1c577b54b0
fix a bozo bug I introduced in r119930, causing a miscompile of
...
20040709-1.c from the gcc testsuite. I was using the size of a
pointer instead of the pointee. This fixes rdar://8713376
llvm-svn: 120519
2010-12-01 01:24:55 +00:00
903add84d9
Enhance DSE to handle the variable index case in PR8657.
...
llvm-svn: 120498
2010-11-30 23:43:23 +00:00
c0f3379ae0
teach DSE to use GetPointerBaseWithConstantOffset to analyze
...
may-aliasing stores that partially overlap with different base
pointers. This implements PR6043 and the non-variable part of
PR8657
llvm-svn: 120485
2010-11-30 23:05:20 +00:00
e28618de59
move GetPointerBaseWithConstantOffset out of GVN into ValueTracking.h
...
llvm-svn: 120476
2010-11-30 22:25:26 +00:00
50162e3c2a
remove a fixed fixme
...
llvm-svn: 120474
2010-11-30 22:18:11 +00:00
6712251f41
Make DeleteDeadInstruction be a static function, move some code around.
...
llvm-svn: 120471
2010-11-30 21:58:14 +00:00
51d67ce2ff
switch RemoveAccessedObjects to use AliasAnalysis::Location to simplify
...
the code. We now get accurate sizes on Loads, though it surely doesn't
matter in practice.
llvm-svn: 120469
2010-11-30 21:47:58 +00:00
f80b39986f
two improvements to RemoveAccessedObjects:
...
1. if the underlying pointer passed in can be resolved
to any argument or alloca, then we don't need to scan.
Previously we would only avoid the scan if the alloca
or byval was actually considered dead.
2. The dead store processing code is itself completely
dead and didn't handle volatile stores right anyway,
so delete it. This allows simplifying the interface
to RemoveAccessedObjects.
llvm-svn: 120467
2010-11-30 21:38:30 +00:00
7fe08b67fa
remove the "undead" terminology, which is nonstandard and never
...
made sense to me. We now have a set of dead stack objects, and
they become live when loaded. Fix a theoretical problem where
we'd pass in the wrong pointer to the alias query.
llvm-svn: 120465
2010-11-30 21:32:12 +00:00
127818d746
move call handling in handleEndBlock up a bit, and simplify it.
...
If the call might read all the allocas, stop scanning early.
Convert a vector to smallvector, shrink SmallPtrSet to 16 instead
of 64 to avoid crazy linear scans.
llvm-svn: 120463
2010-11-30 21:18:46 +00:00
d3a58c8fa1
Avoid exponential growth of a table. It feels like
...
there should be a better way to do this. PR 8679.
llvm-svn: 120457
2010-11-30 20:23:21 +00:00
60a8b3dab8
various cleanups and code simplification
...
llvm-svn: 120454
2010-11-30 19:48:15 +00:00
51c28a93cc
make getPointerSize a static function. Add ivars to DSE for
...
AA and MD pass info instead of using getAnalysis<> all over.
llvm-svn: 120453
2010-11-30 19:34:42 +00:00
77d79fa25f
reduce indentation, clean up TD use a bit.
...
llvm-svn: 120452
2010-11-30 19:28:23 +00:00
b63ba73b1b
enhance isRemovable to refuse to delete volatile mem transfers
...
now that DSE hacks on them. This fixes a regression I introduced,
by generalizing DSE to hack on transfers.
llvm-svn: 120445
2010-11-30 19:12:10 +00:00
58b779e9c2
Rewrite the main DSE loop to be written in terms of reasoning
...
about pairs of AA::Location's instead of looking for MemDep's
"Def" predicate. This is more powerful and general, handling
memset/memcpy/store all uniformly, and implementing PR8701 and
probably obsoleting parts of memcpyoptimizer.
This also fixes an obscure bug with init.trampoline and i8
stores, but I'm not surprised it hasn't been hit yet. Enhancing
init.trampoline to carry the size that it stores would allow
DSE to be much more aggressive about optimizing them.
llvm-svn: 120406
2010-11-30 07:23:21 +00:00
e3ea1cba79
Add a puts optimization that converts puts() to putchar('\n').
...
llvm-svn: 120398
2010-11-30 06:19:18 +00:00
3590ef817c
rename a function and reduce some indentation, no functionality change.
...
llvm-svn: 120391
2010-11-30 05:30:45 +00:00
b438ef236c
remove the pointless check of MemoryUseIntrinsic from
...
is trivially dead, since these have side effects. This makes the
(misnamed) MemoryUseIntrinsic class dead, so remove it.
llvm-svn: 120382
2010-11-30 02:03:47 +00:00
2227a8a192
rename doesClobberMemory -> hasMemoryWrite to be more specific, and
...
remove an actively-wrong comment.
llvm-svn: 120378
2010-11-30 01:37:52 +00:00
9d179d911d
clean up handling of 'free', detangling it from everything else.
...
It can be seriously improved, but at least now it isn't intertwined
with the other logic.
llvm-svn: 120377
2010-11-30 01:28:33 +00:00
9a146372b5
Teach basicaa that memset's modref set is at worst "mod" and never
...
contains "ref".
Enhance DSE to use a modref query instead of a store-specific hack
to generalize the "ignore may-alias stores" optimization to handle
memset and memcpy.
llvm-svn: 120368
2010-11-30 00:28:45 +00:00
c3c754f750
my previous patch would cause us to start deleting some volatile
...
stores, fix and add a testcase.
llvm-svn: 120363
2010-11-30 00:12:39 +00:00
d4f1090948
two changes to DSE that shouldn't affect anything:
...
1. Don't bother trying to optimize:
lifetime.end(ptr)
store(ptr)
as it is undefined, and therefore shouldn't exist.
2. Move the 'storing a loaded pointer' xform up, simplifying
the may-aliased store code.
llvm-svn: 120359
2010-11-30 00:01:19 +00:00
b4df1d5a3e
prune an llvmcontext include and simplify some code.
...
llvm-svn: 120347
2010-11-29 23:35:33 +00:00
2e8793482c
fix PR8677, patch by Jakub Staszak!
...
llvm-svn: 120325
2010-11-29 21:59:31 +00:00
28218aa8f1
Transform (extractvalue (load P), ...) to (load (gep P, 0, ...)) if the load has no other uses, shrinking the load.
...
llvm-svn: 120323
2010-11-29 21:56:20 +00:00
8ba5f39f70
Second attempt at fixing the performance regressions introduced
...
by my recent GVN improvement. Looking through a single layer of
PHI nodes when attempting to sink GEPs, we need to iteratively
look through arbitrary PHI nests.
llvm-svn: 120202
2010-11-27 08:15:55 +00:00
b8de00ee07
Treat a call of function pointer like a load of the pointer when considering
...
whether the pointer can be replaced with the global variable it is a copy of.
Fixes PR8680.
llvm-svn: 120126
2010-11-24 22:04:20 +00:00
0488d564e1
Rename SimplifyDistributed to the more meaningfull name SimplifyByFactorizing.
...
llvm-svn: 120051
2010-11-23 20:42:39 +00:00
94a622af4c
The srem -> urem transform is not safe for any divisor that's not a power of two.
...
E.g. -5 % 5 is 0 with srem and 1 with urem.
Also addresses Frits van Bommel's comments.
llvm-svn: 120049
2010-11-23 20:33:57 +00:00
433c1679cf
Replace calls to ConstantFoldInstruction with calls to SimplifyInstruction
...
in two places that are really interested in simplified instructions, not
constants.
llvm-svn: 120044
2010-11-23 20:26:33 +00:00
bb2cd025a9
Constant folding here is pointless, because InstructionSimplify
...
(which does constant folding and more) is called a few lines
later.
llvm-svn: 120042
2010-11-23 20:24:21 +00:00
b5afa65b0a
InstCombine: Reduce "X shift (A srem B)" to "X shift (A urem B)" iff B is positive.
...
This allows to transform the rem in "1 << ((int)x % 8);" to an and.
llvm-svn: 120028
2010-11-23 18:52:42 +00:00
60813f96e0
Propagate LeftDistributes and RightDistributes into their only uses.
...
Stylistic improvement suggested by Frits van Bommel.
llvm-svn: 120026
2010-11-23 15:28:14 +00:00
22df741687
Fix typo pointed out by Frits van Bommel and Marius Wachtler.
...
llvm-svn: 120025
2010-11-23 15:25:34 +00:00
adc7771f18
Exploit distributive laws (eg: And distributes over Or, Mul over Add, etc) in a
...
fairly systematic way in instcombine. Some of these cases were already dealt
with, in which case I removed the existing code. The case of Add has a bunch of
funky logic which covers some of this plus a few variants (considers shifts to be
a form of multiplication), which I didn't touch. The simplification performed is:
A*B+A*C -> A*(B+C). The improvement is to do this in cases that were not already
handled [such as A*B-A*C -> A*(B-C), which was reported on the mailing list], and
also to do it more often by not checking for "only one use" if "B+C" simplifies.
llvm-svn: 120024
2010-11-23 14:23:47 +00:00
e5afa15b77
duncan's spider sense was right, I completely reversed the condition
...
on this instcombine xform. This fixes a miscompilation of 403.gcc.
llvm-svn: 119988
2010-11-23 02:42:04 +00:00
f1ebb63161
InstCombine: Implement X - A*-B -> X + A*B.
...
llvm-svn: 119984
2010-11-22 20:31:27 +00:00
c133c54426
If a GEP index simply advances by multiples of a type of zero size,
...
then replace the index with zero.
llvm-svn: 119974
2010-11-22 16:32:50 +00:00
8a0f486e36
Move the "gep undef" -> "undef" transform from instcombine to
...
InstructionSimplify.
llvm-svn: 119970
2010-11-22 13:42:49 +00:00
c6648eb4c3
Don't keep track of inserted phis in PromoteMemoryToRegister: the information
...
is never used. Patch by Cameron Zwarich.
llvm-svn: 119963
2010-11-22 09:41:24 +00:00
fc9aead6fd
fix comment
...
llvm-svn: 119948
2010-11-21 19:05:34 +00:00
5957229659
rework some DSE paths to use the newly-public "getPointerDependencyFrom"
...
method in MemDep instead of inserting an instruction, doing a query,
then removing it. Neither operation is effectively cached.
llvm-svn: 119930
2010-11-21 08:06:10 +00:00
e48c31ce33
implement PR8576, deleting dead stores with intervening may-alias stores.
...
llvm-svn: 119927
2010-11-21 07:34:32 +00:00
f7e896138e
optimize:
...
void a(int x) { if (((1<<x)&8)==0) b(); }
into "x != 3", which occurs over 100 times in 403.gcc but in no
other program in llvm-test.
llvm-svn: 119922
2010-11-21 06:44:42 +00:00
58f9f58716
Implement PR8644: forwarding a memcpy value to a byval,
...
allowing the memcpy to be eliminated.
Unfortunately, the requirements on byval's without explicit
alignment are really weak and impossible to predict in the
mid-level optimizer, so this doesn't kick in much with current
frontends. The fix is to change clang to set alignment on all
byval arguments.
llvm-svn: 119916
2010-11-21 00:28:59 +00:00
ddd1b7b801
Simplify code. No change in functionality.
...
llvm-svn: 119908
2010-11-20 18:43:35 +00:00
ea326db47b
Document the new GVN number table structure.
...
llvm-svn: 119865
2010-11-19 22:48:40 +00:00
dfb8c3bbfc
When folding addressing modes in CodeGenPrepare, attempt to look through PHI nodes
...
if all the operands of the PHI are equivalent. This allows CodeGenPrepare to undo
unprofitable PRE transforms.
llvm-svn: 119853
2010-11-19 22:15:03 +00:00
Chris Lattner