[PyTorch] Make IValue::toTensor() inlineable (#53213)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53213
The failure path for toTensor() is fairly long because it has to stringify tagKind() and construct a std::string. Forcibly outlining it should allow inlining the happy path.
ghstack-source-id: 123012703
Test Plan:
1) Compare perf profile on AdIndexer benchmark before/after --
toTensor frames no longer show up, demonstrating inlining
2) Compare perf stat results on AdIndexer benchmark before/after:
Before:
```
17,104.66 msec task-clock # 0.999 CPUs utilized ( +- 0.26% )
3,666 context-switches # 0.214 K/sec ( +- 18.53% )
3 cpu-migrations # 0.000 K/sec ( +- 6.25% )
102,745 page-faults # 0.006 M/sec ( +- 0.47% )
33,860,604,938 cycles # 1.980 GHz ( +- 0.25% ) (50.02%)
69,514,752,652 instructions # 2.05 insn per cycle ( +- 0.06% ) (50.01%)
11,280,877,966 branches # 659.521 M/sec ( +- 0.11% ) (50.01%)
75,739,099 branch-misses # 0.67% of all branches ( +- 0.98% ) (50.03%)
# Table of individual measurements:
17.2467 (+0.1172) #
17.0014 (-0.1280) #
17.2134 (+0.0840) #
17.0951 (-0.0343) #
17.0905 (-0.0389) #
# Final result:
17.1294 +- 0.0447 seconds time elapsed ( +- 0.26% )
```
After:
```
16,910.66 msec task-clock # 0.999 CPUs utilized ( +- 0.27% )
3,495 context-switches # 0.207 K/sec ( +- 18.34% )
3 cpu-migrations # 0.000 K/sec ( +- 6.25% )
101,769 page-faults # 0.006 M/sec ( +- 0.45% )
33,460,776,952 cycles # 1.979 GHz ( +- 0.28% ) (50.03%)
69,243,346,925 instructions # 2.07 insn per cycle ( +- 0.17% ) (50.02%)
11,229,930,860 branches # 664.074 M/sec ( +- 0.14% ) (50.03%)
72,273,324 branch-misses # 0.64% of all branches ( +- 0.55% ) (50.03%)
# Table of individual measurements:
16.9530 (+0.0246) #
17.0898 (+0.1614) #
16.8493 (-0.0791) #
16.8282 (-0.1002) #
16.9217 (-0.0067) #
# Final result:
16.9284 +- 0.0464 seconds time elapsed ( +- 0.27% )
```
1.1% cycles win, 0.38% instructions win, both apparently outside noise level
Reviewed By: smessmer
Differential Revision: D26793481
fbshipit-source-id: b035b3ad20f9e22ae738d91163641031b1130ce6
