JIT: Fix likelihoods in range test bool opts

[amanasifkhalid]

When optimizing compound conditionals into one check and branch, we'd expect the likelihood of the optimized branch to be the product of the conditionals' likelihoods. We currently get this wrong in range test opts if the second comparison is flipped, yielding behavior like this:

---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BBnum BBid ref try hnd preds           weight     IBC [IL range]   [jump]                            [EH region]        [flags]
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BB01 [0000]  1                             1    33784 [000..00C)-> BB12(0.407),BB11(0.593) ( cond )                   i IBC
BB11 [0009]  1       BB01                  0.59 20037 [000..001)                           (throw )                   i IBC gcsafe
BB12 [0010]  1       BB01                  0.41 13747 [000..001)-> BB14(0.899),BB08(0.101) ( cond )                   i IBC
BB14 [0012]  3       BB03,BB08,BB12        0.37 12357 [000..021)                           (return)                   i IBC jmp
BB08 [0006]  1       BB12                  0.45 15138 [000..015)-> BB14(0.2),BB03(0.8)     ( cond )                   i IBC
BB03 [0002]  1       BB08                  0        0 [015..019)-> BB14(1)                 (always)                   i IBC rare
---------------------------------------------------------------------------------------------------------------------------------------------------------------------

*************** In optOptimizeBools()
setting likelihood of BB12 -> BB14 from 0.8988597 to 0.0809122
setting likelihood of BB12 -> BB03 to 0.9190878
fgRemoveBlock BB08, unreachable=true

Removing unreachable BB08


---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BBnum BBid ref try hnd preds           weight     IBC [IL range]   [jump]                            [EH region]        [flags]
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BB01 [0000]  1                             1    33784 [000..00C)-> BB12(0.407),BB11(0.593) ( cond )                   i IBC
BB11 [0009]  1       BB01                  0.59 20037 [000..001)                           (throw )                   i IBC gcsafe
BB12 [0010]  1       BB01                  0.41 13747 [000..001)-> BB14(0.0809),BB03(0.919)  ( cond )                   i IBC
BB14 [0012]  2       BB03,BB12             0.03  1112 [000..021)                           (return)                   i IBC jmp
BB03 [0002]  1       BB12                  0.37 12635 [015..019)-> BB14(1)                 (always)                   i IBC
---------------------------------------------------------------------------------------------------------------------------------------------------------------------

Based on what the profile initially looked like, the likelihoods out of BB12 should be flipped.

[Copilot]

Pull Request Overview

This PR fixes the likelihood computation in range tests for bool optimizations when the second comparison is reversed.

• Moves likelihood assignments outside the if/else branches to ensure a consistent calculation of the branch probabilities.
• Removes redundant likelihood assignments that led to incorrect flipping of probabilities.

Comments suppressed due to low confidence (2)

src/coreclr/jit/optimizebools.cpp:828

• Ensure that moving the likelihood assignments outside the if/else block correctly reflects the intended behavior for both cmp2IsReversed cases. Verify that the new structure accurately flips the likelihoods as expected.

    newEdge->setLikelihood(inRangeLikelihood);

src/coreclr/jit/optimizebools.cpp:841

• Check that the removal of the redundant likelihood assignments in the reversed branch does not inadvertently affect the logic ensuring the branch likelihood is correctly flipped.

        assert(m_b1->TrueTargetIs(notInRangeBb));

[dotnet-policy-service]

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

[amanasifkhalid]

@dotnet/jit-contrib PTAL. Diffs. coreclr_tests on x64 is a big outlier -- this change affects a lot of our HW intrinsics test code:

Top method regressions (percentages):
           2 (10.000% of base) : 529608.dasm - RedundantBranchUnsigned2:Test_Span1(System.Span`1[int],int) (FullOpts)
          57 (8.321% of base) : 441611.dasm - JIT.HardwareIntrinsics.General._Vector128.VectorWidenTest__WidenInt16+DataTable:.ctor(sbyte[],short[],short[],int):this (FullOpts)
          57 (8.321% of base) : 441560.dasm - JIT.HardwareIntrinsics.General._Vector128.VectorWidenTest__WidenUInt16+DataTable:.ctor(byte[],ushort[],ushort[],int):this (FullOpts)
          57 (8.321% of base) : 442963.dasm - JIT.HardwareIntrinsics.General._Vector256.VectorWidenTest__WidenInt16+DataTable:.ctor(sbyte[],short[],short[],int):this (FullOpts)
          57 (8.321% of base) : 442954.dasm - JIT.HardwareIntrinsics.General._Vector256.VectorWidenTest__WidenUInt16+DataTable:.ctor(byte[],ushort[],ushort[],int):this (FullOpts)
          57 (8.321% of base) : 448002.dasm - JIT.HardwareIntrinsics.General._Vector512.VectorWidenTest__WidenInt16+DataTable:.ctor(sbyte[],short[],short[],int):this (FullOpts)
          57 (8.321% of base) : 447993.dasm - JIT.HardwareIntrinsics.General._Vector512.VectorWidenTest__WidenUInt16+DataTable:.ctor(byte[],ushort[],ushort[],int):this (FullOpts)
          57 (8.321% of base) : 457511.dasm - JIT.HardwareIntrinsics.General._Vector64.VectorWidenTest__WidenInt16+DataTable:.ctor(sbyte[],short[],short[],int):this (FullOpts)
          57 (8.321% of base) : 457461.dasm - JIT.HardwareIntrinsics.General._Vector64.VectorWidenTest__WidenUInt16+DataTable:.ctor(byte[],ushort[],ushort[],int):this (FullOpts)
          57 (8.273% of base) : 441577.dasm - JIT.HardwareIntrinsics.General._Vector128.VectorWidenTest__WidenInt32+DataTable:.ctor(short[],int[],int[],int):this (FullOpts)
          57 (8.273% of base) : 441645.dasm - JIT.HardwareIntrinsics.General._Vector128.VectorWidenTest__WidenUInt32+DataTable:.ctor(ushort[],uint[],uint[],int):this (FullOpts)
          57 (8.273% of base) : 442957.dasm - JIT.HardwareIntrinsics.General._Vector256.VectorWidenTest__WidenInt32+DataTable:.ctor(short[],int[],int[],int):this (FullOpts)
          57 (8.273% of base) : 442969.dasm - JIT.HardwareIntrinsics.General._Vector256.VectorWidenTest__WidenUInt32+DataTable:.ctor(ushort[],uint[],uint[],int):this (FullOpts)
          57 (8.273% of base) : 447996.dasm - JIT.HardwareIntrinsics.General._Vector512.VectorWidenTest__WidenInt32+DataTable:.ctor(short[],int[],int[],int):this (FullOpts)
          57 (8.273% of base) : 448008.dasm - JIT.HardwareIntrinsics.General._Vector512.VectorWidenTest__WidenUInt32+DataTable:.ctor(ushort[],uint[],uint[],int):this (FullOpts)
          57 (8.273% of base) : 457477.dasm - JIT.HardwareIntrinsics.General._Vector64.VectorWidenTest__WidenInt32+DataTable:.ctor(short[],int[],int[],int):this (FullOpts)
          57 (8.273% of base) : 457545.dasm - JIT.HardwareIntrinsics.General._Vector64.VectorWidenTest__WidenUInt32+DataTable:.ctor(ushort[],uint[],uint[],int):this (FullOpts)
          57 (8.261% of base) : 442960.dasm - JIT.HardwareIntrinsics.General._Vector256.VectorWidenTest__WidenInt64+DataTable:.ctor(int[],long[],long[],int):this (FullOpts)
          57 (8.261% of base) : 447999.dasm - JIT.HardwareIntrinsics.General._Vector512.VectorWidenTest__WidenInt64+DataTable:.ctor(int[],long[],long[],int):this (FullOpts)
          57 (8.261% of base) : 457528.dasm - JIT.HardwareIntrinsics.General._Vector64.VectorWidenTest__WidenDouble+DataTable:.ctor(float[],double[],double[],int):this (FullOpts)