[JIT] X64 - More replacement sequences for integer multiplication by a constant

[TIHan]

Description

Resolves this issue: #75119 - we are only planning on doing 3-instruction replacement sequences or less, no more than that.

This PR does these two optimizations:

-       mov      edx, eax
-       shl      edx, 2
+       lea      edx, [4*rax]

-       mov      edx, eax
-       shl      edx, 3
+       lea      edx, [8*rax]

This PR also lifts the restriction that only allowed GT_LCL_VAR to be the first operand for "multiply by constant"; now it will create a tmp local if needs to - this ensures that we can take advantage of these optimizations.

I wanted to keep the "multiply by constant" optimizations in one place, and that place is lowering. Doing this required to disable any "multiply by constant" optimizations from happening in Tier-0, which I think is reasonable. This means Tier-0 will always emit imul, except for the cases that can emit lea.

We only do the "multiply by constant" -> lea instruction in codegen.

Notes:
There are other replacement sequences that were considered, but ultimately, those replacement sequences have higher latency totals than the single imul for modern CPUs. These cases are described and tested in the IntMultiply disasm tests.

Microbenchmark Results

    using BenchmarkDotNet.Attributes;
    using BenchmarkDotNet.Running;

    namespace Perf
    {
        public class Multiply
        {
            static ulong fieldValue = 1;

            [Benchmark]
            public void shift_2()
            {
                ulong value = fieldValue;
                for (ulong i = 1; i < 10000000; i++)
                {
                    value = i << 2;
                }
                fieldValue = value;
            }

            [Benchmark]
            public void shift_3()
            {
                ulong value = fieldValue;
                for (ulong i = 1; i < 10000000; i++)
                {
                    value = i << 3;
                }
                fieldValue = value;
            }
        }
        static class Program
        {
            static int Main(string[] args)
            {
                BenchmarkRunner.Run<Multiply>(null, args);
                return 0;
            }
        }
    }

Before:

BenchmarkDotNet=v0.13.2, OS=Windows 11 (10.0.22621.674)
AMD Ryzen 9 7950X, 1 CPU, 32 logical and 16 physical cores
.NET SDK=7.0.100-rc.2.22477.23
  [Host]     : .NET 7.0.0 (7.0.22.47203), X64 RyuJIT AVX2
  DefaultJob : .NET 7.0.0 (7.0.22.47203), X64 RyuJIT AVX2


|  Method |     Mean |     Error |    StdDev |
|-------- |---------:|----------:|----------:|
| shift_2 | 1.842 ms | 0.0032 ms | 0.0030 ms |
| shift_3 | 1.842 ms | 0.0025 ms | 0.0023 ms |

After:

BenchmarkDotNet=v0.13.2, OS=Windows 11 (10.0.22621.674)
AMD Ryzen 9 7950X, 1 CPU, 32 logical and 16 physical cores
.NET SDK=7.0.100-rc.2.22477.23
  [Host]     : .NET 7.0.0 (7.0.22.47203), X64 RyuJIT AVX2
  Job-GILPZC : .NET 8.0.0 (42.42.42.42424), X64 RyuJIT AVX2

Toolchain=CoreRun

|  Method |     Mean |     Error |    StdDev |
|-------- |---------:|----------:|----------:|
| shift_2 | 1.828 ms | 0.0048 ms | 0.0045 ms |
| shift_3 | 1.830 ms | 0.0052 ms | 0.0049 ms |

Acceptance Criteria

• Merge [JIT] X64 - Three instruction replacement sequence for multiply in certain cases #76981

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Description
TBD

Refactors some of the codegen multiply optimizations by moving them to lowering.

Acceptance Criteria

• Merge [JIT] X64 - Three instruction replacement sequence for multiply in certain cases #76981

Author:	TIHan
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[EgorBo]

Can you please run fuzz pipelines yourself once you're done with changes?

[TIHan]

/azp run Fuzzlyn

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[TIHan]

Looks like Fuzzlyn x64 passed. I need to update one of the disasm checks as I forgot a minor thing.

[tannergooding]

What about:

-       mov      edx, eax
-       shl      edx, 1
+       lea      edx, [2*rax]

[TIHan]

@tannergooding We already handle that case for x << 1 in codegen, it emits this:

lea rax, [rcx+rcx]

[TIHan]

@dotnet/jit-contrib This is ready again - I fixed the tests and one of the earlier commits did pass fuzzlyn.

[BruceForstall]

Have you created a BDN microbenchmark to show the performance before/after your optimizations?

[TIHan]

Have you created a BDN microbenchmark to show the performance before/after your optimizations?

Have not, but will do that now.

[TIHan]

@BruceForstall I've provided microbenchmark results in the description of this PR.

[BruceForstall]

Your SuperPMI jobs are failing with:

Traceback (most recent call last):

  File "C:\h\w\B0FF09A8\p\superpmi.py", line 1720, in create_one_artifact

    with open(item_path, 'r') as file_handle:

FileNotFoundError: [Errno 2] No such file or directory: 'C:\\h\\w\\B0FF09A8\\p\\artifacts\\spmi\\asm.coreclr_tests.run.windows.x64.checked\\base\\18886.dasm'



The above exception was the direct cause of the following exception:



Traceback (most recent call last):

  File "C:\h\w\B0FF09A8\p\superpmi.py", line 4451, in <module>

    sys.exit(main(args))

  File "C:\h\w\B0FF09A8\p\superpmi.py", line 4342, in main

    success = asm_diffs.replay_with_asm_diffs()

  File "C:\h\w\B0FF09A8\p\superpmi.py", line 1736, in replay_with_asm_diffs

    subproc_helper.run_to_completion(create_replay_artifacts, self, mch_file, asm_dotnet_vars_full_env, text_differences, base_asm_location, diff_asm_location, ".dasm")

  File "C:\h\w\B0FF09A8\p\superpmi.py", line 601, in run_to_completion

    loop.run_until_complete(self.__run_to_completion__(async_callback, *extra_args))

  File "C:\python3\lib\asyncio\base_events.py", line 647, in run_until_complete

    return future.result()

  File "C:\h\w\B0FF09A8\p\superpmi.py", line 584, in __run_to_completion__

    await asyncio.gather(*tasks)

  File "C:\h\w\B0FF09A8\p\superpmi.py", line 556, in __get_item__

    await async_callback(print_prefix, item, *extra_args)

  File "C:\h\w\B0FF09A8\p\superpmi.py", line 1726, in create_replay_artifacts

    base_txt = await create_one_artifact(self.base_jit_path, base_location, flags + base_option_flags_for_diff_artifact)

  File "C:\h\w\B0FF09A8\p\superpmi.py", line 1723, in create_one_artifact

    raise create_exception() from err

Exception: Failure while creating JitStdOutFile.

Exit code: 0

This has been fixed. Maybe you just need to rebase/re-push to trigger updated CI testing?

[BruceForstall]

@TIHan Note that the formatting job failed. If you don't already, try to get into the habit of running "jit-format -f" before pushing a PR change to GitHub.

Also, it looks like the unix-x64 superpmi-diffs job failed for unknown reasons.

[TIHan]

@dotnet/jit-contrib @BruceForstall looks like CI is passing - is there anything else that I should do in the PR?

[BruceForstall]

Diffs