Summary:
The macros that build `c10::Error` compute the stack trace at the point of throwing, which is then returned as part of the `what()`. If `what()` is never called, which is the case for most exceptions (since logging is throttled), the cost of computing the stack trace was wasted.
By far, the most expensive part of computing the stack trace is its symbolization; just unwinding the stack and collecting the instruction addresses is comparatively cheap. We can thus defer the symbolization to first invocation of `what()`.
Test Plan:
Added unit tests exercising the lazy nature of `what()`.
Ran an adfinder canary: https://www.internalfb.com/intern/ads/canary/460118801509424346
We can see that the cost of symbolization is obliterated (meaning that `what()` is virtually never called, as expected):
{F1496627896}
Differential Revision: D57128632
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125787
Approved by: https://github.com/huydhn
Summary:
The macros that build `c10::Error` compute the stack trace at the point of throwing, which is then returned as part of the `what()`. If `what()` is never called, which is the case for most exceptions (since logging is throttled), the cost of computing the stack trace was wasted.
By far, the most expensive part of computing the stack trace is its symbolization; just unwinding the stack and collecting the instruction addresses is comparatively cheap. We can thus defer the symbolization to first invocation of `what()`.
Test Plan:
Added unit tests exercising the lazy nature of `what()`.
Ran an adfinder canary: https://www.internalfb.com/intern/ads/canary/460118801509424346
We can see that the cost of symbolization is obliterated (meaning that `what()` is virtually never called, as expected):
{F1496627896}
Reviewed By: ezyang
Differential Revision: D56586844
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125682
Approved by: https://github.com/ezyang
