https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102147
Bug ID: 102147
Summary: IRA dependent on 32-bit vs 64-bit register size
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: dje at gcc dot gnu.org
Target Milestone: ---
IRA heuristics chooses different data structure encodings based on the register
size, and this produces different register allocation results.
This was discovered by a GCC bootstrap comparison failure of tree-vect-slp.c
when using a 32 bit compiler to bootstrap a 64 bit compiler.
A difference occurs in ira-conflicts.c: build_object_conflicts(), for
the same object with the same properties (i.e., min, max and px are the same),
the function ira_conflict_vector_profitable_p() will return 1 by
stage1-gcc and 0 by stage2-gcc.
stage1-gcc: build_object_conflict obj140(a140) px=4 min=3 max=139
profitable_p=1
stage2-gcc: build_object_conflict obj140(a140) px=4 min=3 max=139
profitable_p=0
That's because the size of ira_object_t being a pointer is different
in stage1-gcc (which is 32bit) and stage2-gcc (which is 64bit).
My colleagues at ATOS and I aren't completely certain how this difference
causes different conflict / allocation behavior because it seems that it should
be an
optimization.
Should the data structure choice / algorithm choice depend on pointer size?
Are the two algorithms supposed to generate the same results?