[PATCH] D32563: Add LiveRangeShrink pass to shrink live range within BB.

[via llvm-commits]

This patch caused some fairly significant regressions on our GPU target. While I can’t guarantee this is all of the problems, here is a particular case that exemplified to me “things this pass is doing that it should not”.


Original code:

PHI
PHI
PHI
[…]
foo = REG_SEQUENCE […] // inputs come from the phi
STORE foo
bar = REG_SEQUENCE […]
STORE bar
foobar = REG_SEQUENCE […]
STORE foobar
[…]

New code:

PHI
PHI
PHI
[…]
foo = REG_SEQUENCE […]
bar = REG_SEQUENCE […]
foobar = REG_SEQUENCE […]
[…]
STORE bar
STORE foo
STORE foobar

This causes a vast increase in register count because the reg_sequences constrain the registers to be adjacent far longer than necessary.

Two things are going on here that seem dubious.

1. why are we moving them up… to a PHI??? does this make sense for any instruction?
2. why are we moving reg_sequences??? away from their use, at that?

I can’t imagine this makes sense on any target.

—escha