With the patch for PR37535 applied to mainline GCC the bootstrap still fails
due to a miscompilation. The problem is that r52 is assigned to r6 what
collides with an INSN loading r6 with a parameter for a CALL.
I think the problem is created in ira_flattening. Allocno a3 is a parent of
a87. For a87 the conflict with r6 is properly recorded. Afterwards ira_flatting
merges a87 into a3 without propagating the conflict_hard_regs set. So the
conflict with r6 is lost and ira_reassign_pseudos later on assigns r52 to hard
reg r6.
There is already code in ira_flattening which propagates the conflict sets. But
the code is only enabled if propagate_p is true. This in turn seems only to get
set if the loop once merged allocnos with different regnos?! I unfortunately
don't understand this enough to come up with a patch. So I better leave that to
you :) Please contact me if you need more information.
I wasn't able to reduce the testcase (genattrtab) since it is quite difficult
to see from the code if the miscompile occurred or not.
;; a3(r52,l0) conflicts: a1(r87,l0) ......
;; total conflict hard regs: 0-6 14
;; conflict hard regs: 0-5 14
;; a87(r52,l2) conflicts: a86(r48,l2) .....
;; total conflict hard regs: 0-6 14
;; conflict hard regs: 0-6 14
...
Moving ranges of a87r52 to a3r52: [245..248] [234..243] [229..232] [218..227]
[173..189] [126..169]
...
Try assign 52(a3), cost=268: reassign to 6
--
Summary: Bootstrap failure due to miscompilation of genattrtab
Product: gcc
Version: 4.4.0
Status: UNCONFIRMED
Severity: critical
Priority: P3
Component: middle-end
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: krebbel at gcc dot gnu dot org
GCC build triplet: s390x-ibm-linux
GCC host triplet: s390x-ibm-linux
GCC target triplet: s390x-ibm-linux
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37674
