Commit Graph

203 Commits

Author SHA1 Message Date
Nate Begeman 784a0dcbd0 Move virtual function call out of loop to speed up getFreePhysReg by about
20%, shaving 0.1s off hbd compile time on my g5.  Yay.

llvm-svn: 18592
2004-12-07 05:25:53 +00:00
Chris Lattner b75e790a2b * There is no reason for SpillWeights to be an instance var
* Do not put fixed registers into the unhandled set.  This means they will
  never find their way into the inactive, active, or handled sets, so we
  can simplify a bunch of code.

llvm-svn: 17945
2004-11-18 06:01:45 +00:00
Chris Lattner 08ec603bb0 Fix a couple of bugs where we considered physregs past their range as possibly
intersecting an interval.

llvm-svn: 17939
2004-11-18 04:33:31 +00:00
Chris Lattner 49ff5f0477 Start using the iterators in the fixed_ intervals to avoid having to binary
search physreg intervals every time we access it.  This takes another
half second off of linscan.

llvm-svn: 17937
2004-11-18 04:13:02 +00:00
Chris Lattner 850852c789 Add a counter for the number of times linscan has to backtrack. Start using
the iterator hints we have to speed up overlaps().  This speeds linscan up
by about .2s (out of 8.7) on 175.vpr for PPC.

llvm-svn: 17935
2004-11-18 03:49:30 +00:00
Chris Lattner a1f7779876 * Improve comments/documentation substantially
* Eliminate the releaseMemory method, this is not an analysis
* Change the fixed, active, and inactive lists of intervals to maintain an
  iterator for the current position in the interval.  This allows us to do
  constant time increments of the iterator instead of having to do a binary
  search to find our liverange in our liveinterval all of the time, which
  substantially speeds up cases where LiveIntervals have many LiveRanges
  - which is very common for physical registers.  On targets with many
  physregs, this can make a noticable difference.

  With a release build of LLC for PPC, this halves the time in
  processInactiveIntervals and processActiveIntervals, from 1.5s to .75s.

  This also lays the ground for more to come.

llvm-svn: 17933
2004-11-18 02:42:27 +00:00
Chris Lattner a51f5eeac1 Rename some methods, use 'begin' instead of 'start', add new LiveInterval
iterator/begin/end members.

llvm-svn: 17930
2004-11-18 01:29:39 +00:00
Chris Lattner c22ffc26b0 Do not make i have bigger scope that we need
llvm-svn: 17483
2004-11-05 04:47:37 +00:00
Chris Lattner ddd5229f3b Free the VirtRegMap at the end of MachineFunction processing instead of at
the beginning of processing the next one.

llvm-svn: 16605
2004-09-30 02:02:33 +00:00
Alkis Evlogimenos a5c04ee50f Fixes to make LLVM compile with vc7.1.
Patch contributed by Paolo Invernizzi!

llvm-svn: 16152
2004-09-03 18:19:51 +00:00
Alkis Evlogimenos 095c3a80f3 Change the way we choose a free register: instead of picking the first
free allocatable register, we prefer the a free one with the most uses
of inactive intervals. This causes less spills and performes a bit
better compared to gcc:

Program                 | GCC/LLC (Before)| GCC/LLC (After)
164.gzip/164.gzip       | 0.59            | 0.60
175.vpr/175.vpr         | 0.57            | 0.58
176.gcc/176.gcc         | 0.59            | 0.61
181.mcf/181.mcf         | 0.94            | 0.95
186.crafty/186.crafty   | 0.62            | 0.62
197.parser/197.parser   | 0.89            | 0.88
252.eon/252.eon         | 0.61            | 0.66
253.perlbmk/253.perlbmk | 0.79            | 0.84
254.gap/254.gap         | 0.81            | 0.81
255.vortex/255.vortex   | 0.92            | 0.93
256.bzip2/256.bzip2     | 0.69            | 0.69
300.twolf/300.twolf     | 0.91            | 0.90

llvm-svn: 16147
2004-09-02 21:23:32 +00:00
Alkis Evlogimenos dcbf893095 We don't need to sort the added vector as unhandled intervals are
stored in a binary heap.

llvm-svn: 16143
2004-09-02 18:00:38 +00:00
Reid Spencer 7c16caa336 Changes For Bug 352
Move include/Config and include/Support into include/llvm/Config,
include/llvm/ADT and include/llvm/Support. From here on out, all LLVM
public header files must be under include/llvm/.

llvm-svn: 16137
2004-09-01 22:55:40 +00:00
Alkis Evlogimenos fae8f6a4fb Be a bit more efficient when processing the active and inactive
lists. Instead of scanning the vector backwards, scan it forward and
swap each element we want to erase. Then at the end erase all removed
intervals at once. This doesn't save much: 0.08s out of 4s when
compiling 176.gcc.

llvm-svn: 16136
2004-09-01 22:52:29 +00:00
Alkis Evlogimenos e2d962e0fd Now that LiveIntervals::addIntervalsForSpills is fixed, do not require
LiveVariables.

llvm-svn: 16076
2004-08-27 19:00:29 +00:00
Alkis Evlogimenos efd40bce75 Back out this change as it broke the build last night. This should be
investicated further as the linearscan variants don't really need
LiveVariables...

llvm-svn: 16074
2004-08-27 18:01:21 +00:00
Alkis Evlogimenos 11f0fc8eda The linear scan variants do not require the LiveVariables analysis.
llvm-svn: 16071
2004-08-27 04:51:13 +00:00
Alkis Evlogimenos a698308cce Convert indentation to 2 spaces.
llvm-svn: 15489
2004-08-04 09:46:26 +00:00
Alkis Evlogimenos 83d9b62b28 Add some comments to the backtracking code.
llvm-svn: 15200
2004-07-25 08:10:33 +00:00
Alkis Evlogimenos cf72e7f854 Change std::map<unsigned, LiveInterval*> into a std::map<unsigned,
LiveInterval>. This saves some space and removes the pointer
indirection caused by following the pointer.

llvm-svn: 15167
2004-07-24 11:44:15 +00:00
Chris Lattner c51866a20e Completely eliminate the intervals_ list. instead, the r2iMap_ maintains
ownership of the intervals.

llvm-svn: 15155
2004-07-24 03:32:06 +00:00
Chris Lattner 856383326a Rename LiveIntervals.(cpp|h) -> LiveIntervalAnalysis.(cpp|h)
llvm-svn: 15135
2004-07-23 17:56:30 +00:00
Alkis Evlogimenos 1a876fab64 Some compile time improvements resulting in a 1sec speedup in the 5sec
compilation of gcc:

* Use vectors instead of lists for the intervals sets
* Use a heap for the unhandled set to keep intervals always sorted and
  makes insertions back to the heap very fast (compared to scanning a
  list)

llvm-svn: 15103
2004-07-22 08:14:44 +00:00
Brian Gaeke 902dcf0729 These files don't need to include <iostream> since they include "Support/Debug.h".
llvm-svn: 15089
2004-07-21 20:50:33 +00:00
Alkis Evlogimenos 12a71ea03e Remove dead code.
llvm-svn: 15011
2004-07-19 23:35:55 +00:00
Chris Lattner 8c8144b958 Fix assertion to not dereference end!
llvm-svn: 14991
2004-07-19 05:02:09 +00:00
Chris Lattner 70f8dca59b Add some asserts that the list of intervals returned by addIntervalsForSpills
is sorted.  This is not the case currently, which is causing no end of
problems.

llvm-svn: 14990
2004-07-19 04:47:36 +00:00
Alkis Evlogimenos 21b3a5b0de Correctly compute the ration of iterations/#intervals.
llvm-svn: 14626
2004-07-04 17:23:35 +00:00
Alkis Evlogimenos 8f3cc0316c Add efficiency statistic.
llvm-svn: 14590
2004-07-04 07:59:06 +00:00
Chris Lattner 4fbff990b0 Rename Interval class to LiveInterval to avoid conflicting with the already
existing llvm::Interval class.

Patch contributed by Vladimir Prus!
http://mail.cs.uiuc.edu/pipermail/llvmbugs/2004-June/000710.html

llvm-svn: 14281
2004-06-21 13:10:56 +00:00
Alkis Evlogimenos fbe799cfa2 Pull Interval class out of LiveIntervals.
llvm-svn: 13910
2004-05-30 07:46:27 +00:00
Alkis Evlogimenos a5268e8199 When spilling an register, introduce a new temporary for each of its
spills. This allows for more flexibility when allocating registers for
spill code.

llvm-svn: 13907
2004-05-30 07:24:39 +00:00
Alkis Evlogimenos 2a54b5d3e1 Add required header
llvm-svn: 13417
2004-05-08 03:50:03 +00:00
Chris Lattner 991425ae70 numeric_limits::infinity() apparently does not work on all systems. As a
workaround, use the C HUGE_VAL macro instead.

llvm-svn: 13377
2004-05-06 16:25:59 +00:00
Alkis Evlogimenos 02a5354d9b Make the set of fixed (preallocated) intervals be a fixed superset of
unhandled + handled. So unhandled is now including all fixed intervals
and fixed intervals never changes when processing a function.

llvm-svn: 12462
2004-03-17 00:48:59 +00:00
Alkis Evlogimenos 31953c7a10 Add a spiller option to llc. A simple spiller will come soon. When we get CFG in the machine code represenation a global spiller will also be possible. Also document the linear scan register allocator but mark it as experimental for now.
llvm-svn: 12062
2004-03-01 23:18:15 +00:00
Alkis Evlogimenos b76d234ee9 Add the long awaited memory operand folding support for linear scan
llvm-svn: 12058
2004-03-01 20:05:10 +00:00
Alkis Evlogimenos e669caafe0 Rename member function to be consistent with the rest.
llvm-svn: 11898
2004-02-27 06:11:15 +00:00
Alkis Evlogimenos a9f03fba9d Remove asssert since it is breaking cases that it shouldn't.
llvm-svn: 11841
2004-02-25 22:01:06 +00:00
Chris Lattner f5a393a133 Add an assertion
llvm-svn: 11830
2004-02-25 19:37:44 +00:00
Alkis Evlogimenos 1dd872ce94 Move machine code rewriter and spiller outside the register
allocator.

The implementation is completely rewritten and now employs several
optimizations not exercised before. For example for 164.gzip we have
997 loads and 699 stores vs the 1221 loads and 880 stores we have
before.

llvm-svn: 11798
2004-02-24 08:58:30 +00:00
Alkis Evlogimenos ab77b05d8c Remove '4Virt' from member function names as it is obvious.
llvm-svn: 11781
2004-02-23 23:47:10 +00:00
Alkis Evlogimenos c794a9060f Refactor VirtRegMap out of RegAllocLinearScan as the first part of bug
251 (providing a generic machine code rewriter/spiller).

llvm-svn: 11780
2004-02-23 23:08:11 +00:00
Alkis Evlogimenos 52de9e214c Add number of spilled registers statistic.
llvm-svn: 11759
2004-02-23 18:45:32 +00:00
Alkis Evlogimenos d780ce6462 Remove implementation of default constructor as it is useless now.
llvm-svn: 11755
2004-02-23 18:28:35 +00:00
Alkis Evlogimenos f83cb74b57 Improved PhysRegTracker interface. RegAlloc lazily allocates the register tracker using a std::auto_ptr
llvm-svn: 11738
2004-02-23 06:10:13 +00:00
Alkis Evlogimenos d42c2f24ec Fix comments in PhysRegTracker and rename isPhysRegAvail to isRegAvail to be consistent with the other two
llvm-svn: 11723
2004-02-23 01:25:05 +00:00
Alkis Evlogimenos 1410859484 Pull PhysRegTracker out of RegAllocLinearScan as it can be used by other allocators as well
llvm-svn: 11720
2004-02-23 00:53:31 +00:00
Alkis Evlogimenos 486b12b71a Move LiveIntervals.h to lib/CodeGen since it shouldn't be exposed to other parts of the compiler
llvm-svn: 11719
2004-02-23 00:50:15 +00:00
Alkis Evlogimenos cb352300c9 Print basic block boundaries in machine instruction debug output.
llvm-svn: 11704
2004-02-22 05:46:04 +00:00
Alkis Evlogimenos 22de21ea80 Fix crash in debug output.
llvm-svn: 11659
2004-02-20 06:41:12 +00:00
Alkis Evlogimenos bcc99a83ee Fix instruction numbering in debug output.
llvm-svn: 11655
2004-02-20 06:29:51 +00:00
Alkis Evlogimenos 76eca062ea Too many changes in one commit:
1. LiveIntervals now implement a 4 slot per instruction model. Load,
   Use, Def and a Store slot. This is required in order to correctly
   represent caller saved register clobbering on function calls,
   register reuse in the same instruction (def resues last use) and
   also spill code added later by the allocator. The previous
   representation (2 slots per instruction) was insufficient and as a
   result was causing subtle bugs.

2. Fixes in spill code generation. This was the major cause of
   failures in the test suite.

3. Linear scan now has core support for folding memory operands. This
   is untested and not enabled (the live interval update function does
   not attempt to fold loads/stores in instructions).

4. Lots of improvements in the debugging output of both live intervals
   and linear scan. Give it a try... it is beautiful :-)

In summary the above fixes all the issues with the recent reserved
register elimination changes and get the allocator very close to the
next big step: folding memory operands.

llvm-svn: 11654
2004-02-20 06:15:40 +00:00
Alkis Evlogimenos 3cf546037e Fix RA::verifyAssignment()
llvm-svn: 11629
2004-02-19 19:24:17 +00:00
Alkis Evlogimenos d0a60b71fc Rename reloads/spills to loads/stores.
llvm-svn: 11619
2004-02-19 06:19:09 +00:00
Alkis Evlogimenos b490a12501 Implement assignment correctness verification.
llvm-svn: 11609
2004-02-18 23:15:23 +00:00
Alkis Evlogimenos 2c5ddd20ba Eliminate the use of spill (reserved) registers.
llvm-svn: 11476
2004-02-15 10:24:21 +00:00
Alkis Evlogimenos ddcb4ad53e Use std::numeric_limits<float>::infinity() instead of
std::numeric_limits<float>::max() for weighting preallocated
intervals.

llvm-svn: 11427
2004-02-14 00:44:07 +00:00
Alkis Evlogimenos 8cdd0215bf Remove getAllocatedRegNum(). Use getReg() instead.
llvm-svn: 11393
2004-02-13 21:01:20 +00:00
Alkis Evlogimenos 80da865f77 Change MachineBasicBlock's vector of MachineInstr pointers into an
ilist of MachineInstr objects. This allows constant time removal and
insertion of MachineInstr instances from anywhere in each
MachineBasicBlock. It also allows for constant time splicing of
MachineInstrs into or out of MachineBasicBlocks.

llvm-svn: 11340
2004-02-12 02:27:10 +00:00
Chris Lattner 5dd5be3ec3 Do not use MachineOperand::isVirtualRegister either!
llvm-svn: 11283
2004-02-10 21:12:22 +00:00
Alkis Evlogimenos e82a70766e Increase code clarity.
llvm-svn: 11151
2004-02-06 18:08:18 +00:00
Alkis Evlogimenos cd2e7ffcc6 Eliminate uneeded lookups by passing a Virt2PhysMap::iterator instead
of the virtual register to certain functions.

llvm-svn: 11143
2004-02-06 03:15:40 +00:00
Alkis Evlogimenos 327426411e Modify the two address instruction pass to remove the duplicate
operand of the instruction and thus simplify the register allocation.

llvm-svn: 11124
2004-02-04 22:17:40 +00:00
Alkis Evlogimenos 3d100ef0ca When an instruction like: A += B had both A and B virtual registers
spilled, A was loaded from its stack location twice. This fixes the bug.

llvm-svn: 11093
2004-02-03 01:13:07 +00:00
Alkis Evlogimenos 6e34205653 Fix debugging output.
llvm-svn: 11088
2004-02-02 22:00:32 +00:00
Alkis Evlogimenos 2cecace9b2 Create an object for tracking physical register usage. This will look
much better when I get rid of the reserved registers.

llvm-svn: 11066
2004-02-02 07:30:36 +00:00
Alkis Evlogimenos 9b9c0485fd Change weight array into a vector and make it as big as the number of
registers (not as the max number of registers).

Change toSpill from a std::set into a std::vector<bool>.

Use the reverse iterator adapter to do a reverse scan of allocatable
registers.

llvm-svn: 11061
2004-02-01 20:13:26 +00:00
Alkis Evlogimenos 5e05a59bf8 Use std::map::count() instead of std::map::find() != std::map::end()
where appropriate.

llvm-svn: 11060
2004-02-01 18:39:53 +00:00
Alkis Evlogimenos 91dcc03c39 Use MRegisterInfo::isPhysicalRegister and
MRegisterInfo::isVirtualRegister.

llvm-svn: 11045
2004-02-01 01:27:01 +00:00
Alkis Evlogimenos 73df113676 Add option to join live intervals. Two intervals are joined if there
is a move between two registers, at least one of the registers is
virtual and the two live intervals do not overlap.

This results in about 40% reduction in intervals, 30% decrease in the
register allocators running time and a 20% increase in peephole
optimizations (mainly move eliminations).

The option can be enabled by passing -join-liveintervals where
appropriate.

llvm-svn: 10965
2004-01-22 23:08:45 +00:00
Alkis Evlogimenos e1c24c06f9 Remove unneeded check. An interval in active, by definition overlaps
with the current one.

llvm-svn: 10959
2004-01-22 20:07:18 +00:00
Alkis Evlogimenos 4716d76d22 Improve debugging output. Remove unneeded virtReg->0 mapping when
virtReg lives on the stack. Now a virtual register has an entry in the
virtual->physical map or the virtual->stack slot map but never in
both.

llvm-svn: 10958
2004-01-22 19:24:43 +00:00
Alkis Evlogimenos d844972574 Handle printing of intervals that are not assign to any physical
register yet (2nd try).

llvm-svn: 10896
2004-01-16 20:33:13 +00:00
Alkis Evlogimenos 8672af1cc8 Handle printing of intervals that are not assign to any physical
register yet.

llvm-svn: 10895
2004-01-16 20:29:42 +00:00
Alkis Evlogimenos d86933d21d Fold open interval ends handling into
LiveIntervals::Interval::expiredAt() and simplify regalloc code.

llvm-svn: 10894
2004-01-16 20:17:05 +00:00
Alkis Evlogimenos a0865cec3f Improve debugging output.
llvm-svn: 10834
2004-01-14 00:09:36 +00:00
Alkis Evlogimenos 65bc990bda Indentation and whitespace cleanups.
llvm-svn: 10820
2004-01-13 20:42:08 +00:00
Alkis Evlogimenos 6b1ec75da9 Fix bug introduced by previous commit: check if fixed intervals
overlap before adding their spill weight.

llvm-svn: 10819
2004-01-13 20:37:01 +00:00
Alkis Evlogimenos ae5b3d4f74 Add a separate list of fixed intervals. This improves the running time
of the register allocator as follows:

       before   after
mesa   2.3790  1.5994
vpr    2.6008  1.2078
gcc    1.9840  0.5273
mcf    0.2569  0.0470
eon    1.8468  1.4359
twolf  0.9475  0.2004
burg   1.6807  1.3300
lambda 1.2191  0.3764

Speedups range anyware from 30% to over 400% :-)

llvm-svn: 10712
2004-01-07 09:20:58 +00:00
Alkis Evlogimenos a64d71203e Minor cleanups.
llvm-svn: 10711
2004-01-07 05:31:12 +00:00
Alkis Evlogimenos 9f62b96206 Remove declared but undefined method.
llvm-svn: 10710
2004-01-07 02:29:33 +00:00
Alkis Evlogimenos 2fa1aef707 Remove simple coalescing.
llvm-svn: 10695
2004-01-05 08:24:57 +00:00
Alkis Evlogimenos a158dc8104 Reserve ECX and EDI instead of EBX and EDI. Since EBX is a callee
saved register it has a longer free range than ECX (which is defined
every time there is a fnuction call) which makes ECX a better register
to reserve.

llvm-svn: 10635
2003-12-28 18:03:52 +00:00
Alkis Evlogimenos 43b587d94d Add coalescing to register allocator. A hint is added to each interval
which denotes the register we would like to be assigned to (virtual or
physical). In register allocation, if this hint exists and we can map
it to a physical register (it is either a physical register or it is a
virtual register that already got assigned to a physical one) we use
that register if it is available instead of a random one in the free
pool.

llvm-svn: 10634
2003-12-28 17:58:18 +00:00
Alkis Evlogimenos 2e8c984c58 Improve debugging output when choosing a register to spill.
llvm-svn: 10604
2003-12-24 18:53:31 +00:00
Alkis Evlogimenos 7d7d7e89d4 Change the way free regusters are computed and perform better
allocation in the presence of preallocated intervals.

llvm-svn: 10595
2003-12-23 18:00:33 +00:00
Alkis Evlogimenos 690e1b8de2 Remove verifyIntervals() since it doesn't actually work right now.
llvm-svn: 10570
2003-12-21 20:41:26 +00:00
Alkis Evlogimenos 2665d9c6f9 Change weight into a float so that we can take into account the
nesting level when computing it. Right now the allocator uses:

    w = sum_over_defs_uses( 10 ^ nesting level );

llvm-svn: 10569
2003-12-21 20:19:10 +00:00
Alkis Evlogimenos c09b77ea35 Add support for inactive intervals. This effectively reuses registers
for live ranges that fall into assigned registers' holes.

llvm-svn: 10566
2003-12-21 05:43:40 +00:00
Chris Lattner 8f88cdfa22 Prune some #includes
Add a statistic for # reloads

llvm-svn: 10518
2003-12-18 20:25:31 +00:00
Alkis Evlogimenos e5e2bd6777 Modify linear scan register allocator to use the two-address
instruction pass. This also fixes all remaining bugs for this new
allocator to pass all tests under test/Programs.

llvm-svn: 10515
2003-12-18 13:15:02 +00:00
Alkis Evlogimenos 3bd69eaac5 Fix bug in reserved registers. DH actually aliases DX and EDX which
are not reserved registers.

llvm-svn: 10514
2003-12-18 13:12:18 +00:00
Alkis Evlogimenos 4d87219486 Rename LiveIntervals::expired() to LiveIntervals::expiredAt().
llvm-svn: 10511
2003-12-18 08:56:11 +00:00
Alkis Evlogimenos aaba4639f8 Change interface of MachineOperand as follows:
a) remove opIsUse(), opIsDefOnly(), opIsDefAndUse()
    b) add isUse(), isDef()
    c) rename opHiBits32() to isHiBits32(),
              opLoBits32() to isLoBits32(),
              opHiBits64() to isHiBits64(),
              opLoBits64() to isLoBits64().

This results to much more readable code, for example compare
"op.opIsDef() || op.opIsDefAndUse()" to "op.isDef()" a pattern used
very often in the code.

llvm-svn: 10461
2003-12-14 13:24:17 +00:00
Alkis Evlogimenos 4f7f017ab5 When reserving a preallocated register spill the aliases of this
register too.

llvm-svn: 10450
2003-12-13 11:58:10 +00:00
Alkis Evlogimenos ed53b35076 Expire any active intervals left when register allocation is done.
llvm-svn: 10448
2003-12-13 05:50:19 +00:00
Alkis Evlogimenos e6dc614c3c Add instruction numbers to debugging output.
llvm-svn: 10447
2003-12-13 05:48:57 +00:00
Alkis Evlogimenos f8fe176204 Fix bug in register spilling when a preallocated live range overlaps a
potential register assignment.

llvm-svn: 10291
2003-12-05 11:17:55 +00:00
Alkis Evlogimenos 5fe0031c68 Improve debugging output and clean up some code.
llvm-svn: 10288
2003-12-04 03:57:28 +00:00
Alkis Evlogimenos 26f17e73eb Print instructions before register allocation is performed. Also fix
bug where spill instructions were added to the next basic block
instead of the end of the current one if the instruction that required
the spill was the last in the block.

llvm-svn: 10272
2003-11-30 23:40:39 +00:00
Alkis Evlogimenos 0aa40f8db0 Remove "numReloaded" statistic.
llvm-svn: 10268
2003-11-30 05:15:36 +00:00
Alkis Evlogimenos 0e9ded74c0 Merging the linear scan register allocator in trunk. It currently passes most tests under test/Programs/SingleSource/Benchmarks/Shootout so development will continue on trunk. The allocator is not enabled by default. You will need to pass -regallo=linearscan to lli or llc to use it.
llvm-svn: 10103
2003-11-20 03:32:25 +00:00