replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[X86] Fix SLM v2i64 ADD/Sub/CMPEQ instruction schedules 

Noticed while fixing the reduction costs for D59710 - the SLM model doesn't account for the poor throughput of v2i64 ops.

Numbers taken from Intel AOM (+ checked against Agner)
This commit is contained in:
Simon Pilgrim 2019-11-06 19:07:54 +00:00
parent ad70d5f39a
commit 1786047b91
3 changed files with 30 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
llvm/lib/Target/X86/X86ScheduleSLM.td
View File

@ -511,4 +511,20 @@ defm : X86WriteResUnsupported<WriteCvtPS2PHSt>;
defm : X86WriteResUnsupported<WriteCvtPS2PHYSt>;
defm : X86WriteResUnsupported<WriteCvtPS2PHZSt>;
// Remaining SLM instrs.
def SLMWriteResGroup1rr : SchedWriteRes<[SLM_FPC_RSV01]> {
let Latency = 4;
let NumMicroOps = 2;
let ResourceCycles = [4];
}
def: InstRW<[SLMWriteResGroup1rr], (instrs PADDQrr, PSUBQrr, PCMPEQQrr)>;
def SLMWriteResGroup1rm : SchedWriteRes<[SLM_MEC_RSV,SLM_FPC_RSV01]> {
let Latency = 7;
let NumMicroOps = 3;
let ResourceCycles = [1,4];
}
def: InstRW<[SLMWriteResGroup1rm], (instrs PADDQrm, PSUBQrm, PCMPEQQrm)>;
} // SchedModel

18
llvm/test/tools/llvm-mca/X86/SLM/resources-sse2.s
View File

@ -528,8 +528,8 @@ xorpd (%rax), %xmm2
# CHECK-NEXT: 1 4 1.00 * paddd (%rax), %xmm2
# CHECK-NEXT: 1 1 0.50 paddq %mm0, %mm2
# CHECK-NEXT: 1 4 1.00 * paddq (%rax), %mm2
# CHECK-NEXT: 1 1 0.50 paddq %xmm0, %xmm2
# CHECK-NEXT: 1 4 1.00 * paddq (%rax), %xmm2
# CHECK-NEXT: 2 4 2.00 paddq %xmm0, %xmm2
# CHECK-NEXT: 3 7 2.00 * paddq (%rax), %xmm2
# CHECK-NEXT: 1 1 0.50 paddsb %xmm0, %xmm2
# CHECK-NEXT: 1 4 1.00 * paddsb (%rax), %xmm2
# CHECK-NEXT: 1 1 0.50 paddsw %xmm0, %xmm2
@ -626,8 +626,8 @@ xorpd (%rax), %xmm2
# CHECK-NEXT: 1 4 1.00 * psubd (%rax), %xmm2
# CHECK-NEXT: 1 1 0.50 psubq %mm0, %mm2
# CHECK-NEXT: 1 4 1.00 * psubq (%rax), %mm2
# CHECK-NEXT: 1 1 0.50 psubq %xmm0, %xmm2
# CHECK-NEXT: 1 4 1.00 * psubq (%rax), %xmm2
# CHECK-NEXT: 2 4 2.00 psubq %xmm0, %xmm2
# CHECK-NEXT: 3 7 2.00 * psubq (%rax), %xmm2
# CHECK-NEXT: 1 1 0.50 psubsb %xmm0, %xmm2
# CHECK-NEXT: 1 4 1.00 * psubsb (%rax), %xmm2
# CHECK-NEXT: 1 1 0.50 psubsw %xmm0, %xmm2
@ -687,7 +687,7 @@ xorpd (%rax), %xmm2
# CHECK: Resource pressure per iteration:
# CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6] [7]
# CHECK-NEXT: - 412.00 12.00 152.50 90.50 3.00 3.00 134.00
# CHECK-NEXT: - 412.00 12.00 158.50 96.50 3.00 3.00 134.00
# CHECK: Resource pressure by instruction:
# CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6] [7] Instructions:
@ -812,8 +812,8 @@ xorpd (%rax), %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - 1.00 paddd (%rax), %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - - paddq %mm0, %mm2
# CHECK-NEXT: - - - 0.50 0.50 - - 1.00 paddq (%rax), %mm2
# CHECK-NEXT: - - - 0.50 0.50 - - - paddq %xmm0, %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - 1.00 paddq (%rax), %xmm2
# CHECK-NEXT: - - - 2.00 2.00 - - - paddq %xmm0, %xmm2
# CHECK-NEXT: - - - 2.00 2.00 - - 1.00 paddq (%rax), %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - - paddsb %xmm0, %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - 1.00 paddsb (%rax), %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - - paddsw %xmm0, %xmm2
@ -910,8 +910,8 @@ xorpd (%rax), %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - 1.00 psubd (%rax), %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - - psubq %mm0, %mm2
# CHECK-NEXT: - - - 0.50 0.50 - - 1.00 psubq (%rax), %mm2
# CHECK-NEXT: - - - 0.50 0.50 - - - psubq %xmm0, %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - 1.00 psubq (%rax), %xmm2
# CHECK-NEXT: - - - 2.00 2.00 - - - psubq %xmm0, %xmm2
# CHECK-NEXT: - - - 2.00 2.00 - - 1.00 psubq (%rax), %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - - psubsb %xmm0, %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - 1.00 psubsb (%rax), %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - - psubsw %xmm0, %xmm2

10
llvm/test/tools/llvm-mca/X86/SLM/resources-sse41.s
View File

@ -180,8 +180,8 @@ roundss $1, (%rax), %xmm2
# CHECK-NEXT: 1 4 1.00 * pblendvb %xmm0, (%rax), %xmm2
# CHECK-NEXT: 1 1 1.00 pblendw $11, %xmm0, %xmm2
# CHECK-NEXT: 1 4 1.00 * pblendw $11, (%rax), %xmm2
# CHECK-NEXT: 1 1 0.50 pcmpeqq %xmm0, %xmm2
# CHECK-NEXT: 1 4 1.00 * pcmpeqq (%rax), %xmm2
# CHECK-NEXT: 2 4 2.00 pcmpeqq %xmm0, %xmm2
# CHECK-NEXT: 3 7 2.00 * pcmpeqq (%rax), %xmm2
# CHECK-NEXT: 1 1 1.00 pextrb $1, %xmm0, %ecx
# CHECK-NEXT: 2 4 2.00 * pextrb $1, %xmm0, (%rax)
# CHECK-NEXT: 1 1 1.00 pextrd $1, %xmm0, %ecx
@ -264,7 +264,7 @@ roundss $1, (%rax), %xmm2
# CHECK: Resource pressure per iteration:
# CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6] [7]
# CHECK-NEXT: - - - 85.00 22.00 - - 54.00
# CHECK-NEXT: - - - 88.00 25.00 - - 54.00
# CHECK: Resource pressure by instruction:
# CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6] [7] Instructions:
@ -293,8 +293,8 @@ roundss $1, (%rax), %xmm2
# CHECK-NEXT: - - - 1.00 - - - 1.00 pblendvb %xmm0, (%rax), %xmm2
# CHECK-NEXT: - - - 1.00 - - - - pblendw $11, %xmm0, %xmm2
# CHECK-NEXT: - - - 1.00 - - - 1.00 pblendw $11, (%rax), %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - - pcmpeqq %xmm0, %xmm2
# CHECK-NEXT: - - - 0.50 0.50 - - 1.00 pcmpeqq (%rax), %xmm2
# CHECK-NEXT: - - - 2.00 2.00 - - - pcmpeqq %xmm0, %xmm2
# CHECK-NEXT: - - - 2.00 2.00 - - 1.00 pcmpeqq (%rax), %xmm2
# CHECK-NEXT: - - - 1.00 - - - - pextrb $1, %xmm0, %ecx
# CHECK-NEXT: - - - 1.00 - - - 2.00 pextrb $1, %xmm0, (%rax)
# CHECK-NEXT: - - - 1.00 - - - - pextrd $1, %xmm0, %ecx