On 2016-02-12, at 11:28 AM, Bernd Schmidt wrote: > I'm attaching a full patch. John David Anglin is currently running tests on > PA which may take a while. He's confirmed it fixes his testcase, and the > earlier 67781 testcase seems to be cured as well (only looked at generated > assembly, don't have a full cross environment - Thomas, can you verify?) > I'm also doing an x86_64 test run. Ok everywhere if it passes? I'm also > attaching a version of the testcase (LGPL, from ffmpeg) which I'd also apply > to gcc.dg/torture if David (or Thomas) can confirm it works on a big-endian > target after the fix.
I have tested the patch on hppa2.0w-hp-hpux11.11 and hppa-unknown-linux-gnu so far. It fixes the original testcase and I didn't see any regressions related to this change. I also have a private communication that it also fixes the mips testcase from Debian bug #813858. As far as the avcrc.c reduced testcase, it didn't trigger the original bug on hppa-unknown-linux-gnu. > > > There is room for more improvement in this code - hopefully we can get to > that in gcc-7. As mentioned, it does not use the bytepos in the final > replacement. We could extend the code to recognize smaller sub-bswaps at an > offset. Currently, only one of the two rotates involved in a 32-bit bswap is > recognized. > > Also, for the crc testcase, the transformation actually pessimizes the code > by inserting an extra memory load. If we already have loaded a larger value, > we should build the subpart using a shift and convert. I noticed that. I'm testing a patch to add bswap patterns to pa.md. The PA 2.0 architecture manual actually gives some optimized sequences for byte swap operations with half-word, word and double-word operands. The compiler might be able to duplicate the half-word and word sequences, but there is no way it can duplicate the double-word sequence. It uses half-word instructions that aren't exposed in PA backend. Dave -- John David Anglin dave.ang...@bell.net
