https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85537
Thomas Koenig <tkoenig at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |tkoenig at gcc dot gnu.org
--- Comment #7 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
Created attachment 46016
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=46016&action=edit
Assembly for test case
Looking at the assembly of the test case in comment #6 (with -g -fverbose-asm
only):
.LBE2:
# a.f90:29: a = a + 1
.loc 1 29 0
movl (%rbx), %eax # CHAIN.4_10(D)->a, _11
addl $1, %eax #, _12
movl %eax, (%rbx) # _12, CHAIN.4_10(D)->a
When called through a procedure pointer, rbx no longer contains the
address of a.
Looking further up, we have
movq %r10, %rbx # CHAIN.4, CHAIN.4
movq %r10, -552(%rbp) # CHAIN.4,
.LBB2:
so %r10 is clobbered. This happens in the call of _gfortran_st_write.
Looking at the ABI, %r10 is not supposed to be preserved.
Data point: This works on POWER9.
This looks like a target bug to me.
