Remove RewriteTensorPointer from the optimization pipeline

[etiotto]

No description provided.

[etiotto]

Depends on PR #2598

[etiotto]

Performance results (see label "etiotto3"):

http://benchmarks.glados.intel.com/d/1pXX4hUSz/microbenchmarks?orgId=1&var-tag=ci%7Cetiotto3&var-bench=attn&var-bench=gemm&var-bench=gemm-preop-exp&var-bench=gemm-postop-gelu&var-bench=gemm-postop-addmatrix&var-bench=gemm-streamk&var-bench=gemm-splitk&var-bench=gemm-bt&var-bench=gemm-at&var-device=All&var-compiler=triton&var-backend=All&var-baseline_backend=triton-ci-XPU%201550&var-target_backend=triton-ci-XPU%201550&from=now-5d&to=now

There was a degradation in "gemm-at" which has been fixed in this PR by enhancing the lowering code (to LLVM) for loads with blocked pointers, we can now generate vector loads of the appropriate vector length.

[etiotto]

Depends on PR #2598

I have merged that PR in now.

[whitneywhtsang]

Can you please explain why all changes not in compiler.py are needed?

[etiotto]

Can you please explain why all changes not in compiler.py are needed?

You want some comments in the PR to help understand what each change does ? Can you be more specific and put comments on pieces of code you want explained? I'll start to add comments only in the major pieces.

[whitneywhtsang]

Can you please explain why all changes not in compiler.py are needed?

You want some comments in the PR to help understand what each change does ? Can you be more specific and put comments on pieces of code you want explained? I'll start to add comments only in the major pieces.

Discussed offline, one main change is to use intel version of ModuleAxisInfoAnalysis.

[etiotto]

Depends on PR @#2623.