https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113384
--- Comment #3 from John David Anglin <danglin at gcc dot gnu.org> ---
These fails only when generating PA 2.0 code. The fails don't occur when
generating the default PA 1.1 code.
Wrong code is generated for this line in dependency_58.f90:
b%im = a%im - 0.5
b%re is set instead of b%im.
We have the following PA 1.1 assembly code:
0x00010ab0 <MAIN__+852>: fsub,sgl fr23,fr22,fr22
0x00010ab4 <MAIN__+856>: ldo 64(r20),r20
0x00010ab8 <MAIN__+860>: depw,z r20,28,29,r20
0x00010abc <MAIN__+864>: ldo 8(r3),r21
0x00010ac0 <MAIN__+868>: add,l r21,r20,r20
0x00010ac4 <MAIN__+872>: fstw fr22,0(r20)
This is PA 1.1 code:
fsub,sgl %fr23L,%fr22L,%fr22L
ldo 8(%r3),%r20
ldo 800(%r20),%r20
ldo 4(%r20),%r20
zdep %r21,28,29,%r21
addl %r21,%r20,%r20
fstws %fr22L,0(%r20)
We miss this ldo instruction in PA 2.0:
ldo 4(%r20),%r20
It adds 4 to the address to access the imaginary part of b.
