Enable CSE for floating-point constants #44419

briansull · 2020-11-09T18:28:10Z

No description provided.

tannergooding · 2020-11-10T00:25:36Z

src/coreclr/src/jit/gentree.cpp

Unrelated to your change, but this isn't accurate anymore. fldz and fld1 are from the x87 stack and we almost exclusively use SSE/SSE2 now (modulo some ABI things for 32-bit).

Now its that we use xorps for 0 and load values for everything else.

tannergooding · 2020-11-10T00:32:09Z

This is just CSE of constants right? What is the behavior for floating-point values that compare equal but whose bitwise values are different (-0.0 vs +0.0) or values who are bitwise equivalent but compare inequal (double.NaN or any of the NaN representations in general)?

briansull · 2020-11-10T02:23:45Z

The ValueNumbers of the two constants have to be the same, We don't use floating point compare to determine if two constant are equal. Value numbering already has to properly handle floating point constants properly.

Tracing through the value number implementation, It eventually uses this function in jithashtable.h

template <typename T>
struct JitLargePrimitiveKeyFuncs : public JitKeyFuncsDefEquals<T>
{
    static unsigned GetHashCode(const T val)
    {
        // A static cast when T is a float or a double converts the value (i.e. 0.25 converts to 0)
        //
        // Instead we want to use all of the bits of a float to create the hash value
        // So we cast the address of val to a pointer to an equivalent sized unsigned int
        // This allows us to read the actual bit representation of a float type
        //

ViktorHofer · 2020-12-08T14:51:17Z

// Auto-generated message

69e114c which was merged 12/7 removed the intermediate src/coreclr/src/ folder. This PR needs to be updated as it touches files in that directory which causes conflicts.

To update your commits you can use this bash script: https://gist.github.com/ViktorHofer/6d24f62abdcddb518b4966ead5ef3783. Feel free to use the comment section of the gist to improve the script for others.

AndyAyersMS · 2021-01-23T03:07:28Z

@briansull anything holding this up other than a review?

briansull · 2021-02-23T22:14:48Z

This is ready for checkin again:
@dotnet/jit-contrib PTAL
@tannergooding

briansull · 2021-02-23T22:18:05Z

ARM32:

Top file regressions (bytes):
         294 : System.Data.Common.dasm (0.03% of base)
Top file improvements (bytes):
       -3200 : System.Private.CoreLib.dasm (-0.10% of base)
        -496 : Microsoft.VisualBasic.Core.dasm (-0.12% of base)
        -344 : System.Runtime.Numerics.dasm (-0.49% of base)
        -170 : System.Drawing.Common.dasm (-0.05% of base)
         -98 : Microsoft.Diagnostics.Tracing.TraceEvent.dasm (-0.00% of base)
         -28 : FSharp.Core.dasm (-0.00% of base)
         -26 : System.Net.Http.dasm (-0.00% of base)
         -20 : System.Private.Xml.dasm (-0.00% of base)
         -18 : System.Drawing.Primitives.dasm (-0.05% of base)
         -14 : Microsoft.CSharp.dasm (-0.00% of base)
         -12 : System.Linq.Parallel.dasm (-0.00% of base)
         -10 : System.Speech.dasm (-0.00% of base)
         -10 : System.Net.Http.WinHttpHandler.dasm (-0.02% of base)
          -6 : Newtonsoft.Json.dasm (-0.00% of base)

ARM64:

Top file regressions (bytes):
          16 : Microsoft.Diagnostics.Tracing.TraceEvent.dasm (0.00% of base)
          12 : System.Runtime.Numerics.dasm (0.01% of base)
           8 : Microsoft.VisualBasic.Core.dasm (0.00% of base)
           8 : System.Speech.dasm (0.00% of base)

Top file improvements (bytes):
        -412 : System.Private.CoreLib.dasm (-0.01% of base)
         -20 : System.Data.Common.dasm (-0.00% of base)
         -12 : System.Drawing.Common.dasm (-0.00% of base)
          -4 : Microsoft.CSharp.dasm (-0.00% of base)

src/coreclr/jit/gentree.cpp

tannergooding · 2021-02-23T22:28:23Z

src/coreclr/jit/morph.cpp

Should there be an issue tracking fixing this so that LGN2DBL properly sets TYP_DOUBLE?

No, I will change the comment to be more clear here.

For this case the tree's type is TYP_FLOAT and the call to fgMorphCastIntoHelper just leaves the return type as is. It assumes that the caller knows what he is doing. Since this is the code that is deciding to use this helper call and with my change we are now just fixing the return type to be correct.

tannergooding · 2021-02-23T22:29:06Z

src/coreclr/jit/morph.cpp

I don't think we use the FPU stack anywhere except for returns on 32-bit...

tannergooding · 2021-02-23T22:30:20Z

src/coreclr/jit/optcse.cpp

Do we need diffs for x86/x64 as well?

X64 Asm Diffs:

Top file regressions (bytes): 239 : Microsoft.Diagnostics.Tracing.TraceEvent.dasm (0.01% of base) 123 : Microsoft.VisualBasic.Core.dasm (0.03% of base) 99 : System.Speech.dasm (0.03% of base) 75 : System.Runtime.Numerics.dasm (0.10% of base) 46 : System.Drawing.Common.dasm (0.01% of base) 7 : System.Net.Http.WinHttpHandler.dasm (0.01% of base) 4 : System.Private.Xml.dasm (0.00% of base) Top file improvements (bytes): -61 : System.Private.CoreLib.dasm (-0.00% of base) -4 : System.Data.Common.dasm (-0.00% of base) -4 : Microsoft.CSharp.dasm (-0.00% of base) -1 : System.Net.Http.dasm (-0.00% of base)

X86 Asm Diffs:

Top file regressions (bytes): 31 : System.Drawing.Primitives.dasm (0.10% of base) 24 : Microsoft.VisualBasic.Core.dasm (0.01% of base) 18 : System.ComponentModel.TypeConverter.dasm (0.01% of base) 11 : System.Drawing.Common.dasm (0.00% of base) 11 : System.Linq.Expressions.dasm (0.00% of base) 10 : Microsoft.CodeAnalysis.CSharp.dasm (0.00% of base) 4 : System.Private.DataContractSerialization.dasm (0.00% of base) Top file improvements (bytes): -41 : System.Speech.dasm (-0.01% of base) -27 : System.Private.CoreLib.dasm (-0.00% of base) -13 : System.Data.Common.dasm (-0.00% of base) -9 : FSharp.Core.dasm (-0.00% of base) -3 : Microsoft.Diagnostics.Tracing.TraceEvent.dasm (-0.00% of base) -2 : System.Linq.Parallel.dasm (-0.00% of base)

tannergooding

LGTM

Update the gtCostEx and gtCostSz values of float and double constants for Arm and Arm64 Fix the gtCosts for 16-bit ARM32 immediate values

briansull · 2021-02-26T06:46:33Z

/azp run runtime-coreclr jitstress

azure-pipelines · 2021-02-26T06:46:49Z

Azure Pipelines successfully started running 1 pipeline(s).

briansull · 2021-02-26T18:56:01Z

All JitStress failures are due to
#48727

briansull · 2021-02-26T18:56:33Z

This is ready for checkin
@dotnet/jit-contrib PTAL

briansull · 2021-02-26T21:53:21Z

Sample Diff with a CSE

AndyAyersMS

Can you run SPMI diffs (on benchmarks, in particular)?

AndyAyersMS · 2021-02-26T22:01:39Z

src/coreclr/jit/morph.cpp

    else if (((tree->gtFlags & GTF_UNSIGNED) == 0) && (srcType == TYP_LONG) && varTypeIsFloating(dstType))
    {
-        return fgMorphCastIntoHelper(tree, CORINFO_HELP_LNG2DBL, oper);
+        oper = fgMorphCastIntoHelper(tree, CORINFO_HELP_LNG2DBL, oper);


Is this part of this change or some other fix?

It is required. CSE will assert because without this we can make a CSE with mismatched types (TYP_FLOAT and TYP_DOUBLE)

AndyAyersMS · 2021-02-26T22:04:11Z

src/coreclr/jit/optcse.cpp


-        // We don't shared small offset constants when we require a reloc
+        // We don't share small offset constants when we require a reloc
+        //


Maybe explain why not? Is it a legality thing or a profitability thing?

It would be incorrect to CSE,,
You can't eliminate a reloc using a CSE and adding an offset to a different constant.

briansull · 2021-02-26T22:28:41Z

Here is a diff from the SPMI Benchmarks where we hoist a Load of a double constant out of a loop using a CSE

AndyAyersMS · 2021-02-26T22:37:21Z

How many diffs are there in benchmarks?

Can you share out the details so we can cross-correlate with perf improvements we might see in the lab?

AndyAyersMS

Changes look good.

Would still like to see a benchmarks diff summary. Presumably you have the data lying around already?

briansull · 2021-02-27T00:23:42Z

PerfScore Benchmark Improvements:

Top method improvements (percentages):
       -4.30 (-16.20% of base) : SIMD\RayTracer\RayTracer\RayTracer.dasm - <>c:<.cctor>b__3_4(Vector):Color:this
       -4.30 (-16.20% of base) : SIMD\RayTracer\RayTracer\RayTracer.dasm - <>c:<.cctor>b__3_7(Vector):Color:this
      -76.85 (-9.26% of base) : Bytemark\Bytemark\Bytemark.dasm -Fourier:DoFPUTransIteration(System.Double[],System.Double[],int):long
     -142.40 (-7.43% of base) : Bytemark\Bytemark\Bytemark.dasm - NeuralJagged:randomize_wts()
       -7.40 (-7.30% of base) : Benchstones\BenchF\NewtR\NewtR\NewtR.dasm - Benchstone.BenchF.NewtR:Bench():bool
     -142.00 (-6.66% of base) : Bytemark\Bytemark\Bytemark.dasm - Neural:randomize_wts()
       -7.40 (-6.01% of base) : Benchstones\BenchF\Secant\Secant\Secant.dasm - Benchstone.BenchF.Secant:Bench():bool
       -4.30 (-5.95% of base) : SIMD\RayTracer\RayTracer\RayTracer.dasm - Color:Brightness():float:this
       -5.80 (-5.84% of base) : Math\Functions\Functions\Functions.dasm - Functions.MathTests:SqrtDoubleTest()
      -77.80 (-5.07% of base) : Benchstones\BenchF\Lorenz\Lorenz\Lorenz.dasm - Benchstone.BenchF.Lorenz:Bench():bool
     -243.40 (-5.04% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm -Algorithms.ScalarFloatRenderer:RenderSingleThreadedNoADT(float,float,float,float,float):this
       -4.90 (-5.03% of base) : Math\Functions\Functions\Functions.dasm - Functions.MathTests:AbsSingleTest()
       -4.90 (-5.03% of base) : Math\Functions\Functions\Functions.dasm - Functions.MathTests:AbsDoubleTest()
     -243.00 (-5.02% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - Algorithms.ScalarFloatRenderer:RenderSingleThreadedWithADT(float,float,float,float,float):this
     -243.40 (-5.02% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - Algorithms.ScalarDoubleRenderer:RenderSingleThreadedNoADT(float,float,float,float,float):this
     -243.00 (-5.00% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - Algorithms.ScalarDoubleRenderer:RenderSingleThreadedWithADT(float,float,float,float,float):this
      -13.47 (-4.95% of base) : Benchstones\BenchF\Bisect\Bisect\Bisect.dasm - Benchstone.BenchF.Bisect:Inner(byref,byref,byref,byref)
      -33.80 (-4.80% of base) : SIMD\RayTracer\RayTracer\RayTracer.dasm - Color:ChangeHue(float):this
      -27.30 (-4.65% of base) : Benchstones\BenchF\NewtE\NewtE\NewtE.dasm - Benchstone.BenchF.NewtE:Bench():bool
       -5.80 (-4.42% of base) : Math\Functions\Functions\Functions.dasm - Functions.MathTests:SqrtSingleTest()
     -107.80 (-4.24% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - <>c__DisplayClass4_0:<RenderMultiThreadedWithADT>b__0(int):this (2 methods)
     -107.40 (-4.23% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - <>c__DisplayClass5_0:<RenderMultiThreadedNoADT>b__0(int):this (2 methods)
       -4.90 (-4.18% of base) : Math\Functions\Functions\Functions.dasm - Functions.MathTests:RoundSingleTest()
       -4.90 (-4.18% of base) : Math\Functions\Functions\Functions.dasm - Functions.MathTests:CeilingSingleTest()
       -4.90 (-4.18% of base) : Math\Functions\Functions\Functions.dasm - Functions.MathTests:FloorSingleTest()
       -4.90 (-4.18% of base) : Math\Functions\Functions\Functions.dasm - Functions.MathTests:CeilingDoubleTest()
       -4.90 (-4.18% of base) : Math\Functions\Functions\Functions.dasm - Functions.MathTests:FloorDoubleTest()
       -4.90 (-4.18% of base) : Math\Functions\Functions\Functions.dasm - Functions.MathTests:RoundDoubleTest()
       -6.30 (-3.53% of base) : Benchstones\BenchF\Bisect\Bisect\Bisect.dasm - Benchstone.BenchF.Bisect:Bench():bool
     -213.57 (-3.40% of base) : Benchstones\BenchF\DMath\DMath\DMath.dasm - Benchstone.BenchF.DMath:Bench(int):bool
       -1.90 (-3.34% of base) : SciMark\SciMark\SciMark.dasm - SciMark2.FFT:num_flops(int):double
      -18.25 (-2.67% of base) : Burgers\Burgers\Burgers.dasm - Burgers:Main():int
      -31.65 (-2.13% of base) : Burgers\Burgers\Burgers.dasm - Burgers:GetCalculated2(int,int,double,double,double,System.Double[]):System.Double[]
      -31.25 (-2.13% of base) : Burgers\Burgers\Burgers.dasm - Burgers:GetCalculated0(int,int,double,double,double,System.Double[]):System.Double[]
      -31.25 (-2.10% of base) : Burgers\Burgers\Burgers.dasm - Burgers:GetCalculated1(int,int,double,double,double,System.Double[]):System.Double[]
       -1.90 (-1.71% of base) : FractalPerf\FractalPerf\FractalPerf.dasm - FractalPerf.Fractal:.ctor(double,double,double,double):this
      -79.60 (-1.50% of base) : Benchstones\BenchF\LLoops\LLoops\LLoops.dasm - Benchstone.BenchF.LLoops:Init():this
       -3.10 (-1.38% of base) : SciMark\SciMark\SciMark.dasm - SciMark2.kernel:measureSOR(int,double,SciMark2.Random):double
      -49.30 (-1.19% of base) : Bytemark\Bytemark\Bytemark.dasm - NeuralJagged:read_data_file():this
       -5.10 (-1.16% of base) : Bytemark\Bytemark\Bytemark.dasm - Fourier:Run():double:this
      -49.30 (-0.92% of base) : Bytemark\Bytemark\Bytemark.dasm - Neural:read_data_file():this
      -43.20 (-0.85% of base) : BenchmarksGame\mandelbrot\mandelbrot-2\mandelbrot-2.dasm - BenchmarksGame.Mandelbrot_2:DoBench(int,System.IO.MemoryStream,bool)
      -59.70 (-0.85% of base) : Bytemark\Bytemark\Bytemark.dasm - Huffman:DoHuffIteration(System.Byte[],System.Byte[],System.Byte[],int,int,huff_node[]):long
       -2.55 (-0.84% of base) : SIMD\RayTracer\RayTracer\RayTracer.dasm - RayTracer:Shade(ISect,Scene,int):Color:this
       -1.40 (-0.69% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - Algorithms.VectorDoubleStrictRenderer:RenderMultiThreadedWithADT(float,float,float,float,float):this
       -1.40 (-0.69% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - Algorithms.VectorDoubleStrictRenderer:RenderMultiThreadedNoADT(float,float,float,float,float):this
       -1.40 (-0.69% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - Algorithms.VectorDoubleRenderer:RenderMultiThreadedNoADT(float,float,float,float,float):this
       -1.40 (-0.69% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - Algorithms.VectorDoubleRenderer:RenderMultiThreadedWithADT(float,float,float,float,float):this
      -28.13 (-0.56% of base) : Burgers\Burgers\Burgers.dasm - Burgers:GetCalculated3(int,int,double,double,double,System.Double[]):System.Double[]
       -8.60 (-0.54% of base) : Benchstones\BenchF\Whetsto\Whetsto\Whetsto.dasm - Benchstone.BenchF.Whetsto:Bench():bool
       -2.30 (-0.48% of base) : Benchstones\BenchF\Regula\Regula\Regula.dasm - Benchstone.BenchF.Regula:Inner(byref,byref,double,double,int,byref)
       -5.50 (-0.41% of base) : BenchmarksGame\n-body\n-body-3\n-body-3.dasm - BenchmarksGame.NBodySystem:Energy():double:this
       -4.50 (-0.36% of base) : Bytemark\Bytemark\Bytemark.dasm - ByteMark:bench_with_confidence(int,byref,byref,byref):int
      -53.70 (-0.34% of base) : Benchstones\BenchF\FFT\FFT\FFT.dasm - Benchstone.BenchF.FFT:FastFourierT(System.Double[],System.Double[],int)
       -1.40 (-0.02% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - Algorithms.VectorDoubleRenderer:RenderSingleThreadedNoADT(float,float,float,float,float):this
       -1.40 (-0.02% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - Algorithms.VectorDoubleStrictRenderer:RenderSingleThreadedNoADT(float,float,float,float,float):this

Dotnet-GitSync-Bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Nov 9, 2020

tannergooding reviewed Nov 10, 2020

View reviewed changes

EgorBo mentioned this pull request Nov 13, 2020

Double constants usage in a loop can be CSEed #35257

Open

briansull force-pushed the cse-float-constants branch from 8c088cd to f45d38b Compare November 20, 2020 20:00

briansull mentioned this pull request Dec 10, 2020

Allow GT_CNS_DBL to be a CSE candidate #34557

Closed

briansull force-pushed the cse-float-constants branch from f45d38b to 2dbc6d2 Compare January 4, 2021 17:50

JulieLeeMSFT assigned briansull Feb 8, 2021

JulieLeeMSFT added this to the 6.0.0 milestone Feb 8, 2021

JulieLeeMSFT linked an issue Feb 8, 2021 that may be closed by this pull request

Allow GT_CNS_DBL to be a CSE candidate #34557

Closed

briansull force-pushed the cse-float-constants branch from 2dbc6d2 to b62e1fa Compare February 23, 2021 07:00

tannergooding reviewed Feb 23, 2021

View reviewed changes

src/coreclr/jit/gentree.cpp Outdated Show resolved Hide resolved

tannergooding reviewed Feb 23, 2021

View reviewed changes

tannergooding approved these changes Feb 23, 2021

View reviewed changes

briansull added 2 commits February 25, 2021 20:53

Implement CSE for constants of type float and double

34742e9

Update the gtCostEx and gtCostSz values of float and double constants for Arm and Arm64 Fix the gtCosts for 16-bit ARM32 immediate values

Code Review feedback

f101bb5

briansull force-pushed the cse-float-constants branch from b62e1fa to f101bb5 Compare February 26, 2021 06:45

AndyAyersMS reviewed Feb 26, 2021

View reviewed changes

AndyAyersMS approved these changes Feb 26, 2021

View reviewed changes

briansull merged commit 70676fd into dotnet:master Feb 27, 2021

briansull deleted the cse-float-constants branch February 27, 2021 03:13

This was referenced Mar 1, 2021

ARM64 - Assertion failed 'isSharedConst' in 'Test:Main():int' during 'Optimize Valnum CSEs' #48895

Closed

Test failure : JIT/Methodical/fp/exgen/10w5d_cs_ro/10w5d_cs_ro.sh #48892

Closed

DrewScoggins mentioned this pull request Mar 4, 2021

[Perf] Regression in floating point math tests #49158

Open

JulieLeeMSFT mentioned this pull request Mar 26, 2021

What's new in .NET 6 Preview 3 dotnet/core#5890

Closed

ghost locked as resolved and limited conversation to collaborators Mar 29, 2021

Enable CSE for floating-point constants #44419

Enable CSE for floating-point constants #44419

Uh oh!

Conversation

briansull commented Nov 9, 2020

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tannergooding commented Nov 10, 2020

Uh oh!

briansull commented Nov 10, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ViktorHofer commented Dec 8, 2020

Uh oh!

AndyAyersMS commented Jan 23, 2021

Uh oh!

briansull commented Feb 23, 2021

Uh oh!

briansull commented Feb 23, 2021

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

briansull Feb 26, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tannergooding left a comment

Choose a reason for hiding this comment

Uh oh!

briansull commented Feb 26, 2021

Uh oh!

azure-pipelines bot commented Feb 26, 2021

Uh oh!

briansull commented Feb 26, 2021

Uh oh!

briansull commented Feb 26, 2021

Uh oh!

briansull commented Feb 26, 2021

Uh oh!

AndyAyersMS left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

briansull Feb 26, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

briansull commented Feb 26, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

AndyAyersMS commented Feb 26, 2021

Uh oh!

AndyAyersMS left a comment

Choose a reason for hiding this comment

Uh oh!

briansull commented Feb 27, 2021

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

briansull commented Nov 10, 2020 •

edited

Loading

briansull Feb 26, 2021 •

edited

Loading

briansull Feb 26, 2021 •

edited

Loading

briansull commented Feb 26, 2021 •

edited

Loading