1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-30 07:22:55 +01:00
Commit Graph

37 Commits

Author SHA1 Message Date
Chris Lattner
4946ae54df Don't transform in another bad case: if the block is empty, it should be
simplified before we do this xform so that our cost model is accurate.

llvm-svn: 31864
2006-11-18 22:25:39 +00:00
Chris Lattner
b101c06bce Fix another case we *don't* want to do this xform.
llvm-svn: 31861
2006-11-18 21:56:39 +00:00
Chris Lattner
514fc3de5c make the previous change more aggressive, moving any block with no fallthrough.
This speeds up yacr2 by 7% on a core2.

llvm-svn: 31856
2006-11-18 21:30:35 +00:00
Chris Lattner
bddc114cfb Minor code layout tweak: if we have something like this:
if (cond) goto BB2
BB1:
    ...
    return;
BB2:
   ...

Move BB1 to the end of the function so that the code falls through in the
non-return case.  This has the effect of moving assert (and other no-return
call) bodies and return blocks out of loops.

llvm-svn: 31855
2006-11-18 20:47:54 +00:00
Chris Lattner
f71921c79d optimize single MBB loops better. In particular, produce:
LBB1_57:        #bb207.i
        movl 72(%esp), %ecx
        movb (%ecx,%eax), %cl
        movl 80(%esp), %edx
        movb %cl, 1(%edx,%eax)
        incl %eax
        cmpl $143, %eax
        jne LBB1_57     #bb207.i
        jmp LBB1_64     #cond_next255.i

intead of:

LBB1_57:        #bb207.i
        movl 72(%esp), %ecx
        movb (%ecx,%eax), %cl
        movl 80(%esp), %edx
        movb %cl, 1(%edx,%eax)
        incl %eax
        cmpl $143, %eax
        je LBB1_64      #cond_next255.i
        jmp LBB1_57     #bb207.i

This eliminates a branch per iteration of the loop.  This hurted PPC
particularly, because the extra branch meant another dispatch group for each
iteration of the loop.

llvm-svn: 31530
2006-11-08 01:03:21 +00:00
Jeff Cohen
e1003da1a2 Unbreak VC++ build.
llvm-svn: 31464
2006-11-05 19:31:28 +00:00
Reid Spencer
4bafa71dc1 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
f20968bc33 give branch folding a simple heuristic to decide which block to split so that
it inserts an uncond branch where it is less likely to cause a problem.  This
fixes some perf issues on ppc.

llvm-svn: 31354
2006-11-01 19:36:29 +00:00
Chris Lattner
b459c08384 make tail merging more aggressive. If two blocks share a common tail, but the
tail is not an entire block for either of them, pick one, split it, then
merge the common part.

llvm-svn: 31336
2006-11-01 01:16:12 +00:00
Chris Lattner
5651528580 enable branch folding with an option
llvm-svn: 31335
2006-11-01 00:38:31 +00:00
Chris Lattner
e1b2b2127b Make CanFallThrough more intelligent (so it can handle blocks with (e.g.) no
successors), and make island block movement more general.

This compiles CodeGen/X86/2006-04-27-ISelFoldingBug.ll to:

_loadAndRLEsource_no_exit_2E_1_label_2E_0:
        subl $8, %esp
        movl %esi, 4(%esp)
        movl %ebx, (%esp)
        movl 16(%esp), %eax
        movl 12(%esp), %ecx
LBB1_3: #label.0
        movl _last, %edx
        movl %edx, %esi
        incl %esi
        movl %esi, _last
        movl %ecx, %ebx
        # TRUNCATE movb %bl, %bl
        movl _block, %esi
        movb %bl, 1(%esi,%edx)
        cmpl %eax, _last
        jge LBB1_2      #codeRepl5.exitStub
LBB1_4: #label.0
        cmpl $257, %ecx
        je LBB1_2       #codeRepl5.exitStub
LBB1_1: #label.0.no_exit.1_crit_edge.exitStub
        movl $1, %eax
        movl (%esp), %ebx
        movl 4(%esp), %esi
        addl $8, %esp
        ret
LBB1_2: #codeRepl5.exitStub
        xorl %eax, %eax
        movl (%esp), %ebx
        movl 4(%esp), %esi
        addl $8, %esp
        ret

instead of:

_loadAndRLEsource_no_exit_2E_1_label_2E_0:
        subl $8, %esp
        movl %esi, 4(%esp)
        movl %ebx, (%esp)
        movl 16(%esp), %eax
        movl 12(%esp), %ecx
        jmp LBB1_3      #label.0
LBB1_1: #label.0.no_exit.1_crit_edge.exitStub
        movl $1, %eax
        movl (%esp), %ebx
        movl 4(%esp), %esi
        addl $8, %esp
        ret
LBB1_2: #codeRepl5.exitStub
        xorl %eax, %eax
        movl (%esp), %ebx
        movl 4(%esp), %esi
        addl $8, %esp
        ret
LBB1_3: #label.0
        movl _last, %edx
        movl %edx, %esi
        incl %esi
        movl %esi, _last
        movl %ecx, %ebx
        # TRUNCATE movb %bl, %bl
        movl _block, %esi
        movb %bl, 1(%esi,%edx)
        cmpl %eax, _last
        jge LBB1_2      #codeRepl5.exitStub
LBB1_4: #label.0
        cmpl $257, %ecx
        jne LBB1_1      #label.0.no_exit.1_crit_edge.exitStub
        jmp LBB1_2      #codeRepl5.exitStub

... which is much better layout :)

llvm-svn: 31282
2006-10-29 21:05:41 +00:00
Chris Lattner
95eed3bd09 Teach branch folding to fold identical jump tables together and to delete
jump tables that are dead.

llvm-svn: 31273
2006-10-28 18:34:47 +00:00
Chris Lattner
af9b4c33c2 improve deletion of blocks that just contain branches by knowing that
the pred block doesn't fall through into them if it's a jumptable.

llvm-svn: 31263
2006-10-28 17:32:47 +00:00
Chris Lattner
2258020df2 simplify code
llvm-svn: 31188
2006-10-25 22:21:37 +00:00
Chris Lattner
df59442a74 turn off tail merging for now
llvm-svn: 31180
2006-10-25 18:08:50 +00:00
Chris Lattner
533b205309 move single basic blocks that are neither fallen into nor fall out of into
a place more useful.  In particular, if we can put them in a place where code
will be able to fall into it, do so.  Otherwise, put it in a place it can fall
through into a successor.  Otherwise, if preventing a fallthrough, move to the
end of the function, out of the way.

This deletes several hundred unconditional branches from spass.

llvm-svn: 31149
2006-10-24 01:12:32 +00:00
Chris Lattner
0c3fabc9a4 Enable tail merging by default.
llvm-svn: 31140
2006-10-23 22:10:12 +00:00
Jim Laskey
f6a28d7161 More complete solution to deleting blocks and debug info.
llvm-svn: 31129
2006-10-23 14:56:37 +00:00
Chris Lattner
9e7d74961b don't break infinite loops
llvm-svn: 31102
2006-10-21 06:11:43 +00:00
Chris Lattner
210d322b1f Use branch reversal to do stuff like this:
call L_strcmp$stub
        testl %eax, %eax
-       jne LBB26_208   #cond_true6020
-       jmp LBB26_227   #bb7119
+       je LBB26_227    #bb7119
 LBB26_208:     #cond_true6020
        movl $l31_str14, 4(%esp)

        testl %eax, %eax
-       jne LBB26_704   #cond_true13042
-       jmp LBB26_713   #bb13151
+       je LBB26_713    #bb13151
 LBB26_704:     #cond_true13042
        movl $_str52, 4(%esp)

        cmpl 76(%ecx), %eax
-       jge LBB26_1628  #cond_false63.i.i
-       jmp LBB26_1769  #_Z8makeGridP13mrSurfaceListidiidd.exit.i
+       jl LBB26_1769   #_Z8makeGridP13mrSurfaceListidiidd.exit.i
 LBB26_1628:    #cond_false63.i.i
        movl $0, 48964(%esp)

llvm-svn: 31100
2006-10-21 05:54:00 +00:00
Chris Lattner
cf7f500cfd Transform code like:
jle FOO
  jmp BAR
BAR:

into:

  jle FOO
BAR:

... whoa!

llvm-svn: 31098
2006-10-21 05:43:30 +00:00
Chris Lattner
d4cb8dc857 Three changes:
1. Remove a bunch of ifdef'd code.
2. When a block just contains an uncond branch, change all blocks branching
   to it to jump to the destination instead.
3. If branch analysis tells us some edges in the machinecfg are not actually
   possible, remove them.

#2 triggers a suprisingly large number of times.

llvm-svn: 31094
2006-10-21 05:08:28 +00:00
Chris Lattner
52a4558cf2 Add an experimental cross-jumping implementation.
This is currently disabled by default and limited in several ways, but does
have a positive effect.

llvm-svn: 31090
2006-10-21 00:47:49 +00:00
Chris Lattner
cd00328fad Teach the branch folder to update debug info if it removes blocks with line
# notes in it.

llvm-svn: 31026
2006-10-17 23:17:27 +00:00
Chris Lattner
ef7c2e1dc8 Enable deleting branches to successor blocks. With the previous patches,
branch folding can now compile stuff like this:

void foo(int W, int X, int Y, int Z) {
  if (W & 1) {
    for (; X;--X) bar();
  } else if (W & 2) {
    for (; Y;--Y) bar();
  } else if (W & 4) {
    for (; Z;--Z) bar();
  } else if (W & 8) {
    for (; W;--W) bar();
  }

  if (W) {
    bar();
  }
}

contrived testcase where loops exits all end up merging together.  To have
the loop merges be:

...
        cmplw cr0, r30, r27
        bne cr0, LBB1_14        ;bb38
LBB1_16:        ;cond_next48.loopexit
        mr r27, r29
LBB1_20:        ;cond_next48
        cmplwi cr0, r27, 0
        beq cr0, LBB1_22        ;UnifiedReturnBlock
...

instead of:


...
        cmplw cr0, r30, r27
        bne cr0, LBB1_14        ;bb38
LBB1_16:        ;cond_next48.loopexit
        mr r27, r29
        b LBB1_20       ;cond_next48
LBB1_17:        ;cond_next48.loopexit1
        b LBB1_20       ;cond_next48
LBB1_18:        ;cond_next48.loopexit2
        b LBB1_20       ;cond_next48
LBB1_19:        ;cond_next48.loopexit3
LBB1_20:        ;cond_next48
        cmplwi cr0, r27, 0
        beq cr0, LBB1_22        ;UnifiedReturnBlock
...


This is CodeGen/PowerPC/branch-opt.ll

llvm-svn: 31006
2006-10-17 18:16:40 +00:00
Chris Lattner
8464c7e3e1 Reenable this pass, fixing the bugs in it.
It now correctly deletes unreachable blocks and blocks that are empty.

llvm-svn: 31000
2006-10-17 17:13:52 +00:00
Chris Lattner
c4be6ee8b8 disable this pass for now, it's causing issues
llvm-svn: 30951
2006-10-14 00:30:06 +00:00
Chris Lattner
1c35d27fc4 falling off the end of a function is ok with an unreachable instruction.
llvm-svn: 30950
2006-10-14 00:21:48 +00:00
Chris Lattner
ae19ea4b2a disable some objectionable code, maybe we can bring this pass to life
llvm-svn: 30939
2006-10-13 20:43:10 +00:00
Misha Brukman
774e55c446 Remove trailing whitespace
llvm-svn: 21420
2005-04-21 22:36:52 +00:00
Reid Spencer
c4abcbefb1 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
Chris Lattner
786c788d12 Dereferencing end() is bad.
llvm-svn: 15402
2004-08-01 09:51:42 +00:00
Alkis Evlogimenos
1dd3ea5515 Make OptimizeBlock take a MachineFunction::iterator instead of a
MachineBasicBlock* as a parameter so that nxext() and prior() helper
functions can work naturally on it.

llvm-svn: 15376
2004-07-31 19:24:41 +00:00
Chris Lattner
ecca71ce92 Next on a pointer increments the pointer, not an iterator
llvm-svn: 15375
2004-07-31 18:40:36 +00:00
Alkis Evlogimenos
b28bc78d0a Use next() helper to make code more readable. Use
MachineFunction::iterator instead of MachineBasicBlock* to avoid
dereferencing end iterators.

llvm-svn: 15373
2004-07-31 15:14:29 +00:00
Alkis Evlogimenos
143aeb646c Use MachineFunction::iterator instead of a MachineBasicBlock* because
FallThrough maybe == to MF.end().

llvm-svn: 15372
2004-07-31 15:03:52 +00:00
Chris Lattner
7b12b8a13e Implement a simple target-independent CFG cleanup pass
llvm-svn: 15368
2004-07-31 10:01:27 +00:00