167 Commits

Author SHA1 Message Date
Evan Cheng
e742ee1dbe Simplify my (il)logic.
llvm-svn: 43819
2007-11-07 08:08:25 +00:00
Evan Cheng
dd71a5c37b When the allocator rewrite a spill register with new virtual register, it replaces other operands of the same register. Watch out for situations where
only some of the operands are sub-register uses.

llvm-svn: 43776
2007-11-06 21:12:10 +00:00
Evan Cheng
92d23e5204 Fix a bug where a def use operand isn't being detected as a sub-register use.
llvm-svn: 43763
2007-11-06 08:50:44 +00:00
Evan Cheng
a8044084ac Fix PR1187.
llvm-svn: 43692
2007-11-05 00:59:10 +00:00
Evan Cheng
66298e226f There are times when the coalescer would not coalesce away a copy but the copy
can be eliminated by the allocator is the destination and source targets the
same register. The most common case is when the source and destination registers
are in different class. For example, on x86 mov32to32_ targets GR32_ which
contains a subset of the registers in GR32.

The allocator can do 2 things:
1. Set the preferred allocation for the destination of a copy to that of its source.
2. After allocation is done, change the allocation of a copy destination (if
   legal) so the copy can be eliminated.

This eliminates 443 extra moves from 403.gcc.

llvm-svn: 43662
2007-11-03 07:20:12 +00:00
Evan Cheng
0dde6e5761 Apply Chris' suggestions.
llvm-svn: 43069
2007-10-17 06:53:44 +00:00
Evan Cheng
8b8c7c9927 Clean up code that calculate MBB live-in's.
llvm-svn: 43060
2007-10-17 02:10:22 +00:00
Evan Cheng
1410b8512c Did mean to leave this in. INSERT_SUBREG isn't being coalesced yet.
llvm-svn: 42916
2007-10-12 17:16:50 +00:00
Evan Cheng
aa2d6ef81d EXTRACT_SUBREG coalescing support. The coalescer now treats EXTRACT_SUBREG like
(almost) a register copy. However, it always coalesced to the register of the
RHS (the super-register). All uses of the result of a EXTRACT_SUBREG are sub-
register uses which adds subtle complications to load folding, spiller rewrite,
etc.

llvm-svn: 42899
2007-10-12 08:50:34 +00:00
Evan Cheng
21a58a72c5 Kill cycle of an live range is always the last use index + 1.
llvm-svn: 42742
2007-10-08 06:59:30 +00:00
Dan Gohman
c731c97fac Use empty() member functions when that's what's being tested for instead
of comparing begin() and end().

llvm-svn: 42585
2007-10-03 19:26:29 +00:00
Dan Gohman
9da02f5ee2 Remove isReg, isImm, and isMBB, and change all their users to use
isRegister, isImmediate, and isMachineBasicBlock, which are equivalent,
and more popular.

llvm-svn: 41958
2007-09-14 20:33:02 +00:00
Evan Cheng
d059eed1c1 Fix a memory leak.
llvm-svn: 41739
2007-09-06 01:07:24 +00:00
Evan Cheng
db53aef53e Use pool allocator for all the VNInfo's to improve memory access locality. This reduces coalescing time on siod Mac OS X PPC by 35%. Also remove the back ptr from VNInfo to LiveInterval and other tweaks.
llvm-svn: 41729
2007-09-05 21:46:51 +00:00
Evan Cheng
32a0a995c6 Try fold re-materialized load instructions into its uses.
llvm-svn: 41598
2007-08-30 05:53:02 +00:00
Evan Cheng
1ad4a6117b Change LiveRange so it keeps a pointer to the VNInfo rather than an index.
Changes related modules so VNInfo's are not copied. This decrease
copy coalescing time by 45% and overall compilation time by 10% on siod.

llvm-svn: 41579
2007-08-29 20:45:00 +00:00
Evan Cheng
70c2de7bf1 Fix some kill info update bugs; add hidden option -disable-rematerialization to turn off remat for debugging.
llvm-svn: 41118
2007-08-16 07:24:22 +00:00
Evan Cheng
33820da1da Re-implement trivial rematerialization. This allows def MIs whose live intervals that are coalesced to be rematerialized.
llvm-svn: 41060
2007-08-13 23:45:17 +00:00
Evan Cheng
05cc486c7b Code to maintain kill information during register coalescing.
llvm-svn: 41016
2007-08-11 00:59:19 +00:00
Evan Cheng
d771b793fe Adding kill info to val#.
llvm-svn: 40925
2007-08-08 07:03:29 +00:00
Evan Cheng
a8c2f38617 - Each val# can have multiple kills.
- Fix some minor bugs related to special markers on val# def. ~0U means
  undefined, ~1U means dead val#.

llvm-svn: 40916
2007-08-08 03:00:28 +00:00
Evan Cheng
0d0fee269a - LiveInterval value#'s now have 3 components: def instruction #,
kill instruction #, and source register number (iff the value# is defined by a
copy).
- Now def instruction # is set for every value#, not just for copy defined ones.
- Update some outdated code related inactive live ranges.
- Kill info not yet set. That's next patch.

llvm-svn: 40913
2007-08-07 23:49:57 +00:00
Evan Cheng
df0c705d7d If a livein is not used in the block. It's live through.
llvm-svn: 37764
2007-06-27 18:47:28 +00:00
Evan Cheng
6cf1371456 Fix an obvious bug. Old code only worked for the entry block.
llvm-svn: 37743
2007-06-27 01:16:36 +00:00
Dan Gohman
9e82064924 Replace M_REMATERIALIZIBLE and the newly-added isOtherReMaterializableLoad
with a general target hook to identify rematerializable instructions. Some
instructions are only rematerializable with specific operands, such as loads
from constant pools, while others are always rematerializable. This hook
allows both to be identified as being rematerializable with the same
mechanism.

llvm-svn: 37644
2007-06-19 01:48:05 +00:00
Dan Gohman
4a4a8eb00e Add a target hook to allow loads from constant pools to be rematerialized, and an
implementation for x86.

llvm-svn: 37576
2007-06-14 20:50:44 +00:00
David Greene
02f6e9b621 Factor live variable analysis so it does not do register coalescing
simultaneously.  Move that pass to SimpleRegisterCoalescing.

This makes it easier to implement alternative register allocation and
coalescing strategies while maintaining reuse of the existing live
interval analysis.

llvm-svn: 37520
2007-06-08 17:18:56 +00:00
Evan Cheng
e1595b6859 Only worry about intervening kill if there are more than one live ranges in the interval.
llvm-svn: 37052
2007-05-14 21:23:51 +00:00
Evan Cheng
c690cba7d9 Fix for PR1406:
v1 =
r2 = move v1
   = op r2<kill>
...
r2 = move v1
   = op r2<kill>

Clear the first r2 kill if v1 and r2 are joined.

llvm-svn: 37050
2007-05-14 21:10:05 +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
Lauro Ramos Venancio
41223586a2 Fix build error.
llvm-svn: 36648
2007-05-02 20:37:47 +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
Evan Cheng
910c80851e Rename findRegisterUseOperand to findRegisterUseOperandIdx to avoid confusion.
llvm-svn: 36483
2007-04-26 19:00:32 +00:00
Evan Cheng
7818c03c6b Fix for PR1306.
- A register def / use now implicitly affects sub-register liveness but does
not affect liveness information of super-registers.
- Def of a larger register (if followed by a use later) is treated as
read/mod/write of a smaller register.

llvm-svn: 36434
2007-04-25 07:30:23 +00:00
Evan Cheng
4c53d321aa VarInfo::UsedBlocks is no longer used. Remove.
llvm-svn: 36250
2007-04-18 05:04:38 +00:00
Evan Cheng
b408e8f11d Don't populate TryAgainList when coalescing only physical registers with virtual registers.
llvm-svn: 36240
2007-04-18 02:30:19 +00:00
Evan Cheng
b881bdabd9 Copy coalescing change to prevent a physical register from being pin to a
long live interval that has low usage density.
1. Change order of coalescing to join physical registers with virtual
   registers first before virtual register intervals become too long.
2. Check size and usage density to determine if it's worthwhile to join.
3. If joining is aborted, assign virtual register live interval allocation
   preference field to the physical register.
4. Register allocator should try to allocate to the preferred register
   first (if available) to create identify moves that can be eliminated.

llvm-svn: 36218
2007-04-17 20:32:26 +00:00
Evan Cheng
8be98c1572 Re-materialize all loads from fixed stack slots.
llvm-svn: 35660
2007-04-04 07:40:01 +00:00
Evan Cheng
2bf2aadd9a Trivially re-materializable instructions have spill weights that are half of what it would be otherwise.
llvm-svn: 35658
2007-04-04 07:04:55 +00:00
Evan Cheng
476fb6a5c9 Ugh. Copy coalescer does not update register numbers.
llvm-svn: 35600
2007-04-02 18:49:18 +00:00
Evan Cheng
48dd5e7d44 Bug fix for PR1279. When isDead is propagate by copy coalescing, we keep length
of dead def live interval at 1 to avoid multiple def's targeting the same
register. The previous patch missed a case where the source operand is live-in.
In that case, remove the whole interval.

llvm-svn: 35512
2007-03-30 20:18:35 +00:00
Evan Cheng
603e49c6da Move rematerialization out of beta.
llvm-svn: 35419
2007-03-28 08:26:40 +00:00
Evan Cheng
b41697c006 Fix for PR1279. Dead def has a live interval of length 1. Copy coalescing should
not violate that.

llvm-svn: 35396
2007-03-28 01:30:37 +00:00
Evan Cheng
4a09b1b5be Fix for PR1266. Don't mark a two address operand IsKill.
llvm-svn: 35365
2007-03-26 22:40:42 +00:00
Evan Cheng
a20815e157 Fix for PR1257. Bug in live range shortening as a result of copy coalescing
where the destination is dead.

llvm-svn: 35252
2007-03-22 01:26:05 +00:00
Evan Cheng
0e3278e505 First cut trivial re-materialization support.
llvm-svn: 35208
2007-03-20 08:13:50 +00:00
Evan Cheng
25d00d545d Remove -reduce-joining-phys-regs options. Make it on by default.
llvm-svn: 35165
2007-03-19 18:08:26 +00:00
Evan Cheng
c5e74f6404 Minor bug fix.
llvm-svn: 35153
2007-03-19 04:22:35 +00:00
Evan Cheng
d8f2e4fe4c - Merge UsedBlocks info after two virtual registers are coalesced.
- Use distance to closest use to determine whether to abort coalescing.

llvm-svn: 35141
2007-03-18 09:05:55 +00:00