On 03/03/2017 08:36 AM, Bernd Schmidt wrote:
Reload is designed in a way to avoid cycles and to process all reloads
for an insn in order of ascending class so as to avoid this kind of
issue. With LRA I'm really not sure how to fix this properly, but the
following patch seems to cure the PR at least, by recognizing when
we're about to spill a reload reg whose assigned class contains only
contains a single register.
I had conversations with several people who worked on different RAs and
got a conclusion that looping is a pretty general problem in iterative
classical RAs. Unfortunately, it is not discussed in a literature and
ad-hoc approaches are used (that is why it is probably not discussed in
scientific literature).
On first stages of LRA development, looping was a very frequent bug and
also add-hoc approaches were used to fix it. The major way is to
predict that given decision can result in repeating patterns and to
avoid such decision (e.g. lra_constraints.c::process_alt_operands
contains code).
Reasons for looping can have different nature and now I have no
systematic solution on my mind to avoid all of them.
Bootstrapped and tested on x86_64-linux, ok?
Yes. Thank you for working on this, Bernd. The issue mentioned in
FIXME part of the comment can be addressed later if we have cases for
which your fix will be not enough.
