Forcing loop optimization

Project Infrastructure LLVM Dev List Archives
1

Dear all,

Good day to you. Is it possible to have forced application of loop unrolling or any other loop optimization ? I have tried applying loop unrolling on the following code. However, the optimization has not been applied by Clang.

Code (code.c)

int main()
{
int i1, i2, N;
int In[50], A[50], D[50];

N = 25;
A[0] = In[0] + 5;
#pragma clang loop unroll (enable)
for(i1 = 1; i1 <= 25; i1++)
{
A[i1] = In[i1];
D[i1] = A[i1];
}
return D[N];
}

The applied commands

  1. clang -S -emit-llvm code.c -o code.ll -Xclang -disable-O0-optnon
  2. opt -loop-unroll -S code.ll -o code-opt.ll -opt-bisect-limit=300 -unroll-count=3 -print-after=loop-unroll
    Regards,
    Sudakshina
2

Hi Sudakshina,

Add -sroa (or mem2reg) to your second command to simplify the IR in a way loop unroll can deal with:

I removed options that are not needed in the link above.

Hope this helps,
Johannes

3

Thanks a lot Mr. Johannes Doerfert.

4

Dear Prof. Johannes Doerfert,

Sorry to disturb you again. The code LLVM IR that you have shared is not only the result of applying x86-64 clang 12.0.0 applied with -emit-llvm -Xclang -disable-O0-optnone options. You might have applied some other options. The reason I am asking is that the LLVM IR code which is shared has a header and footer which generally do not appear if a code is subjected to godbolt explorer. If the LLVM IR code generated by applying x86-64 clang 12.0.0 applied with -emit-llvm -Xclang -disable-O0-optnone options is subjected to opt with -sroa -loop-unroll -unroll-count=3 options, some error is shown. I am very new to compiler explorer. I shall be grateful if you kindly reply.

Thanks,
Sudakshina

5

Dear Prof. Johannes Doerfert,

Just Johannes.

Sorry to disturb you again. The code LLVM IR that you have shared is not
only the result of applying x86-64 clang 12.0.0 applied with -emit-llvm
-Xclang -disable-O0-optnone options. You might have applied some other
options. The reason I am asking is that the LLVM IR code which is shared
has a header and footer which generally do not appear if a code is
subjected to godbolt explorer. If the LLVM IR code generated by applying
x86-64 clang 12.0.0 applied with -emit-llvm -Xclang -disable-O0-optnone
options is subjected to opt with -sroa -loop-unroll -unroll-count=3
options, some error is shown. I am very new to compiler explorer. I shall
be grateful if you kindly reply.

Godbold has a "filter" function. Look for "Filter" in the clang tab
and disable what you don't want. You should be able to do all of
this locally as well. There are multiple good tutorials on how to
compile and use the tools like opt on youtube.

~ Johannes

6

For forcing transformations via source code annotations, see [1].

However, there may be other reasons for an optimization to be not implied, such as the loop not having been normalized. See [2]

[1] https://clang.llvm.org/docs/LanguageExtensions.html#extensions-for-loop-hint-optimizations
[2] https://llvm.org/docs/LoopTerminology.html

Michael