ian-twilightcoder wrote: > @ian-twilightcoder this change seemed to cause our internal builds to run out > of source location space. > > ``` > remark: source manager location address space usage: [-Rsloc-usage] > note: 408559B in local locations, 2146921126B in locations loaded from AST > files, for a total of 2147329685B (99% of available space) > ``` > > The main offenders from the follow-ups are: > > * module map files for standard library and C/C++ runtime headers, each > entered a few thousand times, > * various headers without header guards, e.g. libc++'s `__config` and others, > * `arm_neon.h`. > Without looking deeper into what the change does, this might be an expected > outcome here. However, let me know if something already looks off from this > small description (e.g. the module maps entered so many times is surprising, > I'll need to see if this happened before). > > I will investigate this further, but we might not be able to change how our > builds work quickly. Would it be ok to revert this change or put the new > behavior under a flag to give us some more time to investigate?
That's surprising, the code to re-enter files is `HeaderSearch::ShouldEnterIncludeFile`, and I only changed the behavior there for `#import`. The `#include` behavior is still governed by `#pragma once` and `FileInfo.getControllingMacro`, and that isn't changed. Are you sure it's this change? Do you have a repro case? https://github.com/llvm/llvm-project/pull/83660 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
