Legalize drops conditions and selects

[llvmbot]

Bugzilla Link	652
Resolution	FIXED
Resolved on	Feb 22, 2010 12:49
Version	trunk
OS	Linux
Reporter	LLVM Bugzilla Contributor

Extended Description

Legalize makes the last block mostly disappear, and that is bad.

; ModuleID = 'bugpoint.test.bc'
target endian = little
target pointersize = 64

implementation ; Functions:

bool %l5_l14_g721_encoder_then_2E_i(int %tmp.46.reload, short* %tmp.5542.out) {
newFuncRoot:
br label %then.i

cond_true.exitStub: ; preds = %then.i
store short %tmp.5542, short* %tmp.5542.out
ret bool true

cond_false.exitStub: ; preds = %then.i
store short %tmp.5542, short* %tmp.5542.out
ret bool false

then.i: ; preds = %newFuncRoot
%tmp.11.i = setne int %tmp.46.reload, 0 ; [#uses=1]
%tmp.5542 = select bool %tmp.11.i, short -32768, short 0
; [#uses=3]
%tmp.5744 = setlt short %tmp.5542, 0 ; [#uses=1]
br bool %tmp.5744, label %cond_true.exitStub, label %cond_false.exitStub
}

Gets codegened into:
newFuncRoot:
%$31 = BR mbb<then.i,0x120f11190>
cond_true.exitStub:
STW %$0, 0, %$17
%$0 = LDA 1, %$31
%$31 = RET %$26, 1
cond_false.exitStub:
STW %$0, 0, %$17
%$0 = LDA 0, %$31
%$31 = RET %$26, 1
then.i:
%$0 = LDA 0, %$31
%$31 = BR mbb<cond_false.exitStub,0x120f110d0>

then.i is clearly in bad shape.

[lattner]

Hrm, I get:

$LBBl5_l14_g721_encoder_then_2E_i_3: # then.i
zapnot $16,15,$0
lda $1,0($31)
cmpeq $0,$1,$0
cmpeq $31,$0,$0
sll $0,15,$0
sextw $0,$1
blt $1,$LBBl5_l14_g721_encoder_then_2E_i_1 # cond_true.exitStub
br $31,$LBBl5_l14_g721_encoder_then_2E_i_2 # cond_false.exitStub
.end l5_l14_g721_encoder_then_2E_i

When I just use LLC. Do I need some special flags?

-Chris

[lattner]

note, it's also possible that legalize is folding things in a different order for you than for me. Can you try
to figure out what part of the code is doing the bad xform? Is it the dag combiner for example?

-Chris

[llvmbot]

here is some relivent debug output, still tracking down the line that does this.

Replacing 0x120f14010: ch = TokenFactor 0x120f13fb0, 0x120f122d0
With: 0x120f13fb0: ch = CopyToReg 0x120f122d0, 0x120f13f40, 0x120f13ee0

Replacing 0x120f13bd0: i16 = select 0x120f13b70, 0x120f13d30, 0x120f13da0
With: 0x120f14430: i16 = shl 0x120f14360, 0x120f143c0

Replacing 0x120f14430: i16 = shl 0x120f14360, 0x120f143c0
With: 0x120f13da0: i16 = Constant <0>

Replacing 0x120f13e80: i1 = setcc 0x120f13da0, 0x120f13da0, 0x120f13e10
With: 0x120f14430: i1 = Constant <0>

Replacing 0x120f14260: ch = brcondtwoway 0x120f13fb0, 0x120f14430, 0x120f14070,
0x120f141f0
With: 0x120f13bd0: ch = br 0x120f13fb0, 0x120f141f0

[llvmbot]

DAGCombiner.cpp:2172 is causing this.

[llvmbot]

wait, how did that constant get positive? DAG shows it is selected to
0x120f13d30: i16 = Constant <32768>
before any dagcombine.

[llvmbot]

Fixed. Thanks Chris.

http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20051031/029214.html