Skip to content

Commit 02f03a6

Browse files
committed
[X86] Match vpmullq latency to uops.info. Correct port usage for 512-bit memory form
uops.info says these should be 15 cycle instructions. Uops.info also shows the 512-bit form uses port 0 and 5 for both register and memory. We had memory using 0 and 1. Differential Revision: https://reviews.llvm.org/D75549
1 parent 1bedb23 commit 02f03a6

File tree

3 files changed

+43
-37
lines changed

3 files changed

+43
-37
lines changed

llvm/lib/Target/X86/X86SchedSkylakeServer.td

Lines changed: 14 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -1963,14 +1963,14 @@ def SKXWriteResGroup171 : SchedWriteRes<[SKXPort06,SKXPort0156]> {
19631963
def: InstRW<[SKXWriteResGroup171], (instrs LOOPE, LOOPNE)>;
19641964

19651965
def SKXWriteResGroup174 : SchedWriteRes<[SKXPort01]> {
1966-
let Latency = 12;
1966+
let Latency = 15;
19671967
let NumMicroOps = 3;
19681968
let ResourceCycles = [3];
19691969
}
19701970
def: InstRW<[SKXWriteResGroup174], (instregex "VPMULLQ(Z128|Z256)rr")>;
19711971

19721972
def SKXWriteResGroup174z : SchedWriteRes<[SKXPort05]> {
1973-
let Latency = 12;
1973+
let Latency = 15;
19741974
let NumMicroOps = 3;
19751975
let ResourceCycles = [3];
19761976
}
@@ -2131,8 +2131,8 @@ def SKXWriteResGroup202 : SchedWriteRes<[SKXPort0,SKXPort1,SKXPort5,SKXPort6,SKX
21312131
}
21322132
def: InstRW<[SKXWriteResGroup202], (instrs XCH_F)>;
21332133

2134-
def SKXWriteResGroup205 : SchedWriteRes<[SKXPort23,SKXPort015]> {
2135-
let Latency = 18;
2134+
def SKXWriteResGroup205 : SchedWriteRes<[SKXPort23,SKXPort01]> {
2135+
let Latency = 21;
21362136
let NumMicroOps = 4;
21372137
let ResourceCycles = [1,3];
21382138
}
@@ -2159,13 +2159,19 @@ def SKXWriteResGroup209 : SchedWriteRes<[SKXPort0,SKXPort23,SKXFPDivider]> {
21592159
}
21602160
def : SchedAlias<WriteFDiv64Ld, SKXWriteResGroup209>; // TODO - convert to ZnWriteResFpuPair
21612161

2162-
def SKXWriteResGroup211 : SchedWriteRes<[SKXPort23,SKXPort015]> {
2163-
let Latency = 19;
2162+
def SKXWriteResGroup211 : SchedWriteRes<[SKXPort23,SKXPort01]> {
2163+
let Latency = 22;
2164+
let NumMicroOps = 4;
2165+
let ResourceCycles = [1,3];
2166+
}
2167+
def: InstRW<[SKXWriteResGroup211], (instregex "VPMULLQZ256rm(b?)")>;
2168+
2169+
def SKXWriteResGroup211_1 : SchedWriteRes<[SKXPort23,SKXPort05]> {
2170+
let Latency = 22;
21642171
let NumMicroOps = 4;
21652172
let ResourceCycles = [1,3];
21662173
}
2167-
def: InstRW<[SKXWriteResGroup211], (instregex "VPMULLQZ256rm(b?)",
2168-
"VPMULLQZrm(b?)")>;
2174+
def: InstRW<[SKXWriteResGroup211_1], (instregex "VPMULLQZrm(b?)")>;
21692175

21702176
def SKXWriteResGroup215 : SchedWriteRes<[SKXPort0]> {
21712177
let Latency = 20;

llvm/test/tools/llvm-mca/X86/SkylakeServer/resources-avx512dq.s

Lines changed: 10 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -252,12 +252,12 @@ vxorps (%rax){1to16}, %zmm17, %zmm19 {z}{k1}
252252
# CHECK-NEXT: 2 9 1.00 * vfpclassss $171, (%rax), %k1
253253
# CHECK-NEXT: 1 4 1.00 vfpclassss $171, %xmm16, %k1 {%k2}
254254
# CHECK-NEXT: 2 9 1.00 * vfpclassss $171, (%rax), %k1 {%k2}
255-
# CHECK-NEXT: 3 12 1.50 vpmullq %zmm16, %zmm17, %zmm19
256-
# CHECK-NEXT: 4 19 1.00 * vpmullq (%rax), %zmm17, %zmm19
257-
# CHECK-NEXT: 3 12 1.50 vpmullq %zmm16, %zmm17, %zmm19 {%k1}
258-
# CHECK-NEXT: 4 19 1.00 * vpmullq (%rax), %zmm17, %zmm19 {%k1}
259-
# CHECK-NEXT: 3 12 1.50 vpmullq %zmm16, %zmm17, %zmm19 {%k1} {z}
260-
# CHECK-NEXT: 4 19 1.00 * vpmullq (%rax), %zmm17, %zmm19 {%k1} {z}
255+
# CHECK-NEXT: 3 15 1.50 vpmullq %zmm16, %zmm17, %zmm19
256+
# CHECK-NEXT: 4 22 1.50 * vpmullq (%rax), %zmm17, %zmm19
257+
# CHECK-NEXT: 3 15 1.50 vpmullq %zmm16, %zmm17, %zmm19 {%k1}
258+
# CHECK-NEXT: 4 22 1.50 * vpmullq (%rax), %zmm17, %zmm19 {%k1}
259+
# CHECK-NEXT: 3 15 1.50 vpmullq %zmm16, %zmm17, %zmm19 {%k1} {z}
260+
# CHECK-NEXT: 4 22 1.50 * vpmullq (%rax), %zmm17, %zmm19 {%k1} {z}
261261
# CHECK-NEXT: 1 1 0.50 vxorpd %zmm16, %zmm17, %zmm19
262262
# CHECK-NEXT: 2 8 0.50 * vxorpd (%rax), %zmm17, %zmm19
263263
# CHECK-NEXT: 2 8 0.50 * vxorpd (%rax){1to8}, %zmm17, %zmm19
@@ -291,7 +291,7 @@ vxorps (%rax){1to16}, %zmm17, %zmm19 {z}{k1}
291291

292292
# CHECK: Resource pressure per iteration:
293293
# CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6] [7] [8] [9]
294-
# CHECK-NEXT: - - 55.50 7.00 37.50 37.50 - 81.50 - -
294+
# CHECK-NEXT: - - 57.00 4.00 37.50 37.50 - 83.00 - -
295295

296296
# CHECK: Resource pressure by instruction:
297297
# CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6] [7] [8] [9] Instructions:
@@ -397,11 +397,11 @@ vxorps (%rax){1to16}, %zmm17, %zmm19 {z}{k1}
397397
# CHECK-NEXT: - - - - - - - 1.00 - - vfpclassss $171, %xmm16, %k1 {%k2}
398398
# CHECK-NEXT: - - - - 0.50 0.50 - 1.00 - - vfpclassss $171, (%rax), %k1 {%k2}
399399
# CHECK-NEXT: - - 1.50 - - - - 1.50 - - vpmullq %zmm16, %zmm17, %zmm19
400-
# CHECK-NEXT: - - 1.00 1.00 0.50 0.50 - 1.00 - - vpmullq (%rax), %zmm17, %zmm19
400+
# CHECK-NEXT: - - 1.50 - 0.50 0.50 - 1.50 - - vpmullq (%rax), %zmm17, %zmm19
401401
# CHECK-NEXT: - - 1.50 - - - - 1.50 - - vpmullq %zmm16, %zmm17, %zmm19 {%k1}
402-
# CHECK-NEXT: - - 1.00 1.00 0.50 0.50 - 1.00 - - vpmullq (%rax), %zmm17, %zmm19 {%k1}
402+
# CHECK-NEXT: - - 1.50 - 0.50 0.50 - 1.50 - - vpmullq (%rax), %zmm17, %zmm19 {%k1}
403403
# CHECK-NEXT: - - 1.50 - - - - 1.50 - - vpmullq %zmm16, %zmm17, %zmm19 {%k1} {z}
404-
# CHECK-NEXT: - - 1.00 1.00 0.50 0.50 - 1.00 - - vpmullq (%rax), %zmm17, %zmm19 {%k1} {z}
404+
# CHECK-NEXT: - - 1.50 - 0.50 0.50 - 1.50 - - vpmullq (%rax), %zmm17, %zmm19 {%k1} {z}
405405
# CHECK-NEXT: - - 0.50 - - - - 0.50 - - vxorpd %zmm16, %zmm17, %zmm19
406406
# CHECK-NEXT: - - 0.50 - 0.50 0.50 - 0.50 - - vxorpd (%rax), %zmm17, %zmm19
407407
# CHECK-NEXT: - - 0.50 - 0.50 0.50 - 0.50 - - vxorpd (%rax){1to8}, %zmm17, %zmm19

llvm/test/tools/llvm-mca/X86/SkylakeServer/resources-avx512dqvl.s

Lines changed: 19 additions & 19 deletions
Original file line numberDiff line numberDiff line change
@@ -344,18 +344,18 @@ vxorps (%rax){1to8}, %ymm17, %ymm19 {z}{k1}
344344
# CHECK-NEXT: 1 4 1.00 vfpclassps $171, %ymm16, %k1 {%k2}
345345
# CHECK-NEXT: 2 11 1.00 * vfpclasspsy $171, (%rax), %k1 {%k2}
346346
# CHECK-NEXT: 2 11 1.00 * vfpclassps $171, (%rax){1to8}, %k1 {%k2}
347-
# CHECK-NEXT: 3 12 1.50 vpmullq %xmm16, %xmm17, %xmm19
348-
# CHECK-NEXT: 4 18 1.00 * vpmullq (%rax), %xmm17, %xmm19
349-
# CHECK-NEXT: 3 12 1.50 vpmullq %xmm16, %xmm17, %xmm19 {%k1}
350-
# CHECK-NEXT: 4 18 1.00 * vpmullq (%rax), %xmm17, %xmm19 {%k1}
351-
# CHECK-NEXT: 3 12 1.50 vpmullq %xmm16, %xmm17, %xmm19 {%k1} {z}
352-
# CHECK-NEXT: 4 18 1.00 * vpmullq (%rax), %xmm17, %xmm19 {%k1} {z}
353-
# CHECK-NEXT: 3 12 1.50 vpmullq %ymm16, %ymm17, %ymm19
354-
# CHECK-NEXT: 4 19 1.00 * vpmullq (%rax), %ymm17, %ymm19
355-
# CHECK-NEXT: 3 12 1.50 vpmullq %ymm16, %ymm17, %ymm19 {%k1}
356-
# CHECK-NEXT: 4 19 1.00 * vpmullq (%rax), %ymm17, %ymm19 {%k1}
357-
# CHECK-NEXT: 3 12 1.50 vpmullq %ymm16, %ymm17, %ymm19 {%k1} {z}
358-
# CHECK-NEXT: 4 19 1.00 * vpmullq (%rax), %ymm17, %ymm19 {%k1} {z}
347+
# CHECK-NEXT: 3 15 1.50 vpmullq %xmm16, %xmm17, %xmm19
348+
# CHECK-NEXT: 4 21 1.50 * vpmullq (%rax), %xmm17, %xmm19
349+
# CHECK-NEXT: 3 15 1.50 vpmullq %xmm16, %xmm17, %xmm19 {%k1}
350+
# CHECK-NEXT: 4 21 1.50 * vpmullq (%rax), %xmm17, %xmm19 {%k1}
351+
# CHECK-NEXT: 3 15 1.50 vpmullq %xmm16, %xmm17, %xmm19 {%k1} {z}
352+
# CHECK-NEXT: 4 21 1.50 * vpmullq (%rax), %xmm17, %xmm19 {%k1} {z}
353+
# CHECK-NEXT: 3 15 1.50 vpmullq %ymm16, %ymm17, %ymm19
354+
# CHECK-NEXT: 4 22 1.50 * vpmullq (%rax), %ymm17, %ymm19
355+
# CHECK-NEXT: 3 15 1.50 vpmullq %ymm16, %ymm17, %ymm19 {%k1}
356+
# CHECK-NEXT: 4 22 1.50 * vpmullq (%rax), %ymm17, %ymm19 {%k1}
357+
# CHECK-NEXT: 3 15 1.50 vpmullq %ymm16, %ymm17, %ymm19 {%k1} {z}
358+
# CHECK-NEXT: 4 22 1.50 * vpmullq (%rax), %ymm17, %ymm19 {%k1} {z}
359359
# CHECK-NEXT: 1 1 0.33 vxorpd %xmm16, %xmm17, %xmm19
360360
# CHECK-NEXT: 2 7 0.50 * vxorpd (%rax), %xmm17, %xmm19
361361
# CHECK-NEXT: 2 7 0.50 * vxorpd (%rax){1to2}, %xmm17, %xmm19
@@ -407,7 +407,7 @@ vxorps (%rax){1to8}, %ymm17, %ymm19 {z}{k1}
407407

408408
# CHECK: Resource pressure per iteration:
409409
# CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6] [7] [8] [9]
410-
# CHECK-NEXT: - - 64.00 64.00 59.00 59.00 - 76.00 - -
410+
# CHECK-NEXT: - - 67.00 67.00 59.00 59.00 - 70.00 - -
411411

412412
# CHECK: Resource pressure by instruction:
413413
# CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6] [7] [8] [9] Instructions:
@@ -544,17 +544,17 @@ vxorps (%rax){1to8}, %ymm17, %ymm19 {z}{k1}
544544
# CHECK-NEXT: - - - - 0.50 0.50 - 1.00 - - vfpclasspsy $171, (%rax), %k1 {%k2}
545545
# CHECK-NEXT: - - - - 0.50 0.50 - 1.00 - - vfpclassps $171, (%rax){1to8}, %k1 {%k2}
546546
# CHECK-NEXT: - - 1.50 1.50 - - - - - - vpmullq %xmm16, %xmm17, %xmm19
547-
# CHECK-NEXT: - - 1.00 1.00 0.50 0.50 - 1.00 - - vpmullq (%rax), %xmm17, %xmm19
547+
# CHECK-NEXT: - - 1.50 1.50 0.50 0.50 - - - - vpmullq (%rax), %xmm17, %xmm19
548548
# CHECK-NEXT: - - 1.50 1.50 - - - - - - vpmullq %xmm16, %xmm17, %xmm19 {%k1}
549-
# CHECK-NEXT: - - 1.00 1.00 0.50 0.50 - 1.00 - - vpmullq (%rax), %xmm17, %xmm19 {%k1}
549+
# CHECK-NEXT: - - 1.50 1.50 0.50 0.50 - - - - vpmullq (%rax), %xmm17, %xmm19 {%k1}
550550
# CHECK-NEXT: - - 1.50 1.50 - - - - - - vpmullq %xmm16, %xmm17, %xmm19 {%k1} {z}
551-
# CHECK-NEXT: - - 1.00 1.00 0.50 0.50 - 1.00 - - vpmullq (%rax), %xmm17, %xmm19 {%k1} {z}
551+
# CHECK-NEXT: - - 1.50 1.50 0.50 0.50 - - - - vpmullq (%rax), %xmm17, %xmm19 {%k1} {z}
552552
# CHECK-NEXT: - - 1.50 1.50 - - - - - - vpmullq %ymm16, %ymm17, %ymm19
553-
# CHECK-NEXT: - - 1.00 1.00 0.50 0.50 - 1.00 - - vpmullq (%rax), %ymm17, %ymm19
553+
# CHECK-NEXT: - - 1.50 1.50 0.50 0.50 - - - - vpmullq (%rax), %ymm17, %ymm19
554554
# CHECK-NEXT: - - 1.50 1.50 - - - - - - vpmullq %ymm16, %ymm17, %ymm19 {%k1}
555-
# CHECK-NEXT: - - 1.00 1.00 0.50 0.50 - 1.00 - - vpmullq (%rax), %ymm17, %ymm19 {%k1}
555+
# CHECK-NEXT: - - 1.50 1.50 0.50 0.50 - - - - vpmullq (%rax), %ymm17, %ymm19 {%k1}
556556
# CHECK-NEXT: - - 1.50 1.50 - - - - - - vpmullq %ymm16, %ymm17, %ymm19 {%k1} {z}
557-
# CHECK-NEXT: - - 1.00 1.00 0.50 0.50 - 1.00 - - vpmullq (%rax), %ymm17, %ymm19 {%k1} {z}
557+
# CHECK-NEXT: - - 1.50 1.50 0.50 0.50 - - - - vpmullq (%rax), %ymm17, %ymm19 {%k1} {z}
558558
# CHECK-NEXT: - - 0.33 0.33 - - - 0.33 - - vxorpd %xmm16, %xmm17, %xmm19
559559
# CHECK-NEXT: - - 0.33 0.33 0.50 0.50 - 0.33 - - vxorpd (%rax), %xmm17, %xmm19
560560
# CHECK-NEXT: - - 0.33 0.33 0.50 0.50 - 0.33 - - vxorpd (%rax){1to2}, %xmm17, %xmm19

0 commit comments

Comments
 (0)