https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81044
Bug ID: 81044
Summary: fold-const change in svn 248447 breaks PowerPC spec
2006 hmmr
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: meissner at gcc dot gnu.org
Target Milestone: ---
Created attachment 41527
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=41527&action=edit
hmmio.i file that was used.
I was running spec 2006 benchmarks on a little endian power8 system, and I
noticed that the 456.hmmer benchmark doesn't run starting with subversion id
248447.
Note, in order to get it to fail, you need to use the -mveclibabi=mass option
to vectorize math functions and use the MASS 8.1.3 libraries. I traced this
down the hmmio.c module getting miscompiled when -O2 or higher is used. Note,
the hmmio.c function does not call any of the vectorized math functions, but it
needs other modules to use them.
I haven't looked at all of the differences, but the changes for the code
generated for line 664 seem to be wrong. It is collapsing two different 32-bit
compares into one 64-bit compare and then doing a MCRF (move condition
register) instruction to duplicate the test.
This is the ChangeLog for the 248446 change:
2017-05-25 Marc Glisse <[email protected]>
* fold-const.c (fold_binary_loc) [(A & C) == D]: Remove transformation.
* match.pd (X == C): Rewrite it here.
(with_possible_nonzero_bits, with_possible_nonzero_bits2,
with_certain_nonzero_bits2): New predicates.
* tree-ssanames.c (get_nonzero_bits): Handle INTEGER_CST.