Commit Graph

299 Commits

Author SHA1 Message Date
Chris Lattner 9023155621 Initial checkin of PRE on LLVM. This implementation is still lacking in
several ways:

 * Load expressions are not PRE'd well.  Alias Analysis should be used to
   get accurate information when computing anticipatibility.
 * The expression collection implementation does not handle PHI nodes properly,
   thus the implementation misses many opportunities to PRE.
 * This code could be sped up quite a bit

Despite these flaws, the code seems to work well, and handles PR's as one
would expect.

llvm-svn: 5759
2003-03-31 19:55:43 +00:00
Chris Lattner c42046f9f1 Move BreakCriticalEdges pass to lib/Transforms/Utils
llvm-svn: 5754
2003-03-21 21:43:19 +00:00
Chris Lattner 7fb29e19f0 Add the following instcombine xforms:
- Implement simple reassociation: (A|c1)|(B|c2) == (A|B)|(c1|c2)
  - (A & C1)+(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
  - (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0

llvm-svn: 5743
2003-03-11 00:12:48 +00:00
Chris Lattner 934a64cfa4 Implement: -A*-B == A*B
llvm-svn: 5740
2003-03-10 23:23:04 +00:00
Chris Lattner 3e327a4334 Add new transformation: // (~A | ~B) == (~(A & B))
llvm-svn: 5738
2003-03-10 23:13:59 +00:00
Chris Lattner bb74e2284a Generalize not and neg comparison testers to allow constant to be considered not'able and neg'able. This
allows optimization of this:
int %test4(int %A, int %B) {
        %a = xor int %A, -1
        %c = and int %a, 5    ; 5 = ~c2
        %d = xor int %c, -1
        ret int %d
}

into this:
int %test4(int %A, int %B) {            ; No predecessors!
        %c.demorgan = or int %A, -6             ; <int> [#uses=1]
        ret int %c.demorgan
}

llvm-svn: 5736
2003-03-10 23:06:50 +00:00
Chris Lattner dcf240a649 Generalize (A+c1)+c2 optimization to work with all associative operators
llvm-svn: 5733
2003-03-10 21:43:22 +00:00
Chris Lattner 3695257724 Minor change, no functionality diff
llvm-svn: 5731
2003-03-10 19:20:30 +00:00
Chris Lattner e8d6c604bb Fix bug: (x << 100) wasn't folded to 0, but (x >> 100) was (when x is unsigned)
Implement new shift optimizations for shifting the result of a shift.

llvm-svn: 5729
2003-03-10 19:16:08 +00:00
Chris Lattner 1bbb7b6f19 Implement: (A|B)^B == A & (~B)
llvm-svn: 5728
2003-03-10 18:24:17 +00:00
Chris Lattner 235af5673b Implement %test7 in InstCombine/getelementptr.ll
llvm-svn: 5704
2003-03-05 22:33:14 +00:00
Chris Lattner 0592bb7654 Convert LICM over to use AliasSetTracker. Besides being nicer, this automatically
allows LICM to use access sizes to help alias analysis be more precise.

llvm-svn: 5693
2003-03-03 23:32:45 +00:00
Chris Lattner e27406eb59 Change the mem2reg interface to accept a TargetData argument
llvm-svn: 5685
2003-03-03 17:25:18 +00:00
Chris Lattner e19a9f8ef7 Fix bug: LICM/2003-02-28-PromoteDifferentType.ll
llvm-svn: 5675
2003-02-28 19:21:40 +00:00
Chris Lattner 49eb0e3026 Fix bug I introduced yesterday :(
llvm-svn: 5669
2003-02-28 16:54:17 +00:00
Chris Lattner 32a39c26e0 Fix bug: LICM/2003-02-27-PreheaderExitNodeUpdate.ll
llvm-svn: 5667
2003-02-28 03:07:54 +00:00
Chris Lattner f2d9f94a33 Fix bug: 2003-02-27-PreheaderExitNodeUpdate.ll by updating exit node info
llvm-svn: 5664
2003-02-27 22:48:57 +00:00
Chris Lattner 10b2b055e8 Add a new assertion to check that stuff is happening right
Ironically the exit block modification code wasn't updating the exit block
information itself.  Fix this.

llvm-svn: 5659
2003-02-27 22:31:07 +00:00
Chris Lattner 1ad80e2511 Fix bug: 2003-02-27-StoreSinkPHIs.ll
llvm-svn: 5658
2003-02-27 21:59:36 +00:00
Chris Lattner 4e2fbfb4c3 * Don't forget to update Loop information!
* Remove bogus assertion: there may be a single outside predecessor and
    still need a new loop-preheader if the predecessor has multiple
    successors.  See bug: LICM/2003-02-27-PreheaderProblem.ll

llvm-svn: 5655
2003-02-27 21:50:19 +00:00
Chris Lattner 650096a061 * Significant changes to the preheader insertion pass:
- Now we perform loop exit-block splitting to ensure exit blocks are
     always dominated by the loop header.
   - We now preserve dominance frontier information
   - This fixes bug: LICM/2003-02-26-LoopExitNotDominated.ll

llvm-svn: 5652
2003-02-27 20:27:08 +00:00
Chris Lattner 71b82bb494 Convert to work with new AliasAnalysis interface by conservatively assuming all pointers are arbitrarily large accesses
llvm-svn: 5636
2003-02-26 19:28:57 +00:00
Chris Lattner 4869f3700e Rename Instruction::hasSideEffects() -> mayWriteToMemory()
llvm-svn: 5620
2003-02-24 20:48:32 +00:00
Chris Lattner 45d67d6ea7 Initial implementation of Loop invariant memory->scalar promotion
llvm-svn: 5613
2003-02-24 03:52:32 +00:00
Chris Lattner 26f159046c Split mem2reg promotion into two parts: a function which does the work, and
a pass which wraps the function.  This allows other passes to use the functionality

llvm-svn: 5610
2003-02-22 23:57:48 +00:00
Chris Lattner 57c8d99bbd 4 new transformations:
* X*C + X --> X * (C+1)
  * X + X*C --> X * (C+1)
  * X - X*C --> X * (1-C)
  * X*C - X --> X * (C-1)

llvm-svn: 5592
2003-02-18 19:57:07 +00:00
Chris Lattner 3082c5a070 Add a variety of new transformations:
* A & ~A == 0
  * A / (2^c) == A >> c  if unsigned
  * 0 / A == 0
  * 1.0 * A == A
  * A * (2^c) == A << c
  * A ^ ~A == -1
  * A | ~A == -1
  * 0 % X = 0
  * A % (2^c) == A & (c-1) if unsigned
  * A - (A & B) == A & ~B
  * -1 - A == ~A

llvm-svn: 5587
2003-02-18 19:28:33 +00:00
Chris Lattner f683599915 Fix typo
llvm-svn: 5458
2003-02-01 04:50:59 +00:00
Chris Lattner 6fa1629c47 Fix bug: ADCE/2003-01-22-PredecessorProblem.ll
llvm-svn: 5405
2003-01-23 02:12:18 +00:00
Chris Lattner 3aa7767352 Fix a huge performance problem in reassociate by introducing a
rank map cache for instruction ranks

llvm-svn: 5030
2002-12-15 03:56:00 +00:00
Chris Lattner f96c8befe0 Minor changes:
* Reword comment to make more clear
  * Don't print out BB's after modification made
  * Don't delete and new an instruction when we need to move something, just move it.

llvm-svn: 5029
2002-12-15 03:49:50 +00:00
Chris Lattner 688ac5b723 Remove dead code
llvm-svn: 4952
2002-12-07 18:42:13 +00:00
Chris Lattner c6509f487b Fix bug: 2002-12-05-MissedConstProp.ll pointed out by Casey Carter
llvm-svn: 4935
2002-12-05 22:41:53 +00:00
Chris Lattner 98cf1f5d64 - Eliminated the deferred symbol table stuff in Module & Function, it really
wasn't an optimization and it was causing lots of bugs.

llvm-svn: 4779
2002-11-20 18:36:02 +00:00
Chris Lattner a2620ac156 Fix warning
llvm-svn: 4649
2002-11-09 00:49:43 +00:00
Chris Lattner 0da31d6648 Fix spelling error
llvm-svn: 4645
2002-11-08 23:18:37 +00:00
Chris Lattner 5d606a09ff Instcombine this away:
%A = getelementptr [4 x int]* %M, long 0, long 0
        %B = getelementptr int* %A, long 2

llvm-svn: 4536
2002-11-04 16:43:32 +00:00
Chris Lattner 1085bdf49f Add a transformation to turn:
malloc Ty, C
int
  malloc [C x Ty], 1

llvm-svn: 4534
2002-11-04 16:18:53 +00:00
Chris Lattner 669c6cf459 Make sure to _delete_ memory allocated by worklists
llvm-svn: 4523
2002-11-04 02:54:22 +00:00
Chris Lattner 8fdf75cde5 Fixes to the reassociate pass to make it respect dominance properties
Huge thanks go to Casey Carter for writing this fix, reassociate is now
reoperational!

llvm-svn: 4471
2002-10-31 17:12:59 +00:00
Chris Lattner 12764c85e8 BreakCriticalEdges should update dominance frontier information as well as
other dominance stuff.  Patch contributed by Casey Carter

llvm-svn: 4457
2002-10-31 02:44:36 +00:00
Misha Brukman 632df28deb Fix spelling of `propagate'.
llvm-svn: 4423
2002-10-29 23:06:16 +00:00
Chris Lattner 820d971233 - Rename AnalysisUsage::preservesAll to getPreservesAll & preservesCFG to
setPreservesCFG to be less confusing.

llvm-svn: 4255
2002-10-21 20:00:28 +00:00
Chris Lattner b7368500dc - Fix bug: cee/2002-10-07-NoImmediateDominator.ll
llvm-svn: 4081
2002-10-08 21:53:51 +00:00
Chris Lattner 736709feda - Checkin LARGE number of Changes to CEE pass that will make it much more
powerful, but that are largely disabled.  The basic idea here is that it
    is trying to forward branches across basic blocks that have PHI nodes in
    it, which are crucial to be able to handle cases like whet.ll.
    Unfortunately we are not updating SSA correctly, causing sim.c to die, and I
    don't have time to fix the regression now, so I must disable the
    functionality.

llvm-svn: 4077
2002-10-08 21:34:15 +00:00
Chris Lattner 5ac72defc9 Expose isCriticalEdge & SplitCriticalEdge methods from crit-edges pass
llvm-svn: 4075
2002-10-08 21:06:27 +00:00
Chris Lattner 394617f38b It is illegal for PHI nodes to have zero values, delete the code to handle them
llvm-svn: 4071
2002-10-08 17:07:39 +00:00
Chris Lattner 2e0fb39d87 Fold ashr -1, X into -1
llvm-svn: 4070
2002-10-08 16:16:40 +00:00
Chris Lattner bf3a099a62 Updates to work with recent Statistic's changes:
* Renamed StatisticReporter.h/cpp to Statistic.h/cpp
    * Broke constructor to take two const char * arguments instead of one, so
      that indendation can be taken care of automatically.
    * Sort the list by pass name when printing
    * Make sure to print all statistics as a group, instead of randomly when
      the statistics dtors are called.
    * Updated ProgrammersManual with new semantics.

llvm-svn: 4002
2002-10-01 22:38:41 +00:00
Chris Lattner 05e86302f4 Fix bug in LICM that caused the previous big win. :(
llvm-svn: 3980
2002-09-29 22:26:07 +00:00