Integrate llvm/llvm-project@4845b3e

[amd-eochoalo]

• Carries revert for 9690a718b.
• Carries revert for fcf79e5. See Remove revert of https://github.com/llvm/llvm-project/pull/160615 on LLVM integrate #22171 for details

[yash-amd]

Due to this pr we are seeing a performance degradation of 9% in decode time for llama-8b-fp8 model.

Repro:

Steps to reproduce:
Mlir is here

Setup IREE and Run the following steps on the sharkmi300x-4 machine

Compile Mlir With:

iree-compile output.mlir \
        --iree-hip-target=gfx942 -o output.vmfb \
        --iree-hal-target-device=hip --iree-opt-level=O3 \
        --iree-hal-indirect-command-buffers=true \
        --iree-stream-resource-memory-model=discrete \
        --iree-hip-enable-tensor-ukernels \
        --iree-hal-memoization=true \
        --iree-dispatch-creation-propagate-collapse-across-expands=true \
        --iree-hip-specialize-dispatches

Test for benchmark(performance):

iree-benchmark-module --hip_use_streams=true \
        --module=output.vmfb \
        --parameters=model=/shark-dev/8b/fp8/attnf8/native_fp8_e4m3fnuz_llama3_8b.irpa \
        --device=hip \
        --function=decode_bs4 \
        --input=4x1xi64 \
        --input=4xi64 \
        --input=4xi64 \
        --input=4x65xi64 \
        --input=261x2097152xf8E4M3FNUZ \
        --benchmark_repetitions=3 \
        --benchmark_out_format=json \
        --benchmark_out=<BENCHMARK_DIR>/llama-8B-FP8_decode_bs4_isl_2048.json

output time is generated in the <benchmark_dir>/llama-8B-FP8_prefill_bs4_isl_2048.json file.