llvm dev - Jan 2022 - Which optimization pass deals with heap values?

All the optimization passes I can find in the documentation, deal with
register, stack or global values; I cannot find any mention of anything
trying to optimize values on the heap. But this:

#include <stdlib.h>

int main(int argc, char **argv) {
  int *a = malloc(10 * sizeof(int));
  for (int i = 0; i < 10; i++)
    a[i] = i;
  return a[5];
}

compiles to 'ret 5' (as it should). Which pass does that?
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20220117/1ff82f5c/attachment.html>

Hi,

   The optimization should be done at InstructionCombine pass. See:
https://godbolt.org/z/zGvrMerKv. Previous cleaning passes are necessary.

   But I feel this optimization is a little bit aggresive. Since it changes the
behavior of the program. The program above might hit a memory leak while the
optimized one would not.

   As a reference, GCC wouldn't do this under O2:
https://godbolt.org/z/5z1GWMWPj. It would only do this under O3.

Thanks,
Chuanqi
------------------------------------------------------------------
From:llvm-dev <llvm-dev at lists.llvm.org>
Send Time:2022年1月17日(星期一) 16:08
To:llvm-dev <llvm-dev at lists.llvm.org>
Subject:[llvm-dev] Which optimization pass deals with heap values?

All the optimization passes I can find in the documentation, deal with register,
stack or global values; I cannot find any mention of anything trying to optimize
values on the heap. But this:

#include <stdlib.h>

int main(int argc, char **argv) {
  int *a = malloc(10 * sizeof(int));
  for (int i = 0; i < 10; i++)
    a[i] = i;
  return a[5];
}

compiles to 'ret 5' (as it should). Which pass does that?

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20220117/16c62777/attachment.html>

> On Jan 17, 2022, at 08:07, Russell Wallace via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> 
> All the optimization passes I can find in the documentation, deal with
register, stack or global values; I cannot find any mention of anything trying
to optimize values on the heap. But this:
> 
> #include <stdlib.h>
> 
> int main(int argc, char **argv) {
>   int *a = malloc(10 * sizeof(int));
>   for (int i = 0; i < 10; i++)
>     a[i] = i;
>   return a[5];
> }
> 
> compiles to 'ret 5' (as it should). Which pass does that?
> _____________________________________________

You can use `-mllvm -print-changed` with clang when built with assertions to
find this out, e.g. see https://clang.godbolt.org/z/aM18Tr7Mr
<https://clang.godbolt.org/z/aM18Tr7Mr>

It will be multiple optimizations working together. There are several passes
that optimize memory operation, like DSE, GVN, EarlyCSE +MemorySSA.


-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20220117/4d02d947/attachment.html>