powerpc: Another umaddditi4 fix [PR108862]
authorJakub Jelinek <jakub@redhat.com>
Mon, 20 Feb 2023 21:07:57 +0000 (22:07 +0100)
committerJakub Jelinek <jakub@redhat.com>
Mon, 20 Feb 2023 21:07:57 +0000 (22:07 +0100)
commitcd8b4fae52d02541c2d8bd2200caad3812f37368
tree6642be63f116051dd941e66f25bba2b3d60bdf57
parente86963813750ca0e457300d51926691ba24efb27
powerpc: Another umaddditi4 fix [PR108862]

The following testcase is miscompiled on powerpc64le-linux with
-O2 -mcpu=power9.  The problem is that gen_umaddditi4 is called with
the same TImode register for both op0 and op3, and maddlddi4
overwrites the low half of op0 before the low half of op3 is read,
so when they are the same register it reads the result of maddlddi4.

The following patch fixes that by swapping maddlddi4 and
umadddi4_highpart{,_le} during expansion, as the latter writes into
a temporary pseudo and so can't change anything maddlddi4 depends on.

2023-02-20  Jakub Jelinek  <jakub2redhat.com>

PR target/108862
* config/rs6000/rs6000.md (umaddditi4): Swap gen_maddlddi4 with
gen_umadddi4_highpart{,_le}.

* gcc.dg/pr108862.c: New test.
* gcc.target/powerpc/pr108862.c: New test.
gcc/config/rs6000/rs6000.md
gcc/testsuite/gcc.dg/pr108862.c [new file with mode: 0644]
gcc/testsuite/gcc.target/powerpc/pr108862.c [new file with mode: 0644]