[clang] [llvm] [Pipelines] Move IPSCCP after inliner pipeline (PR #96620)

[via llvm-commits]

joshua-arch1 wrote:

> @joshua-arch1 I tested this patch on LTO and it is consistent with what you describe. With the `-flto` option enabled is that `digits_2` has two specializations whether the patch is applied or not. However with LTO enabled, LLVM performance is comparable to GCC (in terms of number of instructions ). Here are some figures:
> 
> > on x86_64 ArchLinux
> > via `perf stat ./exchange2_r 0`
> 
> Compiler	Instructions
> gfortran O3 (archlinux gcc-fortran 14.2.1+r134+gab884fffe3fc-2)	54,865,827,741
> gfortran O3 lto (archlinux gcc-fortran 14.2.1+r134+gab884fffe3fc-2)	54,186,161,980
> flang O3 (#62d44fbd)	107,953,750,439
> flang O3 lto (#62d44fbd)	53,391,922,257
> flang O3 (patched #62d44fbd)	53,146,581,727
> flang O3 lto (patched #62d44fbd)	53,391,760,016
> When using LTO, I don't think there is a problem that performance is worse than GCC, so this patch is not specific to LTO.

Hi @sihuan, I have tried to apply yout initial design where IPSCCP is directly moved after inliner. It can work with LTO and `digits_2` has eight specializations just like gfortran. 

https://github.com/llvm/llvm-project/pull/96620