thanks for the fix. It works fine for me. David
On Sat, Jun 23, 2012 at 11:00 AM, Nathan Sidwell <nat...@acm.org> wrote: > This patch changes the coverage unlinking to be less aggressive. As David > pointed out, when experimenting with random optimization options along side > -fuse-coverage and -frandom-seed, it is inconvenient for gcc to remove the > coverage data file. The reason it was doing so, is because -frandom-seed > prevents a unique stamp being placed in the coverage data to distinguish > execution of the newly optimized program from execution of the older > incarnation of it. libgcov would get confused and give errors about > mismatched function checksums. > > This patch tweaks things. With -fuse-coverage we read in the coverage data > file and its stamp. We can feed that stamp into the stamp to be used for > the output coverage data, and hence effectively distinguish generations of > coverage data even when -frandom-seed prevents randomness. We also don't > need to delete the data file if we're not producing coverage data for the > new compilation. > > Manual testing shows this patch is preventing deletion of the data file in > the circumstances I understand David's using. David, care to try this patch > out to see if it resolves your use case? > > The changes to tree.c merely extend the crc32 code to allow it to generate a > crc from a 32-bit unsigned, in addition to the byte variant already present. > > testing on i686-pc-linux-gnu shows no regressions. > > nathan
