This pass does simple forward propagation and simplification when
an operand of an insn can only come from a single def. The pass has a
very good potential of catching simplifications currently done by
inter-basic-block CSE (-fcse-follow-jumps and -fcse-skip-blocks) and
combine: however, it is much faster than inter-basic-block CSE and runs
much earlier than combine.
Unlike other attempts to supersede CSE path following, however, I
did not write a general optimization framework like GCSE: instead,
looking at what kinds of simplifications are only made by CSE, I tried
to implement them in a very generic way. Take for example Bug 13724: my
fix for it was to make CSE recognize the high word of (zero_extend:DI
(reg:SI M)), and simplify it to zero. After committing it, Roger Sayle
also tried adding the same to simplify-rtx.c, but it was ineffective.
This pass, instead, can use simplify-rtx.c much more effectively than CSE.
The patch currently has performance regressions in mesa and in
crafty, but overall slightly improves SPEC with a saving of 2-3% in
bootstrap and compilation time.
Personnel
* Paolo Bonzini
Delivery Date
Stage 1
I actually hoped that the pass itself can go in 4.1, being disabled
by default, but it seems unlikely at this point.
Benefits
Improved compilation time, incentive for moving optimizations from
combine.c to simplify-rtx.c.
Dependencies
* None.
Modifications Required
The regressions are still to be investigated; but most of the
modifications will most likely be in simplify-rtx.c (pouring
simplifications in there from combine), or may involve more
sophisticated analysis of available expressions in the forward
propagation pass.
