[phase ordering] opt --std-compile-opts misses optimizations on first pass

[llvmbot]

Bugzilla Link	2069
Version	trunk
OS	Linux
Attachments	Poorly optimized test case (hacked up pcregrep).
Reporter	LLVM Bugzilla Contributor
CC	@nlewycky

Extended Description

(I'm not sure whether I should consider this a bug, feedback appreciated)

On the given test file opt is significantly more effective (in terms of emmitted bytecode size) when run twice.

--
ddunbar@ozzy:lli$ llvm-as -o=- pcregrep.ll | opt --std-compile-opts | llvm-dis | wc -l
1011
ddunbar@ozzy:lli$ llvm-as -o=- pcregrep.ll | opt --std-compile-opts | opt --std-compile-opts | llvm-dis | wc -l
77

Sorry for the complexity of the test case, it is hand trimmed from an llvm built pcregrep and there is a lot of dead code in the initial test. I don't have an easy way to trim further while still reproducing the bug.

[nlewycky]

According to opt -std-compile-opts -stats (in the second round), the major winner is instcombine with:

408 instcombine - Number of dead inst eliminated

In order to get this (well, 405 of them) to happen, you need opt -globalopt -scalarrepl -instcombine in the second round.

[lattner]

Is this still a problem after r130241?

[nlewycky]

Yep, still a problem.

$ opt -O2 pcregrep.bc -o - -S | wc -l
1033
$ opt -O2 pcregrep.bc -o - | opt -O2 -o - -S | wc -l
26

Stats from the second opt run are:

1 cgscc-passmgr - Maximum CGSCCPassMgr iterations on one SCC
12 correlated-value-propagation - Number of phis propagated
4 deadtypeelim - Number of unused typenames removed from symtab
91 early-cse - Number of instructions simplified or DCE'd
12 globalopt - Number of globals localized
12 globalopt - Number of globals marked unnamed_addr
1 gvn - Number of instructions deleted
1 gvn - Number of loads deleted
5 instcombine - Number of dead inst eliminated
1 jump-threading - Number of jumps threaded
15 jump-threading - Number of terminators folded
12 scalarrepl - Number of allocas promoted
4 simplify-libcalls - Number of library calls simplified
50 simplifycfg - Number of blocks simplified
17 strip-dead-prototypes - Number of dead prototypes removed

It's those 12 globalopt optz'ns we need to nail before early-cse (formerly instcombine) can clean up the mess.