTVER2-NEXT:    vptest (%rdi), %xmm0 # sched: [8:1.00]
 ; BTVER2-NEXT:    setb %cl # sched: [1:0.50]
 ; BTVER2-NEXT:    andb %al, %cl # sched: [1:0.50]
 ; BTVER2-NEXT:    movzbl %cl, %eax # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_ptest:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vptest %xmm1, %xmm0 # sched: [1:1.00]