f594ecc453
Add a new LargeBlockInfo helper, which is just a wrapper around
...
a trivial dense map. Use this in RewriteSingleStoreAlloca to
avoid aggressively rescanning blocks over and over again. This
fixes PR2925, speeding up mem2reg on the testcase in that bug
from 4.56s to 0.02s in a debug build on my machine.
llvm-svn: 58227
2008-10-27 06:05:26 +00:00
e9ecc68d8f
API changes for class Use size reduction, wave 1.
...
Specifically, introduction of XXX::Create methods
for Users that have a potentially variable number of
Uses.
llvm-svn: 49277
2008-04-06 20:25:17 +00:00
7698bfbe16
Update -mem2reg to use succ_iterator instead of iterating across TerminatorInst
...
successors. This makes it support nounwind.
llvm-svn: 48320
2008-03-13 02:42:41 +00:00
a838141957
Make RenamePass faster by making the 'is this a new phi node'
...
check more intelligent. This speeds up mem2reg from 5.29s to
0.79s on a synthetic testcase with tons of predecessors and
phi nodes.
llvm-svn: 46767
2008-02-05 21:26:23 +00:00
f3ebc3f3d2
Remove attribution from file headers, per discussion on llvmdev.
...
llvm-svn: 45418
2007-12-29 20:36:04 +00:00
d568767ecb
Finishing initial docs for all transformations in Passes.html.
...
Also cleaned up some comments in source files.
llvm-svn: 43674
2007-11-04 16:15:04 +00:00
0625bd6472
Merge DenseMapKeyInfo & DenseMapValueInfo into DenseMapInfo
...
Add a new DenseMapInfo::isEqual method to allow clients to redefine
the equality predicate used when probing the hash table.
llvm-svn: 42042
2007-09-17 18:34:04 +00:00
24fb6b2f8c
Don't promote volatile loads/stores. This is needed (for example) to handle setjmp/longjmp properly.
...
This fixes PR1520.
llvm-svn: 41461
2007-08-26 21:43:30 +00:00
edce70d2fe
rewrite the code used to construct pruned SSA form with the IDF method.
...
In the old way, we computed and inserted phi nodes for the whole IDF of
the definitions of the alloca, then computed which ones were dead and
removed them.
In the new method, we first compute the region where the value is live,
and use that information to only insert phi nodes that are live. This
eliminates the need to compute liveness later, and stops the algorithm
from inserting a bunch of phis which it then later removes.
This speeds up the testcase in PR1432 from 2.00s to 0.15s (14x) in a
release build and 6.84s->0.50s (14x) in a debug build.
llvm-svn: 40825
2007-08-04 22:50:14 +00:00
d91576b01e
Factor out a whole bunch of code into it's own method.
...
llvm-svn: 40824
2007-08-04 21:14:29 +00:00
4e1b4140eb
Use getNumPreds(BB) instead of computing them manually. This is a very small but
...
measurable speedup.
llvm-svn: 40823
2007-08-04 21:06:15 +00:00
b6a4ba808b
Change the rename pass to be "tail recursive", only adding N-1 successors
...
to the worklist, and handling the last one with a 'tail call'. This speeds
up PR1432 from 2.0578s to 2.0012s (2.8%)
llvm-svn: 40822
2007-08-04 20:40:27 +00:00
840259c8d3
cache computation of #preds for a BB. This speeds up
...
mem2reg from 2.0742->2.0522s on PR1432.
llvm-svn: 40821
2007-08-04 20:24:50 +00:00
050bac4bed
reserve operand space for phi nodes when we insert them.
...
llvm-svn: 40820
2007-08-04 20:14:34 +00:00
9318785df5
use continue to avoid nesting, no functionality change.
...
llvm-svn: 40819
2007-08-04 20:07:06 +00:00
6b04ecbaf9
Promoting allocas with the 'single store' fastpath is
...
faster than with the 'local to a block' fastpath. This speeds
up PR1432 from 2.1232 to 2.0686s (2.6%)
llvm-svn: 40818
2007-08-04 20:03:23 +00:00
4a930f9444
When PromoteLocallyUsedAllocas promoted allocas, it didn't remember
...
to increment NumLocalPromoted, and didn't actually delete the
dead alloca, leading to an extra iteration of mem2reg.
llvm-svn: 40817
2007-08-04 20:01:43 +00:00
63c039780c
std::map -> DenseMap
...
llvm-svn: 40816
2007-08-04 19:52:20 +00:00
7d382f7680
fix a logic bug where we wouldn't promote single store allocas if the
...
stored value was a non-instruction value. Doh.
This increase the # single store allocas from 8982 to 9026, and
speeds up mem2reg on the testcase in PR1432 from 2.17 to 2.13s.
llvm-svn: 40813
2007-08-04 02:45:02 +00:00
1b215f0661
When we do the single-store optimization, delete both the store
...
and the alloca so they don't get reprocessed.
This speeds up PR1432 from 2.20s to 2.17s.
llvm-svn: 40812
2007-08-04 02:38:38 +00:00
862f125457
Three improvements:
...
1. Check for revisiting a block before checking domination, which is faster.
2. If the stored value isn't an instruction, we don't have to check for domination.
3. If we have a value used in the same block more than once, make sure to remove the
block from the UsingBlocks vector. Not doing so forces us to go through the slow
path for the alloca.
The combination of these improvements increases the number of allocas on the fastpath
from 8935 to 8982 on PR1432. This speeds it up from 2.90s to 2.20s (31%)
llvm-svn: 40811
2007-08-04 02:32:22 +00:00
ae1e00eb36
switch from using a std::set to using a SmallPtrSet. This speeds up the
...
testcase in PR1432 from 6.33s to 2.90s (2.22x)
llvm-svn: 40810
2007-08-04 02:21:22 +00:00
9181801bb7
In mem2reg, when handling the single-store case, make sure to remove
...
a using block from the list if we handle it. Not doing this caused us
to not be able to promote (with the fast path) allocas which have uses (whoops).
This increases the # allocas hitting this fastpath from 4042 to 8935 on the
testcase in PR1432, speeding up mem2reg by 2.6x
llvm-svn: 40809
2007-08-04 02:15:24 +00:00
886a41a007
split rewriting of single-store allocas into its own
...
method.
llvm-svn: 40806
2007-08-04 01:47:41 +00:00
3cede09c67
refactor some code to shrink PromoteMem2Reg::run a bit
...
llvm-svn: 40805
2007-08-04 01:41:18 +00:00
d524537fe9
add a typedef, no other change.
...
llvm-svn: 40804
2007-08-04 01:19:38 +00:00
df138be527
avoid an unneeded vector copy. This speeds up mem2reg on the testcase
...
in PR1432 by 6%
llvm-svn: 40803
2007-08-04 01:07:49 +00:00
fd838f0770
make RenamePassWorkList a local var instead of an ivar.
...
llvm-svn: 40802
2007-08-04 01:04:40 +00:00
fc7fdef7d2
Use DominatorTree instead of ETForest.
...
This allows faster immediate domiantor walk.
llvm-svn: 37500
2007-06-07 21:57:03 +00:00
d3ccc073a2
Mem2Reg does not need TargetData.
...
llvm-svn: 36444
2007-04-25 18:32:35 +00:00
073be55d8e
Remove unused function argument.
...
llvm-svn: 36441
2007-04-25 17:15:20 +00:00
2965adb849
Fix a comment.
...
llvm-svn: 36299
2007-04-21 07:12:44 +00:00
2da606c757
Move more passes to using ETForest instead of DominatorTree.
...
llvm-svn: 36271
2007-04-20 06:27:13 +00:00
4398e242dd
Reduce malloc/free traffic.
...
llvm-svn: 35370
2007-03-26 23:19:29 +00:00
5f50e61d52
Remove dead comments.
...
llvm-svn: 35053
2007-03-09 23:41:03 +00:00
bda1250624
Avoid recursion. Use iterative algorithm for RenamePass().
...
llvm-svn: 35052
2007-03-09 23:39:14 +00:00
430c9217f0
redesign the primary datastructure used by mem2reg to eliminate an
...
std::map of std::vector's (ouch!). This speeds up mem2reg by 10% on 176.gcc.
llvm-svn: 33974
2007-02-07 01:15:04 +00:00
c85e79f3e0
With the last change, we no longer need both directions of mapping from
...
BBNumbers. Instead of using a bi-directional mapping, just use a single
densemap. This speeds up mem2reg on 176.gcc by 8%, from 1.3489 to
1.2485s.
llvm-svn: 33940
2007-02-05 23:37:20 +00:00
52da61fb5c
Simplify use of DFBlocks, this makes no noticable performance difference,
...
but paves the way to eliminate BBNumbers.
llvm-svn: 33938
2007-02-05 23:31:26 +00:00
bf67b1229b
Switch InsertedPHINodes back to SmallPtrSet now that the SmallPtrSet::erase
...
bug is fixed.
llvm-svn: 33932
2007-02-05 23:11:37 +00:00
606dde0093
switch a SmallPtrSet back to an std::set for now, this caused problems.
...
llvm-svn: 33930
2007-02-05 22:28:52 +00:00
1ed84bbd2d
switch an std::set over to a SmallPtrSet, speeding up mem2reg 6% on 176.gcc.
...
llvm-svn: 33929
2007-02-05 22:15:21 +00:00
70fbb9de4c
switch an std::set over to SmallPtrSet, speeding up mem2reg 3.4% on 176.gcc.
...
llvm-svn: 33928
2007-02-05 22:13:11 +00:00
8fbc888d91
eliminate some malloc traffic, this speeds up mem2reg by 3.4%.
...
llvm-svn: 33927
2007-02-05 21:58:48 +00:00
3d27be1333
s|llvm/Support/Visibility.h|llvm/Support/Compiler.h|
...
llvm-svn: 29911
2006-08-27 12:54:02 +00:00
996795b0dd
Use hidden visibility to make symbols in an anonymous namespace get
...
dropped. This shrinks libllvmgcc.dylib another 67K
llvm-svn: 28975
2006-06-28 23:17:24 +00:00
f98b4aa2e7
Fix some nondeterminstic behavior in the mem2reg pass that (in addition to
...
nondeterminism being bad) could cause some trivial missed optimizations (dead
phi nodes being left around for later passes to clean up).
With this, llvm-gcc4 now bootstraps and correctly compares. I don't know
why I never tried to do it before... :)
llvm-svn: 27984
2006-04-27 01:14:43 +00:00
3e9e8bd25c
Implement a refinement to the mem2reg algorithm for cases where an alloca
...
has a single def. In this case, look for uses that are dominated by the def
and attempt to rewrite them to directly use the stored value.
This speeds up mem2reg on these values and reduces the number of phi nodes
inserted. This should address PR665.
llvm-svn: 24411
2005-11-18 07:31:42 +00:00
31dc3827d3
This needs proper dominance
...
llvm-svn: 24410
2005-11-18 07:29:44 +00:00
257efb2ad3
This code can handle non-dominating instructions
...
llvm-svn: 22667
2005-08-05 00:57:45 +00:00
Chris Lattner