| Issue |
143005
|
| Summary |
[RISCV] 507.cactuBSSN_r regression after bidirectional scheduling/register pressure tracking
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
sihuan
|
Initially noticed that commit 7b09d7b (20.1.5), compared to commit c9a6e993, had about a 50% increase in dynamic instruction count on 507.cactuBSSN_r.
Further testing shows that this regression is introduced by commit 9122c52 (#115445).
Below are some of the test results:
Compilation flags used:
```
CFLAGS := -std=c99 -march=rv64gc -O3 -DSPEC -DSPEC_CPU -DNDEBUG -DSPEC_AUTO_SUPPRESS_OPENMP -Iinclude -DCCODE -DSPEC_LP64
CXXFLAGS := -march=rv64gc -O3 -DSPEC -DSPEC_CPU -DNDEBUG -DSPEC_AUTO_SUPPRESS_OPENMP -Iinclude -DCCODE -DSPEC_LP64
FFLAGS := -march=rv64gc -O3 -Iinclude
LDFLAGS := -march=rv64gc -O3
```
On commit 9122c52 :
```
perf stat ./9122c5235ec8 spec_test.par 1>/dev/null
Performance counter stats for './9122c5235ec8 spec_test.par':
27237.56 msec task-clock # 0.991 CPUs utilized
3034 context-switches # 0.111 K/sec
4 cpu-migrations # 0.000 K/sec
25671 page-faults # 0.942 K/sec
50278526154 cycles # 1.846 GHz
39011759196 instructions # 0.78 insn per cycle
350744284 branches # 12.877 M/sec
6958541 branch-misses # 1.98% of all branches
27.494980251 seconds time elapsed
26.937902000 seconds user
0.275487000 seconds sys
```
On commit 5bbe63e (previous commit):
```
perf stat ./5bbe63ec9122 spec_test.par 1>/dev/null
Performance counter stats for './5bbe63ec9122 spec_test.par':
23636.29 msec task-clock # 0.995 CPUs utilized
2099 context-switches # 0.089 K/sec
2 cpu-migrations # 0.000 K/sec
25670 page-faults # 0.001 M/sec
43618927076 cycles # 1.845 GHz
28844201052 instructions # 0.66 insn per cycle
348492599 branches # 14.744 M/sec
8688534 branch-misses # 2.49% of all branches
23.758763429 seconds time elapsed
23.395399000 seconds user
0.227527000 seconds sys
```
It is worth noting that, although the instruction count increased by 35%, the runtime regressed by only 16%.
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
