Re: Question about gimple code during optimizing if-conversion
On Fri, Oct 13, 2023 at 10:16 PM Hanke Zhang via Gcc wrote:
>
> Hi, I'm working on optimizing if-conversion for my own business
> recently. I got a problem here.
>
> I tried to optimize it in such a case, for example, when a conditional
> statement block has only if statement and no else statement, the
> source C code looks like this:
>
> int* foo; // assume this has been initialized
> int c = rand(), t = rand(), size = 1000;
> for (int i = 0; i < size; i++) {
> if (foo[i] & (1 << c)) foo[i] ^= (1 << t);
> }
>
> Part of its corresponding gimple is optimized like this before if-conversion:
>
> :
> # i_71 = PHI
> # ivtmp_9 = PHI
> _5 = (long unsigned int) i_71;
> _6 = _5 * 4;
> _7 = foo_23 + _6;
> _8 = *_7;
> shifttmp_75 = _8 & shifttmp_76;
> if (shifttmp_75 != 0)
> goto ; [50.00%]
> else
> goto ; [50.00%]
>
>[local count: 531502205]:
> goto ; [100.00%]
>
>[local count: 531502204]:
> _12 = _8 ^ _11;
> *_7 = _12;
>
>[local count: 1063004409]:
> i_39 = i_71 + 1;
> ivtmp_73 = ivtmp_9 - 1;
> if (ivtmp_73 != 0)
> goto ; [99.00%]
> else
> goto ; [1.00%]
>
> I want to add some statements to gimple to make it like adding an else
> block to the source code.
>
> // What I expected:
> int* foo; // assume this has been initialized
> int c = rand(), t = rand(), size = 1000;
> for (int i = 0; i < size; i++) {
> if (foo[i] & (1 << c)) foo[i] ^= (1 << t);
> + else foo[i] = foo[i]; // I want to add a statment here !
> }
>
> And of course I can't change the source code for real, so I can only
> add a pass in front of if-conversion to modify the gimple.
>
> For the example above, I know that I have to add them in the block
> '', but what confuses me is that I don't know what kind of
> statement to add to be legal due to my poor experience.
>
> I try to add something like this below, but the compile error just
> happened. So I'm here for help. What kind of statements should I add
> here?
>
> [local count: 531502205]:
> + *_7 = *_7
> goto ; [100.00%]
>
> Finally, The reason I did this was to avoid MASK_STORE generation,
> because it might add an if branch in the final assembly which I don't
> like it to be. And after such a modification, if-conversion should
> have been changed it to the form of a ternary expression, which would
> reduce the occurrence of branches after final vectorization and
> produce more efficient code.
>
> Or there if is a better way to get rid of MASK_STORE, please tell me
> about that. :)
So there are 2 issues with this transformation which you need to take
into account.
1) C11/C++11 threading model (-fallow-store-data-races is needed)
2) foo could be read only and cause a trap if written to. if the
branch is never taken there would be no writes
Thanks,
Andrew Pinski
>
> Thanks
> Hanke Zhang
Re: Question about gimple code during optimizing if-conversion
On 10/14/23 09:49, Andrew Pinski via Gcc wrote:
On Fri, Oct 13, 2023 at 10:16 PM Hanke Zhang via Gcc wrote:
Hi, I'm working on optimizing if-conversion for my own business
recently. I got a problem here.
I tried to optimize it in such a case, for example, when a conditional
statement block has only if statement and no else statement, the
source C code looks like this:
int* foo; // assume this has been initialized
int c = rand(), t = rand(), size = 1000;
for (int i = 0; i < size; i++) {
if (foo[i] & (1 << c)) foo[i] ^= (1 << t);
}
Part of its corresponding gimple is optimized like this before if-conversion:
:
# i_71 = PHI
# ivtmp_9 = PHI
_5 = (long unsigned int) i_71;
_6 = _5 * 4;
_7 = foo_23 + _6;
_8 = *_7;
shifttmp_75 = _8 & shifttmp_76;
if (shifttmp_75 != 0)
goto ; [50.00%]
else
goto ; [50.00%]
[local count: 531502205]:
goto ; [100.00%]
[local count: 531502204]:
_12 = _8 ^ _11;
*_7 = _12;
[local count: 1063004409]:
i_39 = i_71 + 1;
ivtmp_73 = ivtmp_9 - 1;
if (ivtmp_73 != 0)
goto ; [99.00%]
else
goto ; [1.00%]
I want to add some statements to gimple to make it like adding an else
block to the source code.
// What I expected:
int* foo; // assume this has been initialized
int c = rand(), t = rand(), size = 1000;
for (int i = 0; i < size; i++) {
if (foo[i] & (1 << c)) foo[i] ^= (1 << t);
+ else foo[i] = foo[i]; // I want to add a statment here !
}
And of course I can't change the source code for real, so I can only
add a pass in front of if-conversion to modify the gimple.
For the example above, I know that I have to add them in the block
'', but what confuses me is that I don't know what kind of
statement to add to be legal due to my poor experience.
I try to add something like this below, but the compile error just
happened. So I'm here for help. What kind of statements should I add
here?
[local count: 531502205]:
+ *_7 = *_7
goto ; [100.00%]
Finally, The reason I did this was to avoid MASK_STORE generation,
because it might add an if branch in the final assembly which I don't
like it to be. And after such a modification, if-conversion should
have been changed it to the form of a ternary expression, which would
reduce the occurrence of branches after final vectorization and
produce more efficient code.
Or there if is a better way to get rid of MASK_STORE, please tell me
about that. :)
So there are 2 issues with this transformation which you need to take
into account.
1) C11/C++11 threading model (-fallow-store-data-races is needed)
2) foo could be read only and cause a trap if written to. if the
branch is never taken there would be no writes
Right. See ifcvt_memrefs_wont_trap. That probably could be extended
to capture additional cases. But I'm not sure that'll be sufficient for
Hanke's problem.
Jeff
gcc-13-20231014 is now available
Snapshot gcc-13-20231014 is now available on https://gcc.gnu.org/pub/gcc/snapshots/13-20231014/ and on various mirrors, see http://gcc.gnu.org/mirrors.html for details. This snapshot has been generated from the GCC 13 git branch with the following options: git://gcc.gnu.org/git/gcc.git branch releases/gcc-13 revision a9f39860efa645149ec88f0e9a7a31c3362b7e46 You'll find: gcc-13-20231014.tar.xz Complete GCC SHA256=40bf42e54cefefa4a8f35c48e0f290c9ef8118eee9a72800296a0e620dfb0240 SHA1=c7fa5bd8478bbc42556b25bc5e8b0c1c1dff1cdd Diffs from 13-20231007 are available in the diffs/ subdirectory. When a particular snapshot is ready for public consumption the LATEST-13 link is updated and a message is sent to the gcc list. Please do not use a snapshot before it has been announced that way.
