[Bug tree-optimization/18576] missing jump threading because of type changes

[steven at gcc dot gnu dot org]

--- Additional Comments From steven at gcc dot gnu dot org  2004-11-20 
16:44 ---
 
Confirmed, I see similar code on x86: 
 
.L4: 
xorl%ebx, %ebx 
.L7: 
xorl%edx, %edx 
testb   %bl, %bl 
jne .L2 
 
 

-- 
   What|Removed |Added

 Status|UNCONFIRMED |NEW
 Ever Confirmed||1
   Last reconfirmed|-00-00 00:00:00 |2004-11-20 16:44:20
   date||


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18576

[Bug tree-optimization/18576] missing jump threading because of type changes

[steven at gcc dot gnu dot org]

--- Additional Comments From steven at gcc dot gnu dot org  2004-11-20 
17:01 ---
I have the following in the .optimized dump: 
 
flow_bb_inside_loop_p (loop, source_loop) 
{ 
  unsigned char D.1171; 
  struct loop * D.1170; 
  struct loop * * D.1169; 
  struct loop * * D.1168; 
  unsigned int D.1167; 
  unsigned int D.1166; 
  struct loop * * D.1165; 
  int D.1164; 
  int D.1163; 
  int iftmp.0; 
  int D.1161; 
  struct loop * outer; 
  struct loop * loop; 
  unsigned char D.1146; 
  unsigned char D.1145; 
  int iftmp.1; 
  int D.1140; 
  int D.1160; 
 
  # BLOCK 0 
  # PRED: ENTRY [100.0%]  (fallthru,exec) 
  if (loop == source_loop) goto L8; else goto L0; 
  # SUCC: 6 [10.4%]  (true,exec) 1 [89.6%]  (false,exec) 
 
  # BLOCK 1 
  # PRED: 0 [89.6%]  (false,exec) 
L0:; 
  D.1164 = loop-depth; 
  if (source_loop-depth = D.1164) goto L4; else goto L1; 
  # SUCC: 4 [50.0%]  (true,exec) 2 [50.0%]  (false,exec) 
 
  # BLOCK 2 
  # PRED: 1 [50.0%]  (false,exec) 
L1:; 
  if (loop != *(source_loop-pred + (struct loop * *) ((unsigned int) D.1164 * 
4))) goto L4; else goto L2; 
  # SUCC: 4 [81.0%]  (true,exec) 3 [19.0%]  (false,exec) 
 
  # BLOCK 3 
  # PRED: 2 [19.0%]  (false,exec) 
L2:; 
  iftmp.0 = 1; 
  goto bb 7 (L14); 
  # SUCC: 7 [100.0%]  (fallthru,exec) 
 
  # BLOCK 4 
  # PRED: 1 [50.0%]  (true,exec) 2 [81.0%]  (true,exec) 
L4:; 
  iftmp.0 = 0; 
  # SUCC: 7 [100.0%]  (fallthru) 
 
  # BLOCK 7 
  # PRED: 4 [100.0%]  (fallthru) 3 [100.0%]  (fallthru,exec) 
L14:; 
  if ((unsigned char) (int) (unsigned char) iftmp.0 != 0) goto L8; else goto 
L7; 
  # SUCC: 6 [33.0%]  (true,exec) 5 [67.0%]  (false,exec) 
 
  # BLOCK 5 
  # PRED: 7 [67.0%]  (false,exec) 
L7:; 
  iftmp.1 = 0; 
  goto bb 8 (L15); 
  # SUCC: 8 [100.0%]  (fallthru,exec) 
 
  # BLOCK 6 
  # PRED: 0 [10.4%]  (true,exec) 7 [33.0%]  (true,exec) 
L8:; 
  iftmp.1 = 1; 
  # SUCC: 8 [100.0%]  (fallthru) 
 
  # BLOCK 8 
  # PRED: 6 [100.0%]  (fallthru) 5 [100.0%]  (fallthru,exec) 
L15:; 
  return (int) (unsigned char) iftmp.1; 
  # SUCC: EXIT [100.0%] 
} 
 
Control flow graph: 
 
ENTRY 
| 
0 
|\ 
| \ 
1  \ 
|\  \ 
| \  \ 
2  \  \ 
|\ |  | 
| \|  | 
3  4  | 
| /   | 
|/| 
7 / 
|\   / 
| \ / 
5  6 
| / 
|/ 
8 
| 
EXIT 
 
 
 
Dominator tree: 
 
  0 
 /|\ 
6 8 1 
   /|\ 
  4 7 2 
| | 
5 3 
 
 
The problematic part is here: 
 
  # BLOCK 3 
  # PRED: 2 [19.0%]  (false,exec) 
L2:; 
  iftmp.0 = 1; 
  goto bb 7 (L14); 
  # SUCC: 7 [100.0%]  (fallthru,exec) 
 
  # BLOCK 4 
  # PRED: 1 [50.0%]  (true,exec) 2 [81.0%]  (true,exec) 
L4:; 
  iftmp.0 = 0; 
  # SUCC: 7 [100.0%]  (fallthru) 
 
  # BLOCK 7 
  # PRED: 4 [100.0%]  (fallthru) 3 [100.0%]  (fallthru,exec) 
L14:; 
  if ((unsigned char) (int) (unsigned char) iftmp.0 != 0) goto L8; else goto 
L7; 
  # SUCC: 6 [33.0%]  (true,exec) 5 [67.0%]  (false,exec) 
 
 
Possible causes of the missed jump thread: 
1) Are casts getting in the way? 
2) Do we need that one of edges into the condition we want to 
   thread through dominates the block? 
 
 

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18576

[Bug tree-optimization/18576] missing jump threading because of type changes

[pinskia at gcc dot gnu dot org]

--- Additional Comments From pinskia at gcc dot gnu dot org  2004-11-20 
17:11 ---
Note changing flow_loop_nested_p to return an int instead of unsigned char, the 
jump threading works 

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18576

[Bug tree-optimization/18576] missing jump threading because of type changes

[steven at gcc dot gnu dot org]

--- Additional Comments From steven at gcc dot gnu dot org  2004-11-20 
17:11 ---
Situation in DOM3:  
  
  # BLOCK 4  
  # PRED: 3 [100.0%]  (fallthru,exec) 2 [81.0%]  (true,exec) 1 [50.0%]   
(true,exec)  
  # iftmp.0_1 = PHI 0(2), 1(3), 0(1);  
L4:;  
  D.1171_13 = (unsigned char) iftmp.0_1;  
  D.1160_14 = (int) D.1171_13;  
  D.1145_16 = (unsigned char) D.1160_14;  
  if (D.1145_16 != 0) goto L8; else goto L7;  
  # SUCC: 6 [33.0%]  (true,exec) 5 [67.0%]  (false,exec)  
  
So this is almost certainly casts again getting in the way.  
  

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18576

[Bug tree-optimization/18576] missing jump threading because of type changes

[pinskia at gcc dot gnu dot org]

--- Additional Comments From pinskia at gcc dot gnu dot org  2004-11-20 
06:37 ---
The corresponding asm for the tcb:
cmpw cr7,r3,r4
mr r11,r3
li r10,0
li r3,1
beqlr- cr7
lwz r0,0(r11)
lwz r2,0(r4)
slwi r9,r0,2
cmpw cr7,r2,r0
ble- cr7,L7
lwz r2,4(r4)
lwzx r0,r9,r2
xor r10,r11,r0
subfic r2,r10,0
adde r10,r2,r10
L7:
mr r3,r10
blr

the mainline (which is much worse):
cmpw cr7,r3,r4
li r10,0
beq- cr7,L2
lwz r0,0(r3)
li r11,1
lwz r2,0(r4)
slwi r9,r0,2
cmpw cr7,r2,r0
bgt- cr7,L12
L4:
li r11,0
L7:
cmpwi cr7,r11,0
bne- cr7,L2
mr r3,r10
blr
L12:
lwz r2,4(r4)
lwzx r0,r9,r2
cmpw cr7,r3,r0
bne+ cr7,L4
b L7
L2:
li r10,1
mr r3,r10
blr

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18576

[Bug tree-optimization/18576] missing jump threading because of type changes

[pinskia at gcc dot gnu dot org]

--- Additional Comments From pinskia at gcc dot gnu dot org  2004-11-20 
06:43 ---
Oh, even though the asm is fixed on the tcb we are still missing it on the tree 
level.

By the way I noticed this while trying to speedup flow_bb_inside_loop_p in the 
first place.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18576