[polly] r248667 - [FIX] Use the surrounding loop for non-affine SCoP regions

Tobias Grosser via llvm-commits llvm-commits at lists.llvm.org
Sat Sep 26 09:07:30 PDT 2015


On 09/26/2015 05:32 PM, Johannes Doerfert wrote:
> Could you check if it is really caused by non-affine regions that span
> the whole SCoP? (E.g., add an unfeasible assumption for these in the
> domain or schedule generation as they are detected there anyway).

Not fully sure what you mean. I just tried on one benchmark if disabling
support for non-affine branches avoids the miscompile and it indeed
does.

This LNT command fails for me:

mysandbox/bin/lnt runtest nt --sandbox /tmp/bar --cc /home/grosser/Projects/polly/build/bin/clang --cxx /home/grosser/Projects/polly/build/bin/clang++ --test-suite test-suite/ -j1 --only-test MultiSource/Applications/sgefa --cflag="-Xclang -load" --cflag="-Xclang /home/grosser/Projects/polly/build/lib/LLVMPolly.so" --mllvm=-polly --multisample=1  --mllvm=-polly-no-early-exit --mllvm=-polly-detect-unprofitable --mllvm=-polly-allow-nonaffine-branches=true

whereas this one passes:

mysandbox/bin/lnt runtest nt --sandbox /tmp/bar --cc /home/grosser/Projects/polly/build/bin/clang --cxx /home/grosser/Projects/polly/build/bin/clang++ --test-suite test-suite/ -j1 --only-test MultiSource/Applications/sgefa --cflag="-Xclang -load" --cflag="-Xclang /home/grosser/Projects/polly/build/lib/LLVMPolly.so" --mllvm=-polly --multisample=1  --mllvm=-polly-no-early-exit --mllvm=-polly-detect-unprofitable --mllvm=-polly-allow-nonaffine-branches=false

Best,
Tobias


More information about the llvm-commits mailing list