Commit Graph

494 Commits

Author SHA1 Message Date
Dan Gohman d76d71a291 Factor the code for collecting IV users out of LSR into an IVUsers class,
and generalize it so that it can be used by IndVarSimplify. Implement the
base IndVarSimplify transformation code using IVUsers. This removes
TestOrigIVForWrap and associated code, as ScalarEvolution now has enough
builtin overflow detection and folding logic to handle all the same cases,
and more. Run "opt -iv-users -analyze -disable-output" on your favorite
loop for an example of what IVUsers does.

This lets IndVarSimplify eliminate IV casts and compute trip counts in
more cases. Also, this happens to finally fix the remaining testcases
in PR1301.

Now that IndVarSimplify is being more aggressive, it occasionally runs
into the problem where ScalarEvolutionExpander's code for avoiding
duplicate expansions makes it difficult to ensure that all expanded
instructions dominate all the instructions that will use them. As a
temporary measure, IndVarSimplify now uses a FixUsesBeforeDefs function
to fix up instructions inserted by SCEVExpander. Fortunately, this code
is contained, and can be easily removed once a more comprehensive
solution is available.

llvm-svn: 71535
2009-05-12 02:17:14 +00:00
Evan Cheng 78a4eb844b Teach LSR to optimize more loop exit compares, i.e. change them to use postinc iv value. Previously LSR would only optimize those which are in the loop latch block. However, if LSR can prove it is safe (and profitable), it's now possible to change those not in the latch blocks to use postinc values.
Also, if the compare is the only use, LSR would place the iv increment instruction before the compare instead in the latch.

llvm-svn: 71485
2009-05-11 22:33:01 +00:00
Dale Johannesen 02cb2bf2e3 Reverse a loop that is counting up to a maximum to
count down to 0 instead, under very restricted
circumstances.  Adjust 4 testcases in which this
optimization fires.

llvm-svn: 71439
2009-05-11 17:15:42 +00:00
Evan Cheng b9dcc2c0c9 Factor out code that optimize loop terminating condition.
llvm-svn: 71305
2009-05-09 01:08:24 +00:00
Evan Cheng 342053cd27 Unbreak the build.
llvm-svn: 71091
2009-05-06 18:00:56 +00:00
David Greene 0dec5b9a75 Make sure to use signed arithmetic in APInt to fix a regression.
llvm-svn: 71090
2009-05-06 17:39:26 +00:00
Dan Gohman e58fc20f8d Fix a copy+pasto in a comment.
llvm-svn: 71035
2009-05-05 23:02:38 +00:00
Dan Gohman 96b18ccdd3 Delete a FIXME which is no longer relevant, and add a FIXME that is.
llvm-svn: 71033
2009-05-05 22:59:55 +00:00
Bill Wendling 5e2ac0cd9c Temporarily reverting r71008. It was causing this failure:
Running /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm.src/test/
CodeGen/X86/dg.exp ...
FAIL: /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm.src/test/
CodeGen/X86/change-compare-stride-1.ll
Failed with exit(1) at line 2
while running: grep {cmpq       $-478,} change-compare-stride-1.ll.tmp
child process exited abnormally

llvm-svn: 71013
2009-05-05 20:49:46 +00:00
David Greene 246a3dfb10 Handle overflow of 64-bit loop conditions.
llvm-svn: 71008
2009-05-05 20:22:36 +00:00
Dan Gohman 48f8222293 Re-apply 70645, converting ScalarEvolution to use
CallbackVH, with fixes. allUsesReplacedWith need to
walk the def-use chains and invalidate all users of a
value that is replaced. SCEVs of users need to be
recalcualted even if the new value is equivalent. Also,
make forgetLoopPHIs walk def-use chains, since any
SCEV that depends on a PHI should be recalculated when
more information about that PHI becomes available.

llvm-svn: 70927
2009-05-04 22:30:44 +00:00
Dan Gohman a30370bc33 Constify a bunch of SCEV-using code.
llvm-svn: 70919
2009-05-04 22:02:23 +00:00
Dan Gohman 5036695c32 Revert r70645 for now; it's causing a variety of regressions.
llvm-svn: 70661
2009-05-03 05:46:20 +00:00
Dan Gohman e9a38d16fe Convert ScalarEvolution to use CallbackVH for its internal map. This
makes ScalarEvolution::deleteValueFromRecords, and it's code that
subtly needed to be called before ReplaceAllUsesWith, unnecessary.

It also makes ValueDeletionListener unnecessary.

llvm-svn: 70645
2009-05-02 21:19:20 +00:00
Dan Gohman ff08995589 Previously, RecursivelyDeleteDeadInstructions provided an option
of returning a list of pointers to Values that are deleted. This was
unsafe, because the pointers in the list are, by nature of what
RecursivelyDeleteDeadInstructions does, always dangling. Replace this
with a simple callback mechanism. This may eventually be removed if
all clients can reasonably be expected to use CallbackVH.

Use this to factor out the dead-phi-cycle-elimination code from LSR
utility function, and generalize it to use the
RecursivelyDeleteTriviallyDeadInstructions utility function.

This makes LSR more aggressive about eliminating dead PHI cycles;
adjust tests to either be less trivial or to simply expect fewer
instructions.

llvm-svn: 70636
2009-05-02 18:29:22 +00:00
Dan Gohman 6409e7d4e9 Don't split critical edges during the AddUsersIfInteresting phase
of LSR. This makes the AddUsersIfInteresting phase of LSR a pure
analysis instead of a phase that potentially does CFG modifications.

The conditions where this code would actually perform a split are
rare, and in the cases where it actually would do a split the split
is usually undone by CodeGenPrepare, and in cases where splits
actually survive into codegen, they appear to hurt more often than
they help.

llvm-svn: 70625
2009-05-02 05:36:01 +00:00
Dan Gohman 65dbe7874f Make RequiresTypeConversion canonicalize the types before calling the
target hooks canLosslesslyBitCastTo and isTruncateFree. This allows
targets to avoid worrying about handling all combinations of integer
and pointer types.

llvm-svn: 70555
2009-05-01 17:07:43 +00:00
Dan Gohman d3aa4215ef Minor whitespace fix.
llvm-svn: 70551
2009-05-01 16:56:32 +00:00
Dan Gohman 6be8530158 Fix some code to work if TargetLowering is not available.
llvm-svn: 70546
2009-05-01 16:29:14 +00:00
Dale Johannesen f4031bd01e Print correct instruction in dump.
llvm-svn: 70427
2009-04-29 22:57:20 +00:00
Dan Gohman e99f98262c Permit ChangeCompareStride to rewrite a comparison when the factor
between the comparison's iv stride and the candidate stride is
exactly -1.

llvm-svn: 70244
2009-04-27 20:35:32 +00:00
Dan Gohman 4860db61be Factor out a common base class from SCEVTruncateExpr, SCEVZeroExtendExpr,
and SCEVSignExtendExpr.

llvm-svn: 69649
2009-04-21 01:25:57 +00:00
Dan Gohman b397e1a7a2 Introduce encapsulation for ScalarEvolution's TargetData object, and refactor
the code to minimize dependencies on TargetData.

llvm-svn: 69644
2009-04-21 01:07:12 +00:00
Dan Gohman 056857aa21 Use more const qualifiers with SCEV interfaces.
llvm-svn: 69450
2009-04-18 17:56:28 +00:00
Dan Gohman d2d6fd806c Don't create ConstantInts with pointer type. This fixes a
regression in 403.gcc in PIC_CODEGEN=1 and DISABLE_LTO=1
mode.

llvm-svn: 69344
2009-04-17 02:02:52 +00:00
Dan Gohman fec1d086e0 Use TargetData::getTypeSizeInBits instead of getPrimitiveSizeInBits()
to get the correct answer for pointer types.

llvm-svn: 69321
2009-04-16 22:35:57 +00:00
Dan Gohman 8b6ebb1112 Minor code simplifications. Don't attempt LSR on theoretical
targets with pointers larger than 64 bits, due to the code not
yet being APInt clean.

llvm-svn: 69296
2009-04-16 16:49:48 +00:00
Dan Gohman e2ead2c328 LSR is no longer a GEP optimizer. It is now an IV expression
optimizer, which just happen to frequently involve optimizing GEPs.

llvm-svn: 69295
2009-04-16 16:46:01 +00:00
Dan Gohman a8be04b2db Use ConstantExpr::getIntToPtr instead of SCEVExpander::InsertCastOfTo,
since the operand is always a constant.

llvm-svn: 69291
2009-04-16 15:48:38 +00:00
Dan Gohman 71bccd3e0e Use a SCEV expression cast instead of immediately inserting a
new instruction with SCEVExpander::InsertCastOfTo.

llvm-svn: 69290
2009-04-16 15:47:35 +00:00
Dan Gohman 0a40ad93a9 Expand GEPs in ScalarEvolution expressions. SCEV expressions can now
have pointer types, though in contrast to C pointer types, SCEV
addition is never implicitly scaled. This not only eliminates the
need for special code like IndVars' EliminatePointerRecurrence
and LSR's own GEP expansion code, it also does a better job because
it lets the normal optimizations handle pointer expressions just
like integer expressions.

Also, since LLVM IR GEPs can't directly index into multi-dimensional
VLAs, moving the GEP analysis out of client code and into the SCEV
framework makes it easier for clients to handle multi-dimensional
VLAs the same way as other arrays.

Some existing regression tests show improved optimization.
test/CodeGen/ARM/2007-03-13-InstrSched.ll in particular improved to
the point where if-conversion started kicking in; I turned it off
for this test to preserve the intent of the test.

llvm-svn: 69258
2009-04-16 03:18:22 +00:00
Chris Lattner 42e9ca42ce LSR shouldn't ever try to hack on integer IV's larger than 64-bits. Right now
it is not APInt clean, but even when it is it needs to be evaluated carefully
to determine whether it is actually profitable.

This fixes a crash on PR3806

llvm-svn: 67134
2009-03-17 23:58:30 +00:00
Dan Gohman f12436891e Don't record the increment instruction; just recompute it from the Phi
if needed. This simplifies the code a little, and is needed for an
upcoming refactoring.

llvm-svn: 66479
2009-03-09 22:04:01 +00:00
Dan Gohman b855164751 Fix a few more places where induction variable types were used
where memory access types are needed.

llvm-svn: 66470
2009-03-09 21:22:12 +00:00
Dan Gohman 5a4e31666d Use ReplacedTy instead of recomputing the same value.
llvm-svn: 66469
2009-03-09 21:19:58 +00:00
Dan Gohman 34e52ddb7d Use LoopInfo's getLoopLatch() instead of doing what it does manualy.
llvm-svn: 66467
2009-03-09 21:14:16 +00:00
Dan Gohman 70cc9875d8 Don't use an induction variable type as a memory access type.
Use VoidTy instead, to be properly conservative.

llvm-svn: 66463
2009-03-09 21:04:19 +00:00
Dan Gohman 917ffe4592 Factor out the code that determines the memory access type
of an instruction into a helper function.

llvm-svn: 66460
2009-03-09 21:01:17 +00:00
Dan Gohman e201f8ff1d Move the sorting of the StrideOrder array earlier so that it doesn't
have to be done twice.

llvm-svn: 66449
2009-03-09 20:46:50 +00:00
Dan Gohman b5001909b0 Delete the isOnlyStride argument, which is unused.
llvm-svn: 66446
2009-03-09 20:41:15 +00:00
Dan Gohman 85875f7120 Tidy some LSR debug output: announce the loop it's about to process
before it does any processing.

llvm-svn: 66443
2009-03-09 20:34:59 +00:00
Dan Gohman 66476b582d Fix this comment.
llvm-svn: 66065
2009-03-04 20:50:23 +00:00
Dan Gohman ae0035ee15 Add an assertion for a condition that's always true, and not
immediately obvious.

llvm-svn: 66062
2009-03-04 20:49:01 +00:00
Dan Gohman 0bddac16a8 Rename ScalarEvolution's getIterationCount to getBackedgeTakenCount,
to more accurately describe what it does. Expand its doxygen comment
to describe what the backedge-taken count is and how it differs
from the actual iteration count of the loop. Adjust names and
comments in associated code accordingly.

llvm-svn: 65382
2009-02-24 18:55:53 +00:00
Dan Gohman 5d1f458f0f Generalize the ChangeCompareStride code, in preparation for
handling non-constant strides. No functionality change.

llvm-svn: 65363
2009-02-24 01:58:00 +00:00
Dan Gohman f394e58af5 Properly parenthesize this expression, fixing a real bug in the new
-full-lsr code, as well as a GCC warning.

llvm-svn: 65288
2009-02-22 16:40:52 +00:00
Evan Cheng 69decbf0b2 Only try to sink immediate when TLI is not null. It needs to check if immediate would fit in target addressing field.
llvm-svn: 65268
2009-02-22 07:31:19 +00:00
Evan Cheng 107b06c4b9 Teach LSR sink to sink the immediate portion of the common expression back into uses if they fit in address modes of all the uses.
llvm-svn: 65215
2009-02-21 02:06:47 +00:00
Evan Cheng 8a9481d50d Fix strange logic in CollectIVUsers used to determine whether all uses are
addresses, part 1. This fixes an obvious logic bug. Previously if the only
in-loop use is a PHI, it would return AllUsesAreAddresses as true.

llvm-svn: 65178
2009-02-20 22:16:49 +00:00
Dan Gohman 5e309a5bbb Simplify code and reduce indentation. No functionality change.
llvm-svn: 65167
2009-02-20 21:27:23 +00:00
Dan Gohman 2c8cb5b4ec Fix 80-column violations.
llvm-svn: 65159
2009-02-20 21:06:57 +00:00
Dan Gohman addc50b4ee It's not necessary to check if Base is null here.
llvm-svn: 65157
2009-02-20 21:05:23 +00:00
Dan Gohman 1608df5319 Add a comment about how Imm can be used for loop-variant values.
llvm-svn: 65147
2009-02-20 20:29:04 +00:00
Dan Gohman 2a12ae7d1f Implement "superhero" strength reduction, or full strength
reduction of address calculations down to basic pointer arithmetic.
This is currently off by default, as it needs a few other features
before it becomes generally useful. And even when enabled, full
strength reduction is only performed when it doesn't increase
register pressure, and when several other conditions are true.

This also factors out a bunch of exisiting LSR code out of
StrengthReduceStridedIVUsers into separate functions, and tidies
up IV insertion. This actually decreases register pressure even
in non-superhero mode. The change in iv-users-in-other-loops.ll
is an example of this; there are two more adds because there are
two fewer leas, and there is less spilling.

llvm-svn: 65108
2009-02-20 04:17:46 +00:00
Dan Gohman a34d7adefb Use DEBUG() instead of passing *DOUT to WriteAsOperand,
since the latter just passes a null reference when
debugging is not enabled.

llvm-svn: 65060
2009-02-19 19:32:06 +00:00
Dan Gohman 30a2959367 Make the debug output of LSR less cryptic and more informative.
llvm-svn: 65057
2009-02-19 19:23:27 +00:00
Dan Gohman d0b1fbd983 Fix a typo in a comment.
llvm-svn: 64859
2009-02-18 00:08:39 +00:00
Evan Cheng 161861deb0 Strengthen the "non-constant stride must dominate loop preheader" check.
llvm-svn: 64703
2009-02-17 00:13:06 +00:00
Evan Cheng e79841adbb Fix pr3571: If stride is a value defined by an instruction, make sure it dominates the loop preheader. When IV users are strength reduced, the stride is inserted into the preheader. It could create a use before def situation.
llvm-svn: 64579
2009-02-15 06:06:15 +00:00
Evan Cheng fe151ba135 ifdef out unneeded if statement.
llvm-svn: 64575
2009-02-15 03:20:37 +00:00
Dan Gohman a2730abaaa Complete the sentance in this comment. I have reservations
about the code it describes, but at least now the comment
is right.

llvm-svn: 64465
2009-02-13 17:36:42 +00:00
Dan Gohman f71a473720 Fix the code that checked if a SCEVAddRecExpr Start contains an
addrec in a different loop to check the value being added to
the accumulated Start value, not the Start value before it has
the new value added to it. This prevents LSR from going crazy
on the included testcase. Dale, please review.

llvm-svn: 64440
2009-02-13 03:58:31 +00:00
Dan Gohman ba83228cdb Fix LSR's IV sorting function to explicitly sort by bitwidth
after sorting by stride value. This prevents it from missing
IV reuse opportunities in a host-sensitive manner.

llvm-svn: 64415
2009-02-13 00:26:43 +00:00
Dale Johannesen cd19967754 Fix PR 3471, and some cleanups.
llvm-svn: 64177
2009-02-09 22:14:15 +00:00
Dale Johannesen 1f0e0e7c9c Fix the time regression I introduced in 464.h264ref with
my earlier patch to this file.

The issue there was that all uses of an IV inside a loop
are actually references to Base[IV*2], and there was one
use outside that was the same but LSR didn't see the base
or the scaling because it didn't recurse into uses outside
the loop; thus, it used base+IV*scale mode inside the loop
instead of pulling base out of the loop.  This was extra bad
because register pressure later forced both base and IV into
memory.  Doing that recursion, at least enough
to figure out addressing modes, is a good idea in general;
the change in AddUsersIfInteresting does this.  However,
there were side effects....

It is also possible for recursing outside the loop to
introduce another IV where there was only 1 before (if
the refs inside are not scaled and the ref outside is).
I don't think this is a common case, but it's in the testsuite.
It is right to be very aggressive about getting rid of
such introduced IVs (CheckForIVReuse and the handling of
nonzero RewriteFactor in StrengthReduceStridedIVUsers).
In the testcase in question the new IV produced this way
has both a nonconstant stride and a nonzero base, neither
of which was handled before.  And when inserting 
new code that feeds into a PHI, it's right to put such 
code at the original location rather than in the PHI's 
immediate predecessor(s) when the original location is outside 
the loop (a case that couldn't happen before)
(RewriteInstructionToUseNewBase); better to avoid making
multiple copies of it in this case.

Also, the mechanism for keeping SCEV's corresponding to GEP's
no longer works, as the GEP might change after its SCEV
is remembered, invalidating the SCEV, and we might get a bad
SCEV value when looking up the GEP again for a later loop.  
This also couldn't happen before, as we weren't recursing
into GEP's outside the loop.

Also, when we build an expression that involves a (possibly
non-affine) IV from a different loop as well as an IV from
the one we're interested in (containsAddRecFromDifferentLoop),
don't recurse into that.  We can't do much with it and will
get in trouble if we try to create new non-affine IVs or something.

More testcases are coming.

llvm-svn: 62212
2009-01-14 02:35:31 +00:00
Duncan Sands dc020f9c3c Rename getABITypeSize to getTypePaddedSize, as
suggested by Chris.

llvm-svn: 62099
2009-01-12 20:38:59 +00:00
Dale Johannesen 656237beca Revert 61362 and 61402 until SPEC breakage is fixed.
llvm-svn: 61403
2008-12-23 23:21:35 +00:00
Dale Johannesen f8b161bcd1 This fixes the bug in 175.vpr. It doesn't fix the
other SPEC breakage.  I'll be reverting all recent
changes shortly, this checking is mostly so this
change doesn't get lost.

llvm-svn: 61402
2008-12-23 23:05:26 +00:00
Dale Johannesen 93b9aa8799 Fix the time regression I introduced in 464.h264ref with
my last patch to this file.

The issue there was that all uses of an IV inside a loop
are actually references to Base[IV*2], and there was one
use outside that was the same but LSR didn't see the base
or the scaling because it didn't recurse into uses outside
the loop; thus, it used base+IV*scale mode inside the loop
instead of pulling base out of the loop.  This was extra bad
because register pressure later forced both base and IV into
memory.  Doing that recursion, at least enough
to figure out addressing modes, is a good idea in general;
the change in AddUsersIfInteresting does this.  However,
there were side effects....

It is also possible for recursing outside the loop to
introduce another IV where there was only 1 before (if
the refs inside are not scaled and the ref outside is).
I don't think this is a common case, but it's in the testsuite.
It is right to be very aggressive about getting rid of
such introduced IVs (CheckForIVReuse and the handling of
nonzero RewriteFactor in StrengthReduceStridedIVUsers).
In the testcase in question the new IV produced this way
has both a nonconstant stride and a nonzero base, neither
of which was handled before.  And when inserting 
new code that feeds into a PHI, it's right to put such 
code at the original location rather than in the PHI's 
immediate predecessor(s) when the original location is outside 
the loop (a case that couldn't happen before)
(RewriteInstructionToUseNewBase); better to avoid making
multiple copies of it in this case.

Also, the mechanism for keeping SCEV's corresponding to GEP's
no longer works, as the GEP might change after its SCEV
is remembered, invalidating the SCEV, and we might get a bad
SCEV value when looking up the GEP again for a later loop.  
This also couldn't happen before, as we weren't recursing
into GEP's outside the loop.

I owe some testcases for this, want to get it in for nightly runs.

llvm-svn: 61362
2008-12-23 02:12:52 +00:00
Dale Johannesen 3e5843b992 Revert previous patch, appears to break bootstrap.
llvm-svn: 61181
2008-12-18 01:23:41 +00:00
Dale Johannesen 12d031b716 Fix the time regression I introduced in 464.h264ref with
my last patch to this file.

The issue there was that all uses of an IV inside a loop
are actually references to Base[IV*2], and there was one
use outside that was the same but LSR didn't see the base
or the scaling because it didn't recurse into uses outside
the loop; thus, it used base+IV*scale mode inside the loop
instead of pulling base out of the loop.  This was extra bad
because register pressure later forced both base and IV into
memory.  Doing that recursion, at least enough
to figure out addressing modes, is a good idea in general;
the change in AddUsersIfInteresting does this.  However,
there were side effects....

It is also possible for recursing outside the loop to
introduce another IV where there was only 1 before (if
the refs inside are not scaled and the ref outside is).
I don't think this is a common case, but it's in the testsuite.
It is right to be very aggressive about getting rid of
such introduced IVs (CheckForIVReuse and the handling of
nonzero RewriteFactor in StrengthReduceStridedIVUsers).
In the testcase in question the new IV produced this way
has both a nonconstant stride and a nonzero base, neither
of which was handled before.  (This patch does not handle 
all the cases where this can happen.)  And when inserting 
new code that feeds into a PHI, it's right to put such 
code at the original location rather than in the PHI's 
immediate predecessor(s) when the original location is outside 
the loop (a case that couldn't happen before)
(RewriteInstructionToUseNewBase); better to avoid making
multiple copies of it in this case.

Everything above is exercised in
CodeGen/X86/lsr-negative-stride.ll (and ifcvt4 in ARM which is
the same IR).

llvm-svn: 61178
2008-12-18 00:57:22 +00:00
Dale Johannesen 904ce8120d Clarify that the scale factor from CheckForIVReuse
can be negative.  Keep track of whether all uses of
an IV are outside the loop.  Some cosmetics; no
functional change.

llvm-svn: 61109
2008-12-16 22:16:28 +00:00
Chris Lattner 56b20ffc5f Fix a really subtle off-by-one bug that Duncan noticed with valgrind
on test/CodeGen/Generic/2007-06-06-CriticalEdgeLandingPad.

llvm-svn: 60739
2008-12-09 04:47:21 +00:00
Dale Johannesen 9efd2ce55b Make LoopStrengthReduce smarter about hoisting things out of
loops when they can be subsumed into addressing modes.

Change X86 addressing mode check to realize that
some PIC references need an extra register.
(I believe this is correct for Linux, if not, I'm sure
someone will tell me.)

llvm-svn: 60608
2008-12-05 21:47:27 +00:00
Dale Johannesen 4e9e6ea604 Remove an unused field.
llvm-svn: 60508
2008-12-03 22:43:56 +00:00
Dale Johannesen f7a588b909 Fix a misspelled function name.
llvm-svn: 60506
2008-12-03 20:56:12 +00:00
Dale Johannesen d49ceff6ba Fix a really wrong comment.
llvm-svn: 60494
2008-12-03 19:25:46 +00:00
Dale Johannesen 4d2ecb8f68 Minor rewrite per review feedback.
llvm-svn: 60442
2008-12-02 21:17:11 +00:00
Dale Johannesen 70060013d2 Make the code do what the comment says it does.
llvm-svn: 60431
2008-12-02 18:40:09 +00:00
Chris Lattner ead1a61b47 some random comment improvements.
llvm-svn: 60395
2008-12-02 04:52:26 +00:00
Dale Johannesen 069a4eee55 Consider only references to an IV within the loop when
figuring out the base of the IV.  This produces better
code in the example.  (Addresses use (IV) instead of 
(BASE,IV) - a significant improvement on low-register
machines like x86).

llvm-svn: 60374
2008-12-01 22:00:01 +00:00
Chris Lattner 2c2dd15a85 Introduce a new array_pod_sort function and switch LSR to use it
instead of std::sort.  This shrinks the release-asserts LSR.o file
by 1100 bytes of code on my system.

We should start using array_pod_sort where possible.

llvm-svn: 60335
2008-12-01 06:49:59 +00:00
Chris Lattner 2aebea5735 Eliminate use of setvector for the DeadInsts set, just use a smallvector.
This is a lot cheaper and conceptually simpler.

llvm-svn: 60332
2008-12-01 06:27:41 +00:00
Chris Lattner 4da78e3774 DeleteTriviallyDeadInstructions is always passed the
DeadInsts ivar, just use it directly.

llvm-svn: 60330
2008-12-01 06:14:28 +00:00
Chris Lattner a68a5a4784 simplify DeleteTriviallyDeadInstructions again, unlike my previous
buggy rewrite, this notifies ScalarEvolution of a pending instruction
about to be removed and then erases it, instead of erasing it then 
notifying.

llvm-svn: 60329
2008-12-01 06:11:32 +00:00
Bill Wendling 469e3aa696 Temporarily revert r60195. It's causing an optimized bootstrap of llvm-gcc to fail.
llvm-svn: 60233
2008-11-29 03:43:04 +00:00
Chris Lattner c077a2a535 Simplify LoopStrengthReduce::DeleteTriviallyDeadInstructions by
making it use RecursivelyDeleteTriviallyDeadInstructions to do
the heavy lifting.

llvm-svn: 60195
2008-11-27 23:23:35 +00:00
Chris Lattner 96e2dbe008 use continue to reduce indentation
llvm-svn: 60192
2008-11-27 23:00:20 +00:00
Daniel Dunbar 7f39e2d85a Change create*Pass factory functions to return Pass* instead of
LoopPass*.
 - Although less precise, this means they can be used in clients
   without RTTI (who would otherwise need to include LoopPass.h, which
   eventually includes things using dynamic_cast). This was the
   simplest solution that presented itself, but I am happy to use a
   better one if available.

llvm-svn: 58010
2008-10-22 23:32:42 +00:00
Dan Gohman 67d90de2b0 Call ScalarEvolution's deleteValueFromRecords before deleting an
instruction, not after. This fixes some uses of free'd memory.

llvm-svn: 56908
2008-10-01 02:02:03 +00:00
Dan Gohman 68e7735a38 Teach LSR to optimize away SMAX operations for tripcounts in common
cases.  See the comment above OptimizeSMax for the full story, and
the testcase for an example. This cancels out a pessimization
commonly attributed to indvars, and will allow us to lift some of
the artificial throttles in indvars, rather than add new ones.

llvm-svn: 56230
2008-09-15 21:22:06 +00:00
Devang Patel 92c5367705 fix overflow check.
llvm-svn: 56011
2008-09-09 20:54:34 +00:00
Devang Patel 7518f250b9 Remove unused counter.
llvm-svn: 55924
2008-09-08 17:14:54 +00:00
Devang Patel 538a7f479a Remove OptimizeIVType()
llvm-svn: 55913
2008-09-08 16:13:27 +00:00
Dan Gohman a79db30d28 Tidy up several unbeseeming casts from pointer to intptr_t.
llvm-svn: 55779
2008-09-04 17:05:41 +00:00
Devang Patel bcd39345de Add additional check to ensure that iv is canonicalized.
llvm-svn: 55682
2008-09-03 00:29:13 +00:00
Devang Patel b530f08122 Check iteration count.
llvm-svn: 55680
2008-09-03 00:10:56 +00:00
Devang Patel 81fed043c5 While removing PHI, use basicblock to identify incoming value.
llvm-svn: 55678
2008-09-03 00:02:42 +00:00
Devang Patel 43c5a52e07 If all IV uses are extending integer IV then change the type of IV itself, if possible.
llvm-svn: 55674
2008-09-02 22:18:08 +00:00
Devang Patel d6adbb6a0f Do not apply the transformation if the target does not support DestTy natively.
llvm-svn: 55433
2008-08-27 20:55:23 +00:00
Devang Patel cf7ca5d0ba Fix typos and whitespaces. Other cosmetic changes based on feedback.
llvm-svn: 55424
2008-08-27 17:50:18 +00:00
Devang Patel 4310d39844 If IV is used in a int-to-float cast inside the loop then try to eliminate the cast operation.
llvm-svn: 55374
2008-08-26 17:57:54 +00:00
Evan Cheng 5dabe042a6 Revert 54821. It's miscompiling 252.eon and 447.dealII
llvm-svn: 54878
2008-08-17 08:07:31 +00:00
Devang Patel f2a03d5a4b Reapply 54786. Add overflow and number of mantissa bits checks.
llvm-svn: 54821
2008-08-15 21:21:34 +00:00
Evan Cheng 86834d29f3 Revert 54786. It's not checking for overflows, etc.
llvm-svn: 54813
2008-08-15 08:12:11 +00:00
Devang Patel 054a833dd4 If IV is used in a int-to-float cast inside the loop then try to eliminate the cast opeation.
llvm-svn: 54786
2008-08-14 20:58:31 +00:00
Devang Patel 6369a798ba Rename. s/FindIVForUser/FindIVUserForCond/g
llvm-svn: 54754
2008-08-13 20:31:11 +00:00
Devang Patel 97387e6615 Check sign to detect overflow before changing compare stride.
llvm-svn: 54710
2008-08-13 02:05:14 +00:00
Evan Cheng 907dc2bc37 Fix PR2355: bug in ChangeCompareStride. When the loop termination compare is the only use of its iv stride, the stride can be eliminated by moving it to another stride. If the scale is negative, swap the predicate instead of using a inverse predicate.
llvm-svn: 54415
2008-08-06 18:04:43 +00:00
Dan Gohman 7ad3cd8c9d Fix a bug in LSR's dead-PHI cleanup. If a PHI has a def-use chain that
leads into a cycle involving a different PHI, LSR got stuck running
around that cycle looking for the original PHI. To avoid this, keep
track of visited PHIs and stop searching if we see one more than once.
This fixes PR2570.

llvm-svn: 53879
2008-07-21 21:45:02 +00:00
Dan Gohman 162668fa78 Fix uninitialized use of the Changed variable.
llvm-svn: 53564
2008-07-14 17:55:01 +00:00
Evan Cheng 03001cb820 Fix two serious LSR bugs.
1. LSR runOnLoop is always returning false regardless if any transformation is made.
2. AddUsersIfInteresting can create new instructions that are added to DeadInsts. But there is a later early exit which prevents them from being freed.

llvm-svn: 53193
2008-07-07 19:51:32 +00:00
Dan Gohman ac563833ae Fix spelling and grammar in a comment.
llvm-svn: 52648
2008-06-23 22:11:52 +00:00
Dan Gohman 5ca5e02480 Improve LSR's dead-phi detection to handle use-def cycles
with more than two nodes.

llvm-svn: 52617
2008-06-22 20:44:02 +00:00
Dan Gohman be928e3b21 Move LSR's private isZero function to a public SCEV member
function, and make use of it in several places.

llvm-svn: 52463
2008-06-18 16:23:07 +00:00
Dan Gohman ab0dccba6b Refine the change in r52258 for avoiding use-before-def conditions
when changing the stride of a comparison so that it's slightly
more precise, by having it scan the instruction list to determine
if there is a use of the condition after the point where the
condition will be inserted.

llvm-svn: 52371
2008-06-16 22:34:15 +00:00
Evan Cheng 319e9a4f63 Switch over to SetVector to ensure same order of iterations do not vary across runs.
llvm-svn: 52361
2008-06-16 21:08:17 +00:00
Evan Cheng a72cdcd1a2 Iterating over SmallPtrSet is not deterministic.
llvm-svn: 52339
2008-06-16 18:17:09 +00:00
Dan Gohman 9ad8c54aab Protect ChangeCompareStride from situations in which it is possible
for it to generate use-before-def IR, such as in this testcase.

llvm-svn: 52258
2008-06-13 21:43:41 +00:00
Gabor Greif 0babc61631 op_iterator-ify some loops, fix 80col violations
llvm-svn: 52226
2008-06-11 21:38:51 +00:00
Evan Cheng 02912418f1 Remove x86.sse2.loadh.pd and x86.sse2.loadl.pd. These will be lowered into load and shuffle instructions.
llvm-svn: 51521
2008-05-24 00:07:06 +00:00
Dan Gohman e62632e0bb When LSR is replacing an instruction, call
ScalarEvolution::deleteValueFromRecords on it before doing the
replaceAllUsesWith, because ScalarEvolution looks at the instruction's
users to find SCEV references to the instruction's SCEV object in its
internal maps.

Move all of LSR's loop-related state clearing after processing the loop
and before cleaning up dead PHI nodes. This eliminates all of LSR's SCEV
references just before the calls to ScalarEvolution::deleteValueFromRecords
so that when ScalarEvolution drops its own SCEV references, the reference
counts will reach zero and the SCEVs will be deleted immediately.

These changes fix some compiler aborts involving ScalarEvolution holding
onto and reusing SCEV objects for instructions that have been deleted.
No regression test unfortunately; because the symptoms were due to
dangling pointers, reduced testcases ended up being fairly arbitrary.

llvm-svn: 51359
2008-05-21 00:54:12 +00:00
Dan Gohman e5572706e8 Refine the fix in r51169 to only apply when the operand val being
replaced is a PHI. This prevents it from inserting uses before defs
in the case that it isn't a PHI and it depends on other instructions
later in the block. This fixes the 447.dealII regression on x86-64.

llvm-svn: 51292
2008-05-20 03:01:48 +00:00
Dan Gohman 0a0fa7cf78 Fix a bug in LoopStrengthReduce that caused it to emit IR with
use-before-def. The problem comes up in code with multiple PHIs where
one PHI is being rewritten in terms of the other, but the other needs
to be casted first. LLVM rules requre the cast instruction to be
inserted after any PHI instructions, but when instructions were
inserted to replace the second PHI value with a function of the first,
they were ended up going before the cast instruction. Avoid this
problem by remembering the location of the cast instruction, when one
is needed, and inserting the expansion of the new value after it.

This fixes a bug that surfaced in 255.vortex on x86-64 when
instcombine was removed from the middle of the loop optimization
passes. 

llvm-svn: 51169
2008-05-15 23:26:57 +00:00
Dan Gohman d78c400b5b Clean up the use of static and anonymous namespaces. This turned up
several things that were neither in an anonymous namespace nor static
but not intended to be global.

llvm-svn: 51017
2008-05-13 00:00:25 +00:00
Dan Gohman e36714c0b4 Minor whitespace and comment cleanups.
llvm-svn: 49671
2008-04-14 18:26:16 +00:00
Gabor Greif 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
Evan Cheng a90fdc4340 Remove dead options.
llvm-svn: 48556
2008-03-19 22:02:26 +00:00
Dan Gohman 70de4cb1cd Use empty() instead of comparing size() with zero.
llvm-svn: 46514
2008-01-29 13:02:09 +00:00
Chris Lattner f3ebc3f3d2 Remove attribution from file headers, per discussion on llvmdev.
llvm-svn: 45418
2007-12-29 20:36:04 +00:00
Evan Cheng 26ee54eb05 Clean up previous patch: PHI uses should not prevent iv reuse if all other uses are addresses. This trades a constant multiply for one fewer iv.
llvm-svn: 45251
2007-12-20 02:20:53 +00:00
Evan Cheng e2a8ba7fec Allow iv reuse if the user is a PHI node which is in turn used as addresses.
llvm-svn: 45230
2007-12-19 23:33:23 +00:00
Dale Johannesen 7d97662467 Remove indeterminism from a loop. We think this will
fix an occasional nonrepeatable bootstrap failure we've
been seeing on Darwin.

llvm-svn: 44202
2007-11-17 02:48:01 +00:00
Evan Cheng 240c1adade At end of LSR, replace uses of now constant (as result of SplitCriticalEdge) PHI node with the constant value.
llvm-svn: 43533
2007-10-30 23:45:15 +00:00
Evan Cheng c2dbfee43f It's not safe to tell SplitCriticalEdge to merge identical edges. It may delete the phi instruction that's being processed.
llvm-svn: 43524
2007-10-30 22:27:26 +00:00
Evan Cheng b024c4c81d - Bug fixes.
- Allow icmp rewrite using an iv / stride of a smaller integer type.

llvm-svn: 43480
2007-10-29 22:07:18 +00:00
Dan Gohman 7414e21ec0 Update a comment to reflect the current code.
llvm-svn: 43463
2007-10-29 19:32:39 +00:00
Dan Gohman f5feb01056 Remove an unused function argument.
llvm-svn: 43462
2007-10-29 19:31:25 +00:00
Dan Gohman 50d42224d0 Fix a typo in a comment.
llvm-svn: 43461
2007-10-29 19:26:14 +00:00
Dan Gohman 8e8adada83 Avoid calling ValidStride when not all uses are addresses.
llvm-svn: 43460
2007-10-29 19:23:53 +00:00
Evan Cheng 9dbe99dcd6 A number of LSR fixes:
- ChangeCompareStride only reuse stride that is larger than current stride. It
  will let the general reuse mechanism to try to reuse a smaller stride.
- Watch out for multiplication overflow in ChangeCompareStride.
- Replace std::set with SmallPtrSet.

llvm-svn: 43408
2007-10-26 23:08:19 +00:00
Evan Cheng d78a3e5555 Fix a crash. Make sure TLI is not null.
llvm-svn: 43384
2007-10-26 17:24:46 +00:00
Evan Cheng 7f3d02471d Loosen up iv reuse to allow reuse of the same stride but a larger type when truncating from the larger type to smaller type is free.
e.g.
Turns this loop:
LBB1_1: # entry.bb_crit_edge
        xorl    %ecx, %ecx
        xorw    %dx, %dx
        movw    %dx, %si
LBB1_2: # bb
        movl    L_X$non_lazy_ptr, %edi
        movw    %si, (%edi)
        movl    L_Y$non_lazy_ptr, %edi
        movw    %dx, (%edi)
		addw    $4, %dx
		incw    %si
		incl    %ecx
		cmpl    %eax, %ecx
		jne     LBB1_2  # bb
	
into

LBB1_1: # entry.bb_crit_edge
        xorl    %ecx, %ecx
        xorw    %dx, %dx
LBB1_2: # bb
        movl    L_X$non_lazy_ptr, %esi
        movw    %cx, (%esi)
        movl    L_Y$non_lazy_ptr, %esi
        movw    %dx, (%esi)
        addw    $4, %dx
		incl    %ecx
        cmpl    %eax, %ecx
        jne     LBB1_2  # bb

llvm-svn: 43375
2007-10-26 01:56:11 +00:00
Evan Cheng 29e29e63bd Do not rewrite compare instruction using iv of a different stride if the new
stride may be rewritten using the stride of the compare instruction.

llvm-svn: 43367
2007-10-25 22:45:20 +00:00
Evan Cheng 5a38108374 Remove code that's commented out.
llvm-svn: 43356
2007-10-25 18:38:24 +00:00
Evan Cheng 133694db06 If a loop termination compare instruction is the only use of its stride,
and the compaison is against a constant value, try eliminate the stride
by moving the compare instruction to another stride and change its
constant operand accordingly. e.g.

loop:
...
v1 = v1 + 3
v2 = v2 + 1
if (v2 < 10) goto loop
=>
loop:
...
v1 = v1 + 3
if (v1 < 30) goto loop

llvm-svn: 43336
2007-10-25 09:11:16 +00:00
Dan Gohman e0c3d9f338 Strength reduction improvements.
- Avoid attempting stride-reuse in the case that there are users that
   aren't addresses. In that case, there will be places where the
   multiplications won't be folded away, so it's better to try to
   strength-reduce them.

 - Several SSE intrinsics have operands that strength-reduction can
   treat as addresses. The previous item makes this more visible, as
   any non-address use of an IV can inhibit stride-reuse.

 - Make ValidStride aware of whether there's likely to be a base
   register in the address computation. This prevents it from thinking
   that things like stride 9 are valid on x86 when the base register is
   already occupied.

Also, XFAIL the 2007-08-10-LEA16Use32.ll test; the new logic to avoid
stride-reuse elimintes the LEA in the loop, so the test is no longer
testing what it was intended to test.

llvm-svn: 43231
2007-10-22 20:40:42 +00:00
Dan Gohman a37eaf2bf9 Move the SCEV object factors from being static members of the individual
SCEV subclasses to being non-static member functions of the ScalarEvolution
class.

llvm-svn: 43224
2007-10-22 18:31:58 +00:00
Dale Johannesen b6c05b1f90 Fix stride computations for long double arrays.
llvm-svn: 42508
2007-10-01 23:08:35 +00:00
Chris Lattner 2740694450 wrap some long lines. Major offenders that are left include
gvn, gvnpre, dse, and predsimplify.  To see these, use:

  make check-line-length

llvm-svn: 40738
2007-08-02 16:53:43 +00:00
Dan Gohman 34d442f274 More explicit keywords.
llvm-svn: 40673
2007-08-01 15:32:29 +00:00
Dan Gohman 8c4da37b1f Use SCEVExpander::InsertCastOfTo instead of calling new IntToPtrInst
directly, because the insert point used by the SCEVExpander may vary
from what LSR originally computes.

llvm-svn: 40641
2007-07-31 17:22:27 +00:00
Dan Gohman 32f53bbd85 Rename ScalarEvolution::deleteInstructionFromRecords to
deleteValueFromRecords and loosen the types to all it to accept
Value* instead of just Instruction*, since this is what
ScalarEvolution uses internally anyway. This allows more flexibility
for future uses.

llvm-svn: 37657
2007-06-19 14:28:31 +00:00
Dan Gohman cb9e09ad57 Add a SCEV class and supporting code for sign-extend expressions.
This created an ambiguity for expandInTy to decide when to use
sign-extension or zero-extension, but it turns out that most of its callers
don't actually need a type conversion, now that LLVM types don't have
explicit signedness. Drop expandInTy in favor of plain expand, and change
the few places that actually need a type conversion to do it themselves.

llvm-svn: 37591
2007-06-15 14:38:12 +00:00
Devang Patel df6355ccf8 Use DominatorTree instead of ETForest.
llvm-svn: 37499
2007-06-07 21:42:15 +00:00
Chris Lattner 1b7b6e76ec Fix PR1495 and CodeGen/X86/2007-06-05-LSR-Dominator.ll
llvm-svn: 37454
2007-06-06 01:23:55 +00:00
Chris Lattner e8bd53c36a Handle negative strides much more optimally. This compiles X86/lsr-negative-stride.ll
into:

_t:
        movl 8(%esp), %ecx
        movl 4(%esp), %eax
        cmpl %ecx, %eax
        je LBB1_3       #bb17
LBB1_1: #bb
        cmpl %ecx, %eax
        jg LBB1_4       #cond_true
LBB1_2: #cond_false
        subl %eax, %ecx
        cmpl %ecx, %eax
        jne LBB1_1      #bb
LBB1_3: #bb17
        ret
LBB1_4: #cond_true
        subl %ecx, %eax
        cmpl %ecx, %eax
        jne LBB1_1      #bb
        jmp LBB1_3      #bb17

instead of:

_t:
        subl $4, %esp
        movl %esi, (%esp)
        movl 12(%esp), %ecx
        movl 8(%esp), %eax
        cmpl %ecx, %eax
        je LBB1_4       #bb17
LBB1_1: #bb.outer
        movl %ecx, %edx
        negl %edx
LBB1_2: #bb
        cmpl %ecx, %eax
        jle LBB1_5      #cond_false
LBB1_3: #cond_true
        addl %edx, %eax
        cmpl %ecx, %eax
        jne LBB1_2      #bb
LBB1_4: #bb17
        movl (%esp), %esi
        addl $4, %esp
        ret
LBB1_5: #cond_false
        movl %ecx, %edx
        subl %eax, %edx
        movl %eax, %esi
        addl %esi, %esi
        cmpl %ecx, %esi
        je LBB1_4       #bb17
LBB1_6: #cond_false.bb.outer_crit_edge
        movl %edx, %ecx
        jmp LBB1_1      #bb.outer

llvm-svn: 37252
2007-05-19 01:22:21 +00:00
Chris Lattner 1480e16596 significantly improve debug output of lsr
llvm-svn: 36996
2007-05-11 22:40:34 +00:00
Dan Gohman 2bcbd5b7ca Use IntrinsicInst to test for prefetch instructions, which is ever so
slightly nicer than using CallInst with an extra check; thanks Chris.

llvm-svn: 36743
2007-05-04 14:59:09 +00:00
Dan Gohman 3fbb18d1b6 Allow strength reduction to make use of addressing modes for the
address operand in a prefetch intrinsic.

llvm-svn: 36713
2007-05-03 23:20:33 +00:00
Devang Patel 8c78a0bff0 Drop 'const'
llvm-svn: 36662
2007-05-03 01:11:54 +00:00
Devang Patel e95c6ad802 Use 'static const char' instead of 'static const int'.
Due to darwin gcc bug, one version of darwin linker coalesces
static const int, which defauts PassID based pass identification.

llvm-svn: 36652
2007-05-02 21:39:20 +00:00
Devang Patel 09f162ca6a Do not use typeinfo to identify pass in pass manager.
llvm-svn: 36632
2007-05-01 21:15:47 +00:00
Devang Patel 38bc86f057 Fix
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070423/048333.html

llvm-svn: 36380
2007-04-23 22:42:03 +00:00
Owen Anderson f35a1dbc7a Remove ImmediateDominator analysis. The same information can be obtained from DomTree. A lot of code for
constructing ImmediateDominator is now folded into DomTree construction.

This is part of the ongoing work for PR217.

llvm-svn: 36063
2007-04-15 08:47:27 +00:00
Chris Lattner efd3051d60 Now that codegen prepare isn't defeating me, I can finally fix what I set
out to do! :)

This fixes a problem where LSR would insert a bunch of code into each MBB
that uses a particular subexpression (e.g. IV+base+C).  The problem is that
this code cannot be CSE'd back together if inserted into different blocks.

This patch changes LSR to attempt to insert a single copy of this code and
share it, allowing codegenprepare to duplicate the code if it can be sunk
into various addressing modes.  On CodeGen/ARM/lsr-code-insertion.ll,
for example, this gives us code like:

        add r8, r0, r5
        str r6, [r8, #+4]
..
        ble LBB1_4      @cond_next
LBB1_3: @cond_true
        str r10, [r8, #+4]
LBB1_4: @cond_next
...
LBB1_5: @cond_true55
        ldr r6, LCPI1_1
        str r6, [r8, #+4]

instead of:

        add r10, r0, r6
        str r8, [r10, #+4]
...
        ble LBB1_4      @cond_next
LBB1_3: @cond_true
        add r8, r0, r6
        str r10, [r8, #+4]
LBB1_4: @cond_next
...
LBB1_5: @cond_true55
        add r8, r0, r6
        ldr r10, LCPI1_1
        str r10, [r8, #+4]

Besides being smaller and more efficient, this makes it immediately
obvious that it is profitable to predicate LBB1_3 now :)

llvm-svn: 35972
2007-04-13 20:42:26 +00:00
Chris Lattner 780c009756 switch LSR to use isLegalAddressingMode instead of other simpler hooks
llvm-svn: 35837
2007-04-09 22:20:14 +00:00
Owen Anderson 8763ba1b88 Completely purge DomSet. This is the (hopefully) final patch for PR1171.
llvm-svn: 35731
2007-04-07 07:17:27 +00:00
Chris Lattner 81e0707552 split some code out into a helper function
llvm-svn: 35615
2007-04-03 05:11:24 +00:00
Chris Lattner f3197a7d53 allow -1 strides to reuse "1" strides.
llvm-svn: 35607
2007-04-02 22:51:58 +00:00
Chris Lattner 28e0e4e11e Pass the type of the store access, not the type of the store, into the
target hook.  This allows us to codegen a loop as:

LBB1_1: @cond_next
        mov r2, #0
        str r2, [r0, +r3, lsl #2]
        add r3, r3, #1
        cmn r3, #1
        bne LBB1_1      @cond_next

instead of:

LBB1_1: @cond_next
        mov r2, #0
        str r2, [r0], #+4
        add r3, r3, #1
        cmn r3, #1
        bne LBB1_1      @cond_next

This looks the same, but has one fewer induction variable (and therefore,
one fewer register) live in the loop.

llvm-svn: 35592
2007-04-02 06:34:44 +00:00
Chris Lattner 8fe3cbe6bd print the type of an inserted IV in -debug mode.
llvm-svn: 35563
2007-04-01 22:21:39 +00:00
Dale Johannesen e5866e7b89 Look through bitcast when finding IVs. (Chris' patch really.)
llvm-svn: 35347
2007-03-26 03:01:27 +00:00
Dale Johannesen bacf4acf65 do not share old induction variables when this would result in invalid
instructions (that would have to be split later)

llvm-svn: 35227
2007-03-20 21:54:54 +00:00
Jeff Cohen 1baf5c84ab Fix some VC++ warnings.
llvm-svn: 35224
2007-03-20 20:43:18 +00:00
Dale Johannesen e3a02be5f1 use types of loads and stores, not address, in CheckForIVReuse
llvm-svn: 35197
2007-03-20 00:47:50 +00:00
Evan Cheng b5eb932c93 Correct type info for isLegalAddressImmediate() check.
llvm-svn: 35086
2007-03-13 20:34:37 +00:00
Evan Cheng 720acdfb31 Use new TargetLowering addressing modes hooks.
llvm-svn: 35072
2007-03-12 23:27:37 +00:00
Devang Patel 58818c530f Increment iterator now because IVUseShouldUsePostIncValue may remove
User from the list of I users.

llvm-svn: 35051
2007-03-09 21:19:53 +00:00
Devang Patel b0743b5d6a Now LoopStrengthReduce is a LoopPass.
llvm-svn: 34984
2007-03-06 21:14:09 +00:00
Reid Spencer 53a3739c80 Finally get this patch right :)
Replace expensive getZExtValue() == 0 calls with isZero() calls.

llvm-svn: 34861
2007-03-02 23:51:25 +00:00
Reid Spencer ba547cbb2a Dang, I've done that twice now! Undo previous commit.
llvm-svn: 34860
2007-03-02 23:37:53 +00:00
Reid Spencer 558990e189 Use more efficient test for one value in a ConstantInt.
llvm-svn: 34859
2007-03-02 23:35:28 +00:00
Reid Spencer 197adfaa0a Reverse a premature commital.
llvm-svn: 34822
2007-03-02 00:31:39 +00:00
Reid Spencer 2e54a15943 Prefer non-virtual calls to ConstantInt::isZero over virtual calls to
Constant::isNullValue() in situations where it is possible.

llvm-svn: 34821
2007-03-02 00:28:52 +00:00
Chris Lattner c473d8e431 Privatize StructLayout::MemberOffsets, adding an accessor
llvm-svn: 34156
2007-02-10 19:55:17 +00:00
Reid Spencer 557ab15e71 Apply the VISIBILITY_HIDDEN field to the remaining anonymous classes in
the Transforms library. This reduces debug library size by 132 KB, debug
binary size by 376 KB, and reduces link time for llvm tools slightly.

llvm-svn: 33939
2007-02-05 23:32:05 +00:00
Chris Lattner 03c4953cdd rename Type::isIntegral to Type::isInteger, eliminating the old Type::isInteger.
rename Type::getIntegralTypeMask to Type::getIntegerTypeMask.

This makes naming much more consistent.  For example, there are now no longer any
instances of IntegerType that are not considered isInteger! :)

llvm-svn: 33225
2007-01-15 02:27:26 +00:00
Chris Lattner 1942249c5b Eliminate calls to isInteger, generalizing code and tightening checks as needed.
llvm-svn: 33218
2007-01-15 01:55:30 +00:00
Reid Spencer bf96e02a54 For PR1097:
Enable complex addressing modes on 64-bit platforms involving two induction
variables by keeping a size and scale in 64-bits not 32.
Patch by Dan Gohman.

llvm-svn: 33011
2007-01-08 16:17:51 +00:00
Chris Lattner 3fe98ae10a no need to worry about int vs uint any more.
llvm-svn: 32946
2007-01-06 01:37:35 +00:00
Reid Spencer c635f47d9a For PR950:
This patch replaces signed integer types with signless ones:
1. [US]Byte -> Int8
2. [U]Short -> Int16
3. [U]Int   -> Int32
4. [U]Long  -> Int64.
5. Removal of isSigned, isUnsigned, getSignedVersion, getUnsignedVersion
   and other methods related to signedness. In a few places this warranted
   identifying the signedness information from other sources.

llvm-svn: 32785
2006-12-31 05:48:39 +00:00
Reid Spencer 266e42b312 For PR950:
This patch removes the SetCC instructions and replaces them with the ICmp
and FCmp instructions. The SetCondInst instruction has been removed and
been replaced with ICmpInst and FCmpInst.

llvm-svn: 32751
2006-12-23 06:05:41 +00:00
Chris Lattner 79a42ac941 Switch over Transforms/Scalar to use the STATISTIC macro. For each statistic
converted, we lose a static initializer.  This also allows GCC to emit warnings
about unused statistics.

llvm-svn: 32690
2006-12-19 21:40:18 +00:00
Reid Spencer df1f19a8ef Change the interface to SCEVExpander::InsertCastOfTo to take a cast opcode
so the decision of which opcode to use is pushed upward to the caller.
Adjust the callers to pass the expected opcode.

llvm-svn: 32535
2006-12-13 08:06:42 +00:00
Reid Spencer b341b0861d Change inferred getCast into specific getCast. Passes all tests.
llvm-svn: 32469
2006-12-12 05:05:00 +00:00
Bill Wendling f3baad3ee1 Changed llvm_ostream et all to OStream. llvm_cerr, llvm_cout, llvm_null, are
now cerr, cout, and NullStream resp.

llvm-svn: 32298
2006-12-07 01:30:32 +00:00
Chris Lattner 700b873130 Detemplatize the Statistic class. The only type it is instantiated with
is 'unsigned'.

llvm-svn: 32279
2006-12-06 17:46:33 +00:00
Reid Spencer 6c38f0bb07 For PR950:
The long awaited CAST patch. This introduces 12 new instructions into LLVM
to replace the cast instruction. Corresponding changes throughout LLVM are
provided. This passes llvm-test, llvm/test, and SPEC CPUINT2000 with the
exception of 175.vpr which fails only on a slight floating point output
difference.

llvm-svn: 31931
2006-11-27 01:05:10 +00:00
Bill Wendling 5dbf43c983 Removed #include <iostream> and replaced with llvm_* streams.
llvm-svn: 31923
2006-11-26 09:46:52 +00:00
Chris Lattner 21eba2da26 If an indvar with a variable stride is used by the exit condition, go ahead
and handle it like constant stride vars.  This fixes some bad codegen in
variable stride cases.  For example, it compiles this:

void foo(int k, int i) {
  for (k=i+i; k <= 8192; k+=i)
    flags2[k] = 0;
}

to:

LBB1_1: #bb.preheader
        movl %eax, %ecx
        addl %ecx, %ecx
        movl L_flags2$non_lazy_ptr, %edx
LBB1_2: #bb
        movb $0, (%edx,%ecx)
        addl %eax, %ecx
        cmpl $8192, %ecx
        jle LBB1_2      #bb
LBB1_5: #return
        ret

or (if the array is local and we are in dynamic-nonpic or static mode):

LBB3_2: #bb
        movb $0, _flags2(%ecx)
        addl %eax, %ecx
        cmpl $8192, %ecx
        jle LBB3_2      #bb

and:

        lis r2, ha16(L_flags2$non_lazy_ptr)
        lwz r2, lo16(L_flags2$non_lazy_ptr)(r2)
        slwi r3, r4, 1
LBB1_2: ;bb
        li r5, 0
        add r6, r4, r3
        stbx r5, r2, r3
        cmpwi cr0, r6, 8192
        bgt cr0, LBB1_5 ;return

instead of:

        leal (%eax,%eax,2), %ecx
        movl %eax, %edx
        addl %edx, %edx
        addl L_flags2$non_lazy_ptr, %edx
        xorl %esi, %esi
LBB1_2: #bb
        movb $0, (%edx,%esi)
        movl %eax, %edi
        addl %esi, %edi
        addl %ecx, %esi
        cmpl $8192, %esi
        jg LBB1_5       #return

and:

        lis r2, ha16(L_flags2$non_lazy_ptr)
        lwz r2, lo16(L_flags2$non_lazy_ptr)(r2)
        mulli r3, r4, 3
        slwi r5, r4, 1
        li r6, 0
        add r2, r2, r5
LBB1_2: ;bb
        li r5, 0
        add r7, r3, r6
        stbx r5, r2, r6
        add r6, r4, r6
        cmpwi cr0, r7, 8192
        ble cr0, LBB1_2 ;bb

This speeds up Benchmarks/Shootout/sieve from 8.533s to 6.464s and
implements LoopStrengthReduce/var_stride_used_by_compare.ll

llvm-svn: 31809
2006-11-17 06:17:33 +00:00
Reid Spencer de46e48420 For PR786:
Turn on -Wunused and -Wno-unused-parameter. Clean up most of the resulting
fall out by removing unused variables. Remaining warnings have to do with
unused functions (I didn't want to delete code without review) and unused
variables in generated code. Maintainers should clean up the remaining
issues when they see them. All changes pass DejaGnu tests and Olden.

llvm-svn: 31380
2006-11-02 20:25:50 +00:00
Chris Lattner a6eb7e0803 break edges more intelligently
llvm-svn: 31257
2006-10-28 06:45:33 +00:00
Chris Lattner 5191c65485 prepare for a change I'm about to make
llvm-svn: 31248
2006-10-28 00:59:20 +00:00
Reid Spencer e0fc4dfc22 For PR950:
This patch implements the first increment for the Signless Types feature.
All changes pertain to removing the ConstantSInt and ConstantUInt classes
in favor of just using ConstantInt.

llvm-svn: 31063
2006-10-20 07:07:24 +00:00
Chris Lattner c2d3d3112e eliminate RegisterOpt. It does the same thing as RegisterPass.
llvm-svn: 29925
2006-08-27 22:42:52 +00:00
Chris Lattner 3d27be1333 s|llvm/Support/Visibility.h|llvm/Support/Compiler.h|
llvm-svn: 29911
2006-08-27 12:54:02 +00:00
Chris Lattner 3ff620178b Changes:
1. Update an obsolete comment.
  2. Make the sorting by base an explicit (though still N^2) step, so
     that the code is more clear on what it is doing.
  3. Partition uses so that uses inside the loop are handled before uses
     outside the loop.

Note that none of these changes currently changes the code inserted by LSR,
but they are a stepping stone to getting there.

This code is the result of some crazy pair programming with Nate. :)

llvm-svn: 29493
2006-08-03 06:34:50 +00:00
Evan Cheng e9c68f52e1 Only reuse a previous IV if it would not require a type conversion.
llvm-svn: 29186
2006-07-18 19:07:58 +00:00
Chris Lattner 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
Evan Cheng 398f70292c RewriteExpr, either the new PHI node of induction variable or the
post-increment value, should be first cast to the appropriated type (to the
type of the common expr). Otherwise, the rewrite of a use based on (common +
iv) may end up with an incorrect type.

llvm-svn: 28735
2006-06-09 00:12:42 +00:00
Reid Spencer 13a1a7a4a6 Get rid of a signed/unsigned compare warning.
llvm-svn: 27625
2006-04-12 19:28:15 +00:00
Chris Lattner f365f5f0c1 Fix spello
llvm-svn: 27052
2006-03-24 07:14:34 +00:00
Chris Lattner 7d80b4f366 silence a bogus gcc warning
llvm-svn: 26953
2006-03-22 17:27:24 +00:00
Evan Cheng c28282bd87 - Fixed a bogus if condition.
- Added more debugging info.
- Allow reuse of IV of negative stride. e.g. -4 stride == 2 * iv of -2 stride.

llvm-svn: 26841
2006-03-18 08:03:12 +00:00
Evan Cheng f09f0ebd48 Sort StrideOrder so we can process the smallest strides first. This allows
for more IV reuses.

llvm-svn: 26837
2006-03-18 00:44:49 +00:00
Evan Cheng 4520698820 Allow users of iv / stride to be rewritten with expression that is a multiply
of a smaller stride even if they have a common loop invariant expression part.

llvm-svn: 26828
2006-03-17 19:52:23 +00:00
Evan Cheng 3df447d354 For each loop, keep track of all the IV expressions inserted indexed by
stride. For a set of uses of the IV of a stride which is a multiple
of another stride, do not insert a new IV expression. Rather, reuse the
previous IV and rewrite the uses as uses of IV expression multiplied by
the factor.

e.g.
x = 0 ...; x ++
y = 0 ...; y += 4
then use of y can be rewritten as use of 4*x for x86.

llvm-svn: 26803
2006-03-16 21:53:05 +00:00
Evan Cheng c567c4efbb Added target lowering hooks which LSR consults to make more intelligent
transformation decisions.

llvm-svn: 26738
2006-03-13 23:14:23 +00:00
Chris Lattner d30c4991a1 Use SCEVExpander::InsertCastOfTo instead of our own code. This reduces
#LLVM LOC, and auto-cse's cast instructions.

llvm-svn: 25974
2006-02-04 09:52:43 +00:00
Chris Lattner 2959f0003e Fix two significant bugs in LSR:
1. When rewriting code in outer loops, sometimes we would insert code into
   inner loops that is invariant in that loop.
2. Notice that 4*(2+x) is 8+4*x and use that to simplify expressions.

This is a performance neutral change.

llvm-svn: 25964
2006-02-04 07:36:50 +00:00
Chris Lattner c597b8a55e Make iostream #inclusion explicit
llvm-svn: 25514
2006-01-22 23:32:06 +00:00
Chris Lattner cb36710ff9 Switch these to using ETForest instead of DominatorSet to compute itself.
Patch written by Daniel Berlin!

llvm-svn: 25202
2006-01-11 05:10:20 +00:00
Chris Lattner 077200737c getRawValue zero extens for unsigned values, use getsextvalue so that we
know that small negative values fit into the immediate field of addressing
modes.

llvm-svn: 24608
2005-12-05 18:23:57 +00:00
Chris Lattner 5df0e36e98 My previous patch was too conservative. Reject FP and void types, but do
allow pointer types.

llvm-svn: 23859
2005-10-21 05:45:41 +00:00
Chris Lattner 0c0b38bb4c Do NOT touch FP ops with LSR. This fixes a testcase Nate sent me from an
inner loop like this:

LBB_RateConvertMono8AltiVec_2:  ; no_exit
        lis r2, ha16(.CPI_RateConvertMono8AltiVec_0)
        lfs f3, lo16(.CPI_RateConvertMono8AltiVec_0)(r2)
        fmr f3, f3
        fadd f0, f2, f0
        fadd f3, f0, f3
        fcmpu cr0, f3, f1
        bge cr0, LBB_RateConvertMono8AltiVec_2  ; no_exit

to an inner loop like this:

LBB_RateConvertMono8AltiVec_1:  ; no_exit
        fsub f2, f2, f1
        fcmpu cr0, f2, f1
        fmr f0, f2
        bge cr0, LBB_RateConvertMono8AltiVec_1  ; no_exit

Doh! good catch!

llvm-svn: 23838
2005-10-20 04:47:10 +00:00
Chris Lattner 192cd18f53 Fix (hopefully the last) issue where LSR is nondeterminstic. When pulling
out CSE's of base expressions it could build a result whose order was
nondet.

llvm-svn: 23698
2005-10-11 18:41:04 +00:00
Chris Lattner 5c9d63da31 Fix another problem where LSR was being nondeterminstic. Also remove elements
from the end of a vector instead of the beginning

llvm-svn: 23697
2005-10-11 18:30:57 +00:00
Chris Lattner b7a3894e7c Fix another lsr-is-nondeterministic case
llvm-svn: 23695
2005-10-11 18:17:57 +00:00
Chris Lattner eb4be8b942 Hrm, you didn't see this.
llvm-svn: 23673
2005-10-09 06:24:02 +00:00
Chris Lattner 4ea0a3eaac Fix a source of non-determinism in the backend: the order of processing
IV strides dependend on the pointer order of the strides in memory.
Non-determinism is bad.

llvm-svn: 23672
2005-10-09 06:20:55 +00:00
Chris Lattner f07a587c79 Make IVUseShouldUsePostIncValue more aggressive when the use is a PHI. In
particular, it should realize that phi's use their values in the pred block
not the phi block itself.  This change turns our em3d loop from this:

_test:
        cmpwi cr0, r4, 0
        bgt cr0, LBB_test_2     ; entry.no_exit_crit_edge
LBB_test_1:     ; entry.loopexit_crit_edge
        li r2, 0
        b LBB_test_6    ; loopexit
LBB_test_2:     ; entry.no_exit_crit_edge
        li r6, 0
LBB_test_3:     ; no_exit
        or r2, r6, r6
        lwz r6, 0(r3)
        cmpw cr0, r6, r5
        beq cr0, LBB_test_6     ; loopexit
LBB_test_4:     ; endif
        addi r3, r3, 4
        addi r6, r2, 1
        cmpw cr0, r6, r4
        blt cr0, LBB_test_3     ; no_exit
LBB_test_5:     ; endif.loopexit.loopexit_crit_edge
        addi r3, r2, 1
        blr
LBB_test_6:     ; loopexit
        or r3, r2, r2
        blr

into:

_test:
        cmpwi cr0, r4, 0
        bgt cr0, LBB_test_2     ; entry.no_exit_crit_edge
LBB_test_1:     ; entry.loopexit_crit_edge
        li r2, 0
        b LBB_test_5    ; loopexit
LBB_test_2:     ; entry.no_exit_crit_edge
        li r6, 0
LBB_test_3:     ; no_exit
        lwz r2, 0(r3)
        cmpw cr0, r2, r5
        or r2, r6, r6
        beq cr0, LBB_test_5     ; loopexit
LBB_test_4:     ; endif
        addi r3, r3, 4
        addi r6, r6, 1
        cmpw cr0, r6, r4
        or r2, r6, r6
        blt cr0, LBB_test_3     ; no_exit
LBB_test_5:     ; loopexit
        or r3, r2, r2
        blr


Unfortunately, this is actually worse code, because the register coallescer
is getting confused somehow.  If it were doing its job right, it could turn the
code into this:

_test:
        cmpwi cr0, r4, 0
        bgt cr0, LBB_test_2     ; entry.no_exit_crit_edge
LBB_test_1:     ; entry.loopexit_crit_edge
        li r6, 0
        b LBB_test_5    ; loopexit
LBB_test_2:     ; entry.no_exit_crit_edge
        li r6, 0
LBB_test_3:     ; no_exit
        lwz r2, 0(r3)
        cmpw cr0, r2, r5
        beq cr0, LBB_test_5     ; loopexit
LBB_test_4:     ; endif
        addi r3, r3, 4
        addi r6, r6, 1
        cmpw cr0, r6, r4
        blt cr0, LBB_test_3     ; no_exit
LBB_test_5:     ; loopexit
        or r3, r6, r6
        blr

... which I'll work on next. :)

llvm-svn: 23604
2005-10-03 02:50:05 +00:00
Chris Lattner e4ed42a426 Refactor some code into a function
llvm-svn: 23603
2005-10-03 01:04:44 +00:00
Chris Lattner 360928dbed This break is bogus and I have no idea why it was there. Basically it prevents
memoizing code when IV's are used by phinodes outside of loops.  In a simple
example, we were getting this code before (note that r6 and r7 are isomorphic
IV's):

        li r6, 0
        or r7, r6, r6
LBB_test_3:     ; no_exit
        lwz r2, 0(r3)
        cmpw cr0, r2, r5
        or r2, r7, r7
        beq cr0, LBB_test_5     ; loopexit
LBB_test_4:     ; endif
        addi r2, r7, 1
        addi r7, r7, 1
        addi r3, r3, 4
        addi r6, r6, 1
        cmpw cr0, r6, r4
        blt cr0, LBB_test_3     ; no_exit

Now we get:

        li r6, 0
LBB_test_3:     ; no_exit
        or r2, r6, r6
        lwz r6, 0(r3)
        cmpw cr0, r6, r5
        beq cr0, LBB_test_6     ; loopexit
LBB_test_4:     ; endif
        addi r3, r3, 4
        addi r6, r2, 1
        cmpw cr0, r6, r4
        blt cr0, LBB_test_3     ; no_exit

this was noticed in em3d.

llvm-svn: 23602
2005-10-03 00:37:33 +00:00
Chris Lattner 8fcce170cf when checking if we should move a split edge block outside of a loop,
check the presplit pred, not the post-split pred.  This was causing us
to make the wrong decision in some cases, leaving the critical edge block
in the loop.

llvm-svn: 23601
2005-10-03 00:31:52 +00:00
Chris Lattner 92233d2175 Make the pass name simpler
llvm-svn: 23476
2005-09-27 21:10:32 +00:00
Chris Lattner fd018c8dfe Fix an issue where LSR would miss rewriting a use of an IV expression by a PHI node that is not the original PHI.
This fixes up a dot-product loop in galgel, speeding it up from 18.47s to
16.13s.

llvm-svn: 23327
2005-09-13 02:09:55 +00:00
Chris Lattner 8048b85e8f Fix a regression from last night, which caused this pass to create invalid
code for IV uses outside of loops that are not dominated by the latch block.
We should only convert these uses to use the post-inc value if they ARE
dominated by the latch block.

Also use a new LoopInfo method to simplify some code.

This fixes Transforms/LoopStrengthReduce/2005-09-12-UsesOutOutsideOfLoop.ll

llvm-svn: 23318
2005-09-12 17:11:27 +00:00
Chris Lattner a67648396a _test:
li r2, 0
LBB_test_1:     ; no_exit.2
        li r5, 0
        stw r5, 0(r3)
        addi r2, r2, 1
        addi r3, r3, 4
        cmpwi cr0, r2, 701
        blt cr0, LBB_test_1     ; no_exit.2
LBB_test_2:     ; loopexit.2.loopexit
        addi r2, r2, 1
        stw r2, 0(r4)
        blr
[zion ~/llvm]$ cat > ~/xx
Uses of IV's outside of the loop should use hte post-incremented version
of the IV, not the preincremented version.  This helps many loops (e.g. in sixtrack)
which used to generate code like this (this is the code from the
dont-hoist-simple-loop-constants.ll testcase):

_test:
        li r2, 0                 **** IV starts at 0
LBB_test_1:     ; no_exit.2
        or r5, r2, r2            **** Copy for loop exit
        li r2, 0
        stw r2, 0(r3)
        addi r3, r3, 4
        addi r2, r5, 1
        addi r6, r5, 2           **** IV+2
        cmpwi cr0, r6, 701
        blt cr0, LBB_test_1     ; no_exit.2
LBB_test_2:     ; loopexit.2.loopexit
        addi r2, r5, 2       ****  IV+2
        stw r2, 0(r4)
        blr

And now generated code like this:

_test:
        li r2, 1               *** IV starts at 1
LBB_test_1:     ; no_exit.2
        li r5, 0
        stw r5, 0(r3)
        addi r2, r2, 1
        addi r3, r3, 4
        cmpwi cr0, r2, 701     *** IV.postinc + 0
        blt cr0, LBB_test_1
LBB_test_2:     ; loopexit.2.loopexit
        stw r2, 0(r4)          *** IV.postinc + 0
        blr

llvm-svn: 23313
2005-09-12 06:04:47 +00:00
Chris Lattner 530fe6ab30 implement Transforms/LoopStrengthReduce/dont-hoist-simple-loop-constants.ll.
We used to emit this code for it:

_test:
        li r2, 1     ;; Value tying up a register for the whole loop
        li r5, 0
LBB_test_1:     ; no_exit.2
        or r6, r5, r5
        li r5, 0
        stw r5, 0(r3)
        addi r5, r6, 1
        addi r3, r3, 4
        add r7, r2, r5  ;; should be addi r7, r5, 1
        cmpwi cr0, r7, 701
        blt cr0, LBB_test_1     ; no_exit.2
LBB_test_2:     ; loopexit.2.loopexit
        addi r2, r6, 2
        stw r2, 0(r4)
        blr

now we emit this:

_test:
        li r2, 0
LBB_test_1:     ; no_exit.2
        or r5, r2, r2
        li r2, 0
        stw r2, 0(r3)
        addi r3, r3, 4
        addi r2, r5, 1
        addi r6, r5, 2   ;; whoa, fold those adds!
        cmpwi cr0, r6, 701
        blt cr0, LBB_test_1     ; no_exit.2
LBB_test_2:     ; loopexit.2.loopexit
        addi r2, r5, 2
        stw r2, 0(r4)
        blr

more improvement coming.

llvm-svn: 23306
2005-09-10 01:18:45 +00:00
Chris Lattner ea7dfd53d6 Fix Transforms/LoopStrengthReduce/2005-08-17-OutOfLoopVariant.ll, a crash
on 177.mesa

llvm-svn: 22843
2005-08-17 21:22:41 +00:00
Chris Lattner 2bf7cb5213 Use a new helper to split critical edges, making the code simpler.
Do not claim to not change the CFG.  We do change the cfg to split critical
edges.  This isn't causing us a problem now, but could likely do so in the
future.

llvm-svn: 22824
2005-08-17 06:35:16 +00:00
Chris Lattner 5cf983ee0f Fix a bad case in gzip where we put lots of things in registers across the
loop, because a IV-dependent value was used outside of the loop and didn't
have immediate-folding capability

llvm-svn: 22798
2005-08-16 00:38:11 +00:00
Chris Lattner 47d3ec3525 Ooops, don't forget to clear this. The real inner loop is now:
.LBB_foo_3:     ; no_exit.1
        lfd f2, 0(r9)
        lfd f3, 8(r9)
        fmul f4, f1, f2
        fmadd f4, f0, f3, f4
        stfd f4, 8(r9)
        fmul f3, f1, f3
        fmsub f2, f0, f2, f3
        stfd f2, 0(r9)
        addi r9, r9, 16
        addi r8, r8, 1
        cmpw cr0, r8, r4
        ble .LBB_foo_3  ; no_exit.1

llvm-svn: 22782
2005-08-13 07:42:01 +00:00
Chris Lattner 5949d49032 Recursively scan scev expressions for common subexpressions. This allows us
to handle nested loops much better, for example, by being able to tell that
these two expressions:

{( 8 + ( 16 * ( 1 +  %Tmp11 +  %Tmp12)) +  %c_),+,( 16 *  %Tmp 12)}<loopentry.1>

{(( 16 * ( 1 +  %Tmp11 +  %Tmp12)) +  %c_),+,( 16 *  %Tmp12)}<loopentry.1>

Have the following common part that can be shared:
{(( 16 * ( 1 +  %Tmp11 +  %Tmp12)) +  %c_),+,( 16 *  %Tmp12)}<loopentry.1>

This allows us to codegen an important inner loop in 168.wupwise as:

.LBB_foo_4:     ; no_exit.1
        lfd f2, 16(r9)
        fmul f3, f0, f2
        fmul f2, f1, f2
        fadd f4, f3, f2
        stfd f4, 8(r9)
        fsub f2, f3, f2
        stfd f2, 16(r9)
        addi r8, r8, 1
        addi r9, r9, 16
        cmpw cr0, r8, r4
        ble .LBB_foo_4  ; no_exit.1

instead of:

.LBB_foo_3:     ; no_exit.1
        lfdx f2, r6, r9
        add r10, r6, r9
        lfd f3, 8(r10)
        fmul f4, f1, f2
        fmadd f4, f0, f3, f4
        stfd f4, 8(r10)
        fmul f3, f1, f3
        fmsub f2, f0, f2, f3
        stfdx f2, r6, r9
        addi r9, r9, 16
        addi r8, r8, 1
        cmpw cr0, r8, r4
        ble .LBB_foo_3  ; no_exit.1

llvm-svn: 22781
2005-08-13 07:27:18 +00:00
Chris Lattner 8447b49526 When splitting critical edges, make sure not to leave the new block in the
middle of the loop.  This turns a critical loop in gzip into this:

.LBB_test_1:    ; loopentry
        or r27, r28, r28
        add r28, r3, r27
        lhz r28, 3(r28)
        add r26, r4, r27
        lhz r26, 3(r26)
        cmpw cr0, r28, r26
        bne .LBB_test_8 ; loopentry.loopexit_crit_edge
.LBB_test_2:    ; shortcirc_next.0
        add r28, r3, r27
        lhz r28, 5(r28)
        add r26, r4, r27
        lhz r26, 5(r26)
        cmpw cr0, r28, r26
        bne .LBB_test_7 ; shortcirc_next.0.loopexit_crit_edge
.LBB_test_3:    ; shortcirc_next.1
        add r28, r3, r27
        lhz r28, 7(r28)
        add r26, r4, r27
        lhz r26, 7(r26)
        cmpw cr0, r28, r26
        bne .LBB_test_6 ; shortcirc_next.1.loopexit_crit_edge
.LBB_test_4:    ; shortcirc_next.2
        add r28, r3, r27
        lhz r26, 9(r28)
        add r28, r4, r27
        lhz r25, 9(r28)
        addi r28, r27, 8
        cmpw cr7, r26, r25
        mfcr r26, 1
        rlwinm r26, r26, 31, 31, 31
        add r25, r8, r27
        cmpw cr7, r25, r7
        mfcr r25, 1
        rlwinm r25, r25, 29, 31, 31
        and. r26, r26, r25
        bne .LBB_test_1 ; loopentry

instead of this:

.LBB_test_1:    ; loopentry
        or r27, r28, r28
        add r28, r3, r27
        lhz r28, 3(r28)
        add r26, r4, r27
        lhz r26, 3(r26)
        cmpw cr0, r28, r26
        beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2:    ; loopentry.loopexit_crit_edge
        add r2, r30, r27
        add r8, r29, r27
        b .LBB_test_9   ; loopexit
.LBB_test_3:    ; shortcirc_next.0
        add r28, r3, r27
        lhz r28, 5(r28)
        add r26, r4, r27
        lhz r26, 5(r26)
        cmpw cr0, r28, r26
        beq .LBB_test_5 ; shortcirc_next.1
.LBB_test_4:    ; shortcirc_next.0.loopexit_crit_edge
        add r2, r11, r27
        add r8, r12, r27
        b .LBB_test_9   ; loopexit
.LBB_test_5:    ; shortcirc_next.1
        add r28, r3, r27
        lhz r28, 7(r28)
        add r26, r4, r27
        lhz r26, 7(r26)
        cmpw cr0, r28, r26
        beq .LBB_test_7 ; shortcirc_next.2
.LBB_test_6:    ; shortcirc_next.1.loopexit_crit_edge
        add r2, r9, r27
        add r8, r10, r27
        b .LBB_test_9   ; loopexit
.LBB_test_7:    ; shortcirc_next.2
        add r28, r3, r27
        lhz r26, 9(r28)
        add r28, r4, r27
        lhz r25, 9(r28)
        addi r28, r27, 8
        cmpw cr7, r26, r25
        mfcr r26, 1
        rlwinm r26, r26, 31, 31, 31
        add r25, r8, r27
        cmpw cr7, r25, r7
        mfcr r25, 1
        rlwinm r25, r25, 29, 31, 31
        and. r26, r26, r25
        bne .LBB_test_1 ; loopentry

Next up, improve the code for the loop.

llvm-svn: 22769
2005-08-12 22:22:17 +00:00
Chris Lattner 4fec86d348 Fix a FIXME: if we are inserting code for a PHI argument, split the critical
edge so that the code is not always executed for both operands.  This
prevents LSR from inserting code into loops whose exit blocks contain
PHI uses of IV expressions (which are outside of loops).  On gzip, for
example, we turn this ugly code:

.LBB_test_1:    ; loopentry
        add r27, r3, r28
        lhz r27, 3(r27)
        add r26, r4, r28
        lhz r26, 3(r26)
        add r25, r30, r28    ;; Only live if exiting the loop
        add r24, r29, r28    ;; Only live if exiting the loop
        cmpw cr0, r27, r26
        bne .LBB_test_5 ; loopexit

into this:

.LBB_test_1:    ; loopentry
        or r27, r28, r28
        add r28, r3, r27
        lhz r28, 3(r28)
        add r26, r4, r27
        lhz r26, 3(r26)
        cmpw cr0, r28, r26
        beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2:    ; loopentry.loopexit_crit_edge
        add r2, r30, r27
        add r8, r29, r27
        b .LBB_test_9   ; loopexit
.LBB_test_2:    ; shortcirc_next.0
        ...
        blt .LBB_test_1


into this:

.LBB_test_1:    ; loopentry
        or r27, r28, r28
        add r28, r3, r27
        lhz r28, 3(r28)
        add r26, r4, r27
        lhz r26, 3(r26)
        cmpw cr0, r28, r26
        beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2:    ; loopentry.loopexit_crit_edge
        add r2, r30, r27
        add r8, r29, r27
        b .LBB_t_3:    ; shortcirc_next.0
.LBB_test_3:    ; shortcirc_next.0
        ...
        blt .LBB_test_1


Next step: get the block out of the loop so that the loop is all
fall-throughs again.

llvm-svn: 22766
2005-08-12 22:06:11 +00:00
Chris Lattner edff91a49a Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:

void foo(float *a, float *b, int n, int stride_a, int stride_b) {
  int i;
  for (i=0; i<n; i++)
      a[i*stride_a] = b[i*stride_b];
}

we now emit:

.LBB_foo2_2:    ; no_exit
        lfs f0, 0(r4)
        stfs f0, 0(r3)
        addi r7, r7, 1
        add r4, r2, r4
        add r3, r6, r3
        cmpw cr0, r7, r5
        blt .LBB_foo2_2 ; no_exit

instead of:

.LBB_foo_2:     ; no_exit
        mullw r8, r2, r7     ;; multiply!
        slwi r8, r8, 2
        lfsx f0, r4, r8
        mullw r8, r2, r6     ;; multiply!
        slwi r8, r8, 2
        stfsx f0, r3, r8
        addi r2, r2, 1
        cmpw cr0, r2, r5
        blt .LBB_foo_2  ; no_exit

loops with variable strides occur pretty often.  For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.

Now we can allow indvars to turn functions written like this:

void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
  int i, ai = 0, bi = 0;
  for (i=0; i<n; i++)
    {
      a[ai] = b[bi];
      ai += stride_a;
      bi += stride_b;
    }
}

into code like the above for better analysis.  With this patch, they generate
identical code.

llvm-svn: 22740
2005-08-10 00:45:21 +00:00
Chris Lattner dde7dc525e Fix Regression/Transforms/LoopStrengthReduce/phi_node_update_multiple_preds.ll
by being more careful about updating PHI nodes

llvm-svn: 22739
2005-08-10 00:35:32 +00:00
Chris Lattner c6c4d99a21 Fix some 80 column violations.
Once we compute the evolution for a GEP, tell SE about it.  This allows users
of the GEP to know it, if the users are not direct.  This allows us to compile
this testcase:

void fbSolidFillmmx(int w, unsigned char *d) {
    while (w >= 64) {
        *(unsigned long long *) (d +  0) = 0;
        *(unsigned long long *) (d +  8) = 0;
        *(unsigned long long *) (d + 16) = 0;
        *(unsigned long long *) (d + 24) = 0;
        *(unsigned long long *) (d + 32) = 0;
        *(unsigned long long *) (d + 40) = 0;
        *(unsigned long long *) (d + 48) = 0;
        *(unsigned long long *) (d + 56) = 0;
        w -= 64;
        d += 64;
    }
}

into:

.LBB_fbSolidFillmmx_2:  ; no_exit
        li r2, 0
        stw r2, 0(r4)
        stw r2, 4(r4)
        stw r2, 8(r4)
        stw r2, 12(r4)
        stw r2, 16(r4)
        stw r2, 20(r4)
        stw r2, 24(r4)
        stw r2, 28(r4)
        stw r2, 32(r4)
        stw r2, 36(r4)
        stw r2, 40(r4)
        stw r2, 44(r4)
        stw r2, 48(r4)
        stw r2, 52(r4)
        stw r2, 56(r4)
        stw r2, 60(r4)
        addi r4, r4, 64
        addi r3, r3, -64
        cmpwi cr0, r3, 63
        bgt .LBB_fbSolidFillmmx_2       ; no_exit

instead of:

.LBB_fbSolidFillmmx_2:  ; no_exit
        li r11, 0
        stw r11, 0(r4)
        stw r11, 4(r4)
        stwx r11, r10, r4
        add r12, r10, r4
        stw r11, 4(r12)
        stwx r11, r9, r4
        add r12, r9, r4
        stw r11, 4(r12)
        stwx r11, r8, r4
        add r12, r8, r4
        stw r11, 4(r12)
        stwx r11, r7, r4
        add r12, r7, r4
        stw r11, 4(r12)
        stwx r11, r6, r4
        add r12, r6, r4
        stw r11, 4(r12)
        stwx r11, r5, r4
        add r12, r5, r4
        stw r11, 4(r12)
        stwx r11, r2, r4
        add r12, r2, r4
        stw r11, 4(r12)
        addi r4, r4, 64
        addi r3, r3, -64
        cmpwi cr0, r3, 63
        bgt .LBB_fbSolidFillmmx_2       ; no_exit

llvm-svn: 22737
2005-08-09 23:39:36 +00:00