commit b63b2677bc7e3aecff671a23f387004d2abc8e71
author Geoffrey Martin-Noble <gcmn@google.com> Fri May 05 08:28:26 2023 -0700
committer GitHub <noreply@github.com> Fri May 05 08:28:26 2023 -0700
tree e0cbe989de2285d60f226017b7ff28a7db833d26
parent 8bb2d31f39ec346c7cdc9a50b244aebdbdfdd159

Use remote ccache in mac x86_64 build (#13418)

Fully cached rebuild appears to take 20m for the build step,
compared to 1h8m for the uncached build with remote cache writes
(attempts 2 and 1 of
[this workflow
run](https://github.com/openxla/iree/actions/runs/4886873002/jobs/8724117607)).
That's unfortunately quite a bit slower than when we had caching that
restored a full ccache directory previously (e.g.
[this
run](https://github.com/openxla/iree/actions/runs/4658305324/jobs/8243824546)).
The latter was disabled in https://github.com/openxla/iree/pull/12905
because the actions/cache restore was extremely unreliable, timing out
most of the time.

The latency difference appears to be due to the cache hit rate:

```
Cacheable calls:   4850 / 4865 (99.69%)
  Hits:            4734 / 4850 (97.61%)
    Direct:        4387 / 4734 (92.67%)
    Preprocessed:   347 / 4734 ( 7.33%)
  Misses:           116 / 4850 ( 2.39%)
Uncacheable calls:   15 / 4865 ( 0.31%)
Local storage:
  Cache size (GB): 2.94 / 4.00 (73.50%)
```

vs

```
Cacheable calls:    4922 / 4937 (99.70%)
  Hits:             4080 / 4922 (82.89%)
    Direct:         3729 / 4080 (91.40%)
    Preprocessed:    351 / 4080 ( 8.60%)
  Misses:            842 / 4922 (17.11%)
Uncacheable calls:    15 / 4937 ( 0.30%)
Local storage:
  Cache size (GiB):  0.0 /  5.0 ( 0.00%)
Remote storage:
  Hits:             4080 / 4922 (82.89%)
  Misses:            842 / 4922 (17.11%)
```

We'll need to debug why we're seeing the 17% cache miss rate (I could
wildly speculate), but this is still a sizable improvement.

Part of https://github.com/openxla/iree/issues/13028