[PATCH] D15208: Patch for inline abort code generation
Dan Austin via cfe-commits
cfe-commits at lists.llvm.org
Tue Dec 22 10:57:04 PST 2015
danielaustin updated this revision to Diff 43459.
danielaustin added a comment.
Test case added which looks for the existence of multiple trap basic blocks in generated LLVM-IR with optimization turned on. Flag renamed (-fmerge-traps/-fno-merge-traps) and placed in a more sensible location (CodeGenOptions.def). Trap inlining logic changed to work with most recent version of LLVM (as opposed to the one packaged with AOSP). Without turning optimization off in the function containing the trap, there were cases that the abort calls were merged to the end of the function. Turning off inlining and optimization for a function which is not to merge traps appears to be working, but further analysis is in process.
This version results in slightly bigger code (~10% per function) than the original version, but has the benefit of not breaking basic blocks.
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 6444 bytes
Desc: not available
More information about the cfe-commits