[llvm-commits] [LLVM, loop-unswitch, bugfix for #11429] Wrong behaviour for switches.

[Stepan Dyatkovskiy]

Hi Dan. This bug is described in details (with examples) here:
http://llvm.org/bugs/show_bug.cgi?id=11429

Regarding to your questions..

> It's even possible that this bug is
> accidentally helping the code by preventing it from unswitching too
> much in the presence of switches.
> Do you have an idea on what impact
> this patch has on code size, and performance, in general?

Happily this bug helps to keep the code size small. But 
LoopUnswitch::UnswitchIfProfitable method already controls the produced 
code size. Please, see LoopUnswitch.cpp, string #446 for more details. I 
think that if we need to improve the "restrictioning" of produced code 
size we need to implement this improvement instead of keeping some 
strange code.

About impact on code size.
For switch with N cases (+ 1 default) we got N new loops. If you wish I 
can present the .ll code that should be produced after optimization.

Impact on performance.
The main purpose of this optimization is to move out of loop the 
switches. Each unswitched case increases the performance.

About releaseMemory and CloneUnswitchedVals.
This methods the part of unswitch info cloning for new loops. It is also 
described in bug #11429.

 > CloneUnswitchedVals doesn't actually need to iterate over the
 > instructions in the block to find the SwitchInst. If there's a
 > SwitchInst present, it'll be the Terminator instruction.
OK. You're right. Please find the fixed patch.

Thanks!
-Stepan.

>
> Thanks,
>
> Dan
>
> On Dec 2, 2011, at 12:31 AM, Stepan Dyatkovskiy wrote:
>
>>
>> ping.
>>
>> -Stepan.
>> Stepan Dyatkovskiy wrote:
>>> ping.
>>>
>>> -Stepan.
>>>
>>> Stepan Dyatkovskiy wrote:
>>>> ping.
>>>>
>>>> -Stepan.
>>>>
>>>> Stepan Dyatkovskiy wrote:
>>>>> ping.
>>>>>
>>>>> -Stepan
>>>>>
>>>>> Stepan Dyatkovskiy wrote:
>>>>>> Hi all. Please find the patch in attachment for review. Regression tests
>>>>>> are attached in separated patch.
>>>>>>
>>>>>> Short tests description:
>>>>>>
>>>>>> 2011-11-18-SimpleSwitch.ll
>>>>>> Check simple that simple switch will totally unswitched:
>>>>>>
>>>>>> for (...)
>>>>>> switch(c) {
>>>>>> case 0: inc(); break;
>>>>>> case 1: dec(); break;
>>>>>> default: goto loop_exit;
>>>>>> }
>>>>>>
>>>>>> loop_exit:
>>>>>> ...
>>>>>>
>>>>>> Result of processing should be 2 additional loops for c == 0 and for c
>>>>>> == 1.
>>>>>>
>>>>>>
>>>>>> 2011-11-18-TwoSwitches.ll
>>>>>> Check that second switch will unswitched too. Check that switches will
>>>>>> not unswitched again in new loop:
>>>>>>
>>>>>> Initially we have the next:
>>>>>>
>>>>>> for (...) {
>>>>>> switch(c) {
>>>>>> case 0: inc(); break;
>>>>>> }
>>>>>> switch(d) {
>>>>>> case 0: inc(); break;
>>>>>> }
>>>>>> }
>>>>>>
>>>>>> After optimization we should got 3 additional loops: when (c == 0&&     d
>>>>>> == 0), when (c == 0&&     d != 0) and when (c != 0&&     d == 0). Original
>>>>>> loop will activated for (c != 0&&     d != 0):
>>>>>>
>>>>>> if (c == 0&&     d == 0) {
>>>>>> for (...) ... // All is clear here. Two "switch(0)" instructions.
>>>>>> } else if (c == 0&&     d != 0) {
>>>>>> for (...) {
>>>>>> switch (0) { // c == 0
>>>>>> case 0: inc(); break;
>>>>>> }
>>>>>> switch (d) {
>>>>>> case 0: goto unreachable; // CHECK: That it will not unswitched
>>>>>> // again, since it looks "trivial",
>>>>>> // see LoopUnswitch::IsTrivialUnswitchCondition
>>>>>> }
>>>>>> }
>>>>>> } else if (c != 0&&     d == 0) {
>>>>>> // the same...
>>>>>> ...
>>>>>> } else { // if (c != 0&&     d != 0)
>>>>>> // Original totally unswitched loop:
>>>>>> switch (c) {
>>>>>> case 0: goto unreachable;
>>>>>> }
>>>>>> switch (d) {
>>>>>> case 0: goto unreachable;
>>>>>> }
>>>>>> }
>>>>>>
>>>>>> Stepan.
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> llvm-commits mailing list
>>>>>> llvm-commits at cs.uiuc.edu
>>>>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>>>>
>>>>> _______________________________________________
>>>>> llvm-commits mailing list
>>>>> llvm-commits at cs.uiuc.edu
>>>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>>>
>>>> _______________________________________________
>>>> llvm-commits mailing list
>>>> llvm-commits at cs.uiuc.edu
>>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>>
>>> _______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: loop-unswitch-unswitchvals.patch
Type: text/x-patch
Size: 6362 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20111205/bfcf2325/attachment.bin>