Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Disable folding of implementation-defined casts #53782

Merged
merged 11 commits into from
Jul 7, 2021
Merged

Disable folding of implementation-defined casts #53782

merged 11 commits into from
Jul 7, 2021

Conversation

SingleAccretion
Copy link
Contributor

@SingleAccretion SingleAccretion commented Jun 6, 2021 •
edited
Loading

With #51440 merged, it is now easy to do the intended second part of that PR and disable folding of all implementation-defined float/double -> integer casts. This is a temporary solution to #47478, but without it I will not be able to move forward with the casts work as a lot of this has so far been masked by the fact that problematic conversions were being morphed into calls and thus not getting folded.

For the diffs, the expectation is that there will be some regressions in coreclr_tests.

Windows x64 
Running asm diffs of coreclr_tests.pmi.windows.x64.checked.mch

Summary of Code Size diffs:
(Lower is better)

Total bytes of base: 49048
Total bytes of diff: 50724
Total bytes of delta: 1676 (3.42% of base)
    diff is a regression.


Top method regressions (bytes):
         116 (39.59% of base) - FloatOvfToInt:TestValuesFloatByte():int
         116 (39.59% of base) - FloatOvfToInt:TestValuesDoubleByte():int
         110 (37.54% of base) - FloatOvfToInt:TestValuesDoubleShort():int
         110 (37.54% of base) - FloatOvfToInt:TestValuesFloatShort():int
          95 (30.16% of base) - FloatOvfToInt:TestValuesDoubleInt():int
          95 (30.16% of base) - FloatOvfToInt:TestValuesFloatInt():int
          76 (133.33% of base) - ILGEN_0x44f8f663:Method_0x2e75f3b4(byte,long,short):int
          72 (36.00% of base) - ILGEN_0x1ad7535f:Method_0x2ddbcfb4(short):int
          59 ( 0.23% of base) - ILGEN_0x372a9ae6:Method_0xdc6ff1a4(byte,byte,int,long,ushort,double,long,long):int
          49 (13.14% of base) - ILGEN_0xe696696:main():int
          42 (15.05% of base) - ILGEN_0xc53351bb:Method_0x6cf4(short,byte,int,long,int,long,short,short,ubyte,long,float,ubyte,long,ubyte):float
          42 (700.00% of base) - ILGEN_1806854430:main():int
          41 (33.61% of base) - FloatOvfToInt:TestValuesFloatByte():int
          41 (33.61% of base) - FloatOvfToInt:TestValuesDoubleByte():int
          39 (31.97% of base) - FloatOvfToInt:TestValuesDoubleShort():int
          39 (31.97% of base) - FloatOvfToInt:TestValuesFloatShort():int
          38 (19.29% of base) - ILGEN_622380794:main():int
          33 (27.50% of base) - ILGEN_0x11c02e62:Method_0x7f2e741b(ubyte,ubyte,float,float,int):int
          32 (25.00% of base) - FloatOvfToInt:TestValuesFloatInt():int
          32 (25.00% of base) - FloatOvfToInt:TestValuesDoubleInt():int

Top method improvements (bytes):
         -22 (-4.14% of base) - ILGEN_0x7971e09f:Method_0xd939d2c1(int,int):int
         -10 (-0.43% of base) - ILGEN_0x537f7b0:Method_0x323f83b5(double,byte,int,short,ubyte,double,short,ushort,double,long,int,long,ushort,long,int):float
          -4 (-5.71% of base) - C:L(float):float
          -3 (-5.77% of base) - ConvTest:test4():long
          -3 (-5.77% of base) - ConvTest:test4():long
          -2 (-0.42% of base) - FloatOvfToInt:TestValuesFloatLong():int
          -2 (-0.42% of base) - FloatOvfToInt:TestValuesDoubleLong():int
          -1 (-0.35% of base) - ILGEN_0x19b572e2:Method_0x1a700c9c(double,byte):int
          -1 (-0.28% of base) - DevDiv_544985:Test(int,int,int,short,short,short):int

Top method regressions (percentages):
          42 (700.00% of base) - ILGEN_1806854430:main():int
          76 (133.33% of base) - ILGEN_0x44f8f663:Method_0x2e75f3b4(byte,long,short):int
           3 (50.00% of base) - ILGEN_0x34e68074:Method_0x1f01(float):int
          16 (44.44% of base) - ConvTest:test2():short
          16 (44.44% of base) - ConvTest:test1():byte
          16 (44.44% of base) - ConvTest:test7():ubyte
          16 (44.44% of base) - ConvTest:test7():ubyte
          16 (44.44% of base) - ConvTest:test2():short
          16 (44.44% of base) - ConvTest:test1():byte
         116 (39.59% of base) - FloatOvfToInt:TestValuesFloatByte():int
         116 (39.59% of base) - FloatOvfToInt:TestValuesDoubleByte():int
          14 (38.89% of base) - ConvTest:test8():ushort
          14 (38.89% of base) - ConvTest:test8():ushort
          17 (37.78% of base) - ILGEN_0x64f58bd0:Method_0x1bb0(int,long,float,int,int):int
         110 (37.54% of base) - FloatOvfToInt:TestValuesDoubleShort():int
         110 (37.54% of base) - FloatOvfToInt:TestValuesFloatShort():int
          72 (36.00% of base) - ILGEN_0x1ad7535f:Method_0x2ddbcfb4(short):int
          41 (33.61% of base) - FloatOvfToInt:TestValuesFloatByte():int
          41 (33.61% of base) - FloatOvfToInt:TestValuesDoubleByte():int
          39 (31.97% of base) - FloatOvfToInt:TestValuesDoubleShort():int

Top method improvements (percentages):
          -3 (-5.77% of base) - ConvTest:test4():long
          -3 (-5.77% of base) - ConvTest:test4():long
          -4 (-5.71% of base) - C:L(float):float
         -22 (-4.14% of base) - ILGEN_0x7971e09f:Method_0xd939d2c1(int,int):int
         -10 (-0.43% of base) - ILGEN_0x537f7b0:Method_0x323f83b5(double,byte,int,short,ubyte,double,short,ushort,double,long,int,long,ushort,long,int):float
          -2 (-0.42% of base) - FloatOvfToInt:TestValuesFloatLong():int
          -2 (-0.42% of base) - FloatOvfToInt:TestValuesDoubleLong():int
          -1 (-0.35% of base) - ILGEN_0x19b572e2:Method_0x1a700c9c(double,byte):int
          -1 (-0.28% of base) - DevDiv_544985:Test(int,int,int,short,short,short):int

67 total methods with Code Size differences (9 improved, 58 regressed), 0 unchanged.
Windows x86 
Running asm diffs of coreclr_tests.pmi.windows.x86.checked.mch

Summary of Code Size diffs:
(Lower is better)

Total bytes of base: 50715
Total bytes of diff: 52539
Total bytes of delta: 1824 (3.60% of base)
    diff is a regression.


Top method regressions (bytes):
         106 (27.39% of base) - FloatOvfToInt:TestValuesDoubleByte():int
         106 (27.39% of base) - FloatOvfToInt:TestValuesDoubleShort():int
          96 (24.81% of base) - FloatOvfToInt:TestValuesFloatByte():int
          96 (24.81% of base) - FloatOvfToInt:TestValuesFloatShort():int
          94 (27.41% of base) - ILGEN_0x1ad7535f:Method_0x2ddbcfb4(short):int
          94 (142.42% of base) - ILGEN_0x68ef9cc3:Method_0x21f5(long):float
          86 (23.50% of base) - ILGEN_0xe696696:main():int
          81 (78.64% of base) - ILGEN_0x11c02e62:Method_0x7f2e741b(ubyte,ubyte,float,float,int):int
          78 (31.84% of base) - ILGEN_0xc53351bb:Method_0x6cf4(short,byte,int,long,int,long,short,short,ubyte,long,float,ubyte,long,ubyte):float
          50 (35.46% of base) - FloatOvfToInt:TestValuesDoubleShort():int
          50 (35.46% of base) - FloatOvfToInt:TestValuesDoubleByte():int
          48 (12.66% of base) - DevDiv_590771:ILGEN_METHOD(long,ushort,int,int,ushort,int):byte
          47 ( 8.58% of base) - FloatOvfToInt:TestValuesDoubleInt():int
          40 (28.37% of base) - FloatOvfToInt:TestValuesFloatShort():int
          40 (28.37% of base) - FloatOvfToInt:TestValuesFloatByte():int
          37 ( 6.75% of base) - FloatOvfToInt:TestValuesFloatInt():int
          36 (600.00% of base) - ILGEN_1806854430:main():int
          36 (13.90% of base) - ILGEN_0x19b572e2:Method_0x1a700c9c(double,byte):int
          34 ( 5.87% of base) - DevDiv_544985:Test(int,int,int,short,short,short):int
          32 (12.85% of base) - ILGEN_622380794:main():int

Top method improvements (bytes):
          -8 (-5.23% of base) - C:Main():int
          -8 (-7.02% of base) - C:L(float):float

Top method regressions (percentages):
          36 (600.00% of base) - ILGEN_1806854430:main():int
          24 (400.00% of base) - ILGEN_622380794:main():int
          94 (142.42% of base) - ILGEN_0x68ef9cc3:Method_0x21f5(long):float
          81 (78.64% of base) - ILGEN_0x11c02e62:Method_0x7f2e741b(ubyte,ubyte,float,float,int):int
          21 (75.00% of base) - ILGEN_0xb6f84258:Main():int
          14 (56.00% of base) - ConvTest:test8():ushort
          14 (56.00% of base) - ConvTest:test2():short
          14 (56.00% of base) - ConvTest:test8():ushort
          14 (56.00% of base) - ConvTest:test1():byte
          14 (56.00% of base) - ConvTest:test2():short
          14 (56.00% of base) - ConvTest:test7():ubyte
          14 (56.00% of base) - ConvTest:test1():byte
          14 (56.00% of base) - ConvTest:test7():ubyte
           3 (37.50% of base) - ILGEN_0x34e68074:Method_0x1f01(float):int
          50 (35.46% of base) - FloatOvfToInt:TestValuesDoubleShort():int
          50 (35.46% of base) - FloatOvfToInt:TestValuesDoubleByte():int
          78 (31.84% of base) - ILGEN_0xc53351bb:Method_0x6cf4(short,byte,int,long,int,long,short,short,ubyte,long,float,ubyte,long,ubyte):float
          40 (28.37% of base) - FloatOvfToInt:TestValuesFloatShort():int
          40 (28.37% of base) - FloatOvfToInt:TestValuesFloatByte():int
          94 (27.41% of base) - ILGEN_0x1ad7535f:Method_0x2ddbcfb4(short):int

Top method improvements (percentages):
          -8 (-7.02% of base) - C:L(float):float
          -8 (-5.23% of base) - C:Main():int

61 total methods with Code Size differences (2 improved, 59 regressed), 0 unchanged.
Linux x64 
Running asm diffs of coreclr_tests.pmi.Linux.x64.checked.mch

Summary of Code Size diffs:
(Lower is better)

Total bytes of base: 48680
Total bytes of diff: 50694
Total bytes of delta: 2014 (4.14% of base)
    diff is a regression.


Top method regressions (bytes):
         137 (49.46% of base) - FloatOvfToInt:TestValuesFloatByte():int
         137 (49.46% of base) - FloatOvfToInt:TestValuesDoubleByte():int
         131 (47.29% of base) - FloatOvfToInt:TestValuesDoubleShort():int
         131 (47.29% of base) - FloatOvfToInt:TestValuesFloatShort():int
         115 (37.95% of base) - FloatOvfToInt:TestValuesFloatInt():int
         115 (37.95% of base) - FloatOvfToInt:TestValuesDoubleInt():int
          78 (36.11% of base) - ILGEN_0x1ad7535f:Method_0x2ddbcfb4(short):int
          74 (132.14% of base) - ILGEN_0x44f8f663:Method_0x2e75f3b4(byte,long,short):int
          60 (15.67% of base) - ILGEN_0xe696696:main():int
          58 (52.73% of base) - FloatOvfToInt:TestValuesDoubleByte():int
          58 (52.73% of base) - FloatOvfToInt:TestValuesFloatByte():int
          56 (50.91% of base) - FloatOvfToInt:TestValuesFloatShort():int
          56 (50.91% of base) - FloatOvfToInt:TestValuesDoubleShort():int
          53 ( 0.22% of base) - ILGEN_0x372a9ae6:Method_0xdc6ff1a4(byte,byte,int,long,ushort,double,long,long):int
          52 (44.83% of base) - FloatOvfToInt:TestValuesFloatInt():int
          52 (44.83% of base) - FloatOvfToInt:TestValuesDoubleInt():int
          44 (733.33% of base) - ILGEN_1806854430:main():int
          42 (16.03% of base) - ILGEN_0xc53351bb:Method_0x6cf4(short,byte,int,long,int,long,short,short,ubyte,long,float,ubyte,long,ubyte):float
          38 (16.24% of base) - ILGEN_622380794:main():int
          34 (26.77% of base) - ILGEN_0x11c02e62:Method_0x7f2e741b(ubyte,ubyte,float,float,int):int

Top method improvements (bytes):
         -22 (-4.03% of base) - ILGEN_0x7971e09f:Method_0xd939d2c1(int,int):int
         -15 (-0.62% of base) - ILGEN_0x537f7b0:Method_0x323f83b5(double,byte,int,short,ubyte,double,short,ushort,double,long,int,long,ushort,long,int):float
          -8 (-13.11% of base) - C:L(float):float
          -4 (-0.83% of base) - FloatOvfToInt:TestValuesDoubleLong():int
          -4 (-0.83% of base) - FloatOvfToInt:TestValuesFloatLong():int
          -1 (-0.67% of base) - ILGEN_CLASS:ILGEN_METHOD(ubyte,long,long,ushort):long
          -1 (-0.29% of base) - DevDiv_544985:Test(int,int,int,short,short,short):int
          -1 (-0.32% of base) - ILGEN_0x19b572e2:Method_0x1a700c9c(double,byte):int

Top method regressions (percentages):
          44 (733.33% of base) - ILGEN_1806854430:main():int
          74 (132.14% of base) - ILGEN_0x44f8f663:Method_0x2e75f3b4(byte,long,short):int
          23 (69.70% of base) - ConvTest:test1():byte
          23 (69.70% of base) - ConvTest:test1():byte
          23 (69.70% of base) - ConvTest:test2():short
          23 (69.70% of base) - ConvTest:test2():short
          22 (66.67% of base) - ConvTest:test7():ubyte
          22 (66.67% of base) - ConvTest:test7():ubyte
          21 (63.64% of base) - ConvTest:test8():ushort
          21 (63.64% of base) - ConvTest:test8():ushort
          18 (54.55% of base) - ConvTest:test9():int
          18 (54.55% of base) - ConvTest:test11():int
          18 (54.55% of base) - ConvTest:test11():int
          18 (54.55% of base) - ConvTest:test9():int
          58 (52.73% of base) - FloatOvfToInt:TestValuesDoubleByte():int
          58 (52.73% of base) - FloatOvfToInt:TestValuesFloatByte():int
          56 (50.91% of base) - FloatOvfToInt:TestValuesFloatShort():int
          56 (50.91% of base) - FloatOvfToInt:TestValuesDoubleShort():int
           3 (50.00% of base) - ILGEN_0x34e68074:Method_0x1f01(float):int
         137 (49.46% of base) - FloatOvfToInt:TestValuesFloatByte():int

Top method improvements (percentages):
          -8 (-13.11% of base) - C:L(float):float
         -22 (-4.03% of base) - ILGEN_0x7971e09f:Method_0xd939d2c1(int,int):int
          -4 (-0.83% of base) - FloatOvfToInt:TestValuesDoubleLong():int
          -4 (-0.83% of base) - FloatOvfToInt:TestValuesFloatLong():int
          -1 (-0.67% of base) - ILGEN_CLASS:ILGEN_METHOD(ubyte,long,long,ushort):long
         -15 (-0.62% of base) - ILGEN_0x537f7b0:Method_0x323f83b5(double,byte,int,short,ubyte,double,short,ushort,double,long,int,long,ushort,long,int):float
          -1 (-0.32% of base) - ILGEN_0x19b572e2:Method_0x1a700c9c(double,byte):int
          -1 (-0.29% of base) - DevDiv_544985:Test(int,int,int,short,short,short):int

65 total methods with Code Size differences (8 improved, 57 regressed), 2 unchanged.
Linux ARM64 
Running asm diffs of coreclr_tests.pmi.Linux.arm64.checked.mch

Summary of Code Size diffs:
(Lower is better)

Total bytes of base: 51232
Total bytes of diff: 54080
Total bytes of delta: 2848 (5.56% of base)
    diff is a regression.


Top method regressions (bytes):
         192 (84.21% of base) - FloatOvfToInt:TestValuesDoubleShort():int
         192 (84.21% of base) - FloatOvfToInt:TestValuesFloatByte():int
         192 (84.21% of base) - FloatOvfToInt:TestValuesDoubleByte():int
         192 (84.21% of base) - FloatOvfToInt:TestValuesFloatShort():int
         108 (39.13% of base) - FloatOvfToInt:TestValuesFloatInt():int
         108 (35.53% of base) - ILGEN_0x1ad7535f:Method_0x2ddbcfb4(short):int
         108 (39.13% of base) - FloatOvfToInt:TestValuesDoubleInt():int
          80 (71.43% of base) - FloatOvfToInt:TestValuesFloatByte():int
          80 (71.43% of base) - FloatOvfToInt:TestValuesDoubleByte():int
          80 (71.43% of base) - FloatOvfToInt:TestValuesDoubleShort():int
          80 (71.43% of base) - FloatOvfToInt:TestValuesFloatShort():int
          64 (12.40% of base) - ILGEN_0x7971e09f:Method_0xd939d2c1(int,int):int
          60 (18.52% of base) - FloatOvfToInt:TestValuesFloatLong():int
          60 (18.52% of base) - FloatOvfToInt:TestValuesDoubleLong():int
          60 (14.15% of base) - ILGEN_0xe696696:main():int
          52 (260.00% of base) - ILGEN_1806854430:main():int
          44 (26.83% of base) - ILGEN_0x44f8f663:Method_0x2e75f3b4(byte,long,short):int
          44 (12.36% of base) - ILGEN_0xc53351bb:Method_0x6cf4(short,byte,int,long,int,long,short,short,ubyte,long,float,ubyte,long,ubyte):float
          44 (27.50% of base) - ILGEN_0x11c02e62:Method_0x7f2e741b(ubyte,ubyte,float,float,int):int
          40 ( 7.75% of base) - ILGEN_0x57cc69ce:Method_0x183df1b1():int

Top method improvements (bytes):
          -8 (-0.83% of base) - ILGEN_0x1bd95bae:Method_0x40637edd(int,byte,short,short,double,float,ubyte,float):short

Top method regressions (percentages):
          52 (260.00% of base) - ILGEN_1806854430:main():int
         192 (84.21% of base) - FloatOvfToInt:TestValuesDoubleShort():int
         192 (84.21% of base) - FloatOvfToInt:TestValuesFloatByte():int
         192 (84.21% of base) - FloatOvfToInt:TestValuesDoubleByte():int
         192 (84.21% of base) - FloatOvfToInt:TestValuesFloatShort():int
          80 (71.43% of base) - FloatOvfToInt:TestValuesFloatByte():int
          80 (71.43% of base) - FloatOvfToInt:TestValuesDoubleByte():int
          80 (71.43% of base) - FloatOvfToInt:TestValuesDoubleShort():int
          80 (71.43% of base) - FloatOvfToInt:TestValuesFloatShort():int
          28 (53.85% of base) - ILGEN_0x77e65197:Main():int
          20 (41.67% of base) - ConvTest:test2():short
          20 (41.67% of base) - ConvTest:test8():ushort
          20 (41.67% of base) - ConvTest:test1():byte
          20 (41.67% of base) - ConvTest:test1():byte
          20 (41.67% of base) - ConvTest:test7():ubyte
          20 (41.67% of base) - ConvTest:test8():ushort
          20 (41.67% of base) - ConvTest:test2():short
          20 (41.67% of base) - ConvTest:test7():ubyte
         108 (39.13% of base) - FloatOvfToInt:TestValuesFloatInt():int
         108 (39.13% of base) - FloatOvfToInt:TestValuesDoubleInt():int

Top method improvements (percentages):
          -8 (-0.83% of base) - ILGEN_0x1bd95bae:Method_0x40637edd(int,byte,short,short,double,float,ubyte,float):short

78 total methods with Code Size differences (1 improved, 77 regressed), 2 unchanged.

Running asm diffs of libraries_tests.pmi.Linux.arm64.checked.mch

Summary of Code Size diffs:
(Lower is better)

Total bytes of base: 932
Total bytes of diff: 968
Total bytes of delta: 36 (3.86% of base)
    diff is a regression.


Top method regressions (bytes):
          36 ( 5.29% of base) - ManagedTests.DynamicCSharp.Conformance.dynamic.dynamicType.conversions.numeric002.numeric002.Program:MainMethod(System.String[]):int

Top method regressions (percentages):
          36 ( 5.29% of base) - ManagedTests.DynamicCSharp.Conformance.dynamic.dynamicType.conversions.numeric002.numeric002.Program:MainMethod(System.String[]):int

1 total methods with Code Size differences (0 improved, 1 regressed), 6 unchanged.
Linux ARM 
Running asm diffs of coreclr_tests.pmi.Linux.arm.checked.mch

Summary of Code Size diffs:
(Lower is better)

Total bytes of base: 49502
Total bytes of diff: 52992
Total bytes of delta: 3490 (7.05% of base)
    diff is a regression.


Top method regressions (bytes):
         294 (61.00% of base) - FloatOvfToInt:TestValuesDoubleInt():int
         216 (55.96% of base) - FloatOvfToInt:TestValuesFloatInt():int
         198 (43.23% of base) - FloatOvfToInt:TestValuesDoubleShort():int
         198 (43.23% of base) - FloatOvfToInt:TestValuesDoubleByte():int
         154 (42.78% of base) - FloatOvfToInt:TestValuesFloatShort():int
         154 (42.78% of base) - FloatOvfToInt:TestValuesFloatByte():int
         122 (26.18% of base) - ILGEN_0x1ad7535f:Method_0x2ddbcfb4(short):int
          96 (106.67% of base) - ILGEN_0x68ef9cc3:Method_0x21f5(long):float
          86 (34.96% of base) - ILGEN_622380794:main():int
          84 (62.69% of base) - ILGEN_0x11c02e62:Method_0x7f2e741b(ubyte,ubyte,float,float,int):int
          84 (28.77% of base) - ILGEN_0xc53351bb:Method_0x6cf4(short,byte,int,long,int,long,short,short,ubyte,long,float,ubyte,long,ubyte):float
          80 (54.79% of base) - FloatOvfToInt:TestValuesDoubleByte():int
          80 (54.79% of base) - FloatOvfToInt:TestValuesDoubleShort():int
          76 (49.35% of base) - FloatOvfToInt:TestValuesDoubleInt():int
          70 ( 0.29% of base) - ILGEN_0x372a9ae6:Method_0xdc6ff1a4(byte,byte,int,int,ushort,double,long,long):int
          66 (16.10% of base) - ILGEN_0xe696696:main():int
          66 (1,100.00% of base) - ILGEN_1806854430:main():int
          66 (47.83% of base) - FloatOvfToInt:TestValuesFloatInt():int
          66 (12.94% of base) - ILGEN_0x7971e09f:Method_0xd939d2c1(int,int):int
          60 (1,000.00% of base) - ILGEN_622380794:main():int

Top method improvements (bytes):
         -10 (-11.36% of base) - C:L(float):float

Top method regressions (percentages):
          66 (1,100.00% of base) - ILGEN_1806854430:main():int
          60 (1,000.00% of base) - ILGEN_622380794:main():int
          12 (150.00% of base) - ILGEN_0x34e68074:Method_0x1f01(float):int
          44 (146.67% of base) - ILGEN_0xb6f84258:Main():int
          96 (106.67% of base) - ILGEN_0x68ef9cc3:Method_0x21f5(long):float
          84 (62.69% of base) - ILGEN_0x11c02e62:Method_0x7f2e741b(ubyte,ubyte,float,float,int):int
         294 (61.00% of base) - FloatOvfToInt:TestValuesDoubleInt():int
          26 (59.09% of base) - ConvTest:test2():short
          26 (59.09% of base) - ConvTest:test8():ushort
          26 (59.09% of base) - ConvTest:test1():byte
          26 (59.09% of base) - ConvTest:test7():ubyte
         216 (55.96% of base) - FloatOvfToInt:TestValuesFloatInt():int
          80 (54.79% of base) - FloatOvfToInt:TestValuesDoubleByte():int
          80 (54.79% of base) - FloatOvfToInt:TestValuesDoubleShort():int
          24 (54.55% of base) - ConvTest:test1():byte
          24 (54.55% of base) - ConvTest:test2():short
          24 (54.55% of base) - ConvTest:test7():ubyte
          24 (54.55% of base) - ConvTest:test8():ushort
          76 (49.35% of base) - FloatOvfToInt:TestValuesDoubleInt():int
          66 (47.83% of base) - FloatOvfToInt:TestValuesFloatInt():int

Top method improvements (percentages):
         -10 (-11.36% of base) - C:L(float):float

66 total methods with Code Size differences (1 improved, 65 regressed), 1 unchanged.

Which the replays confirm.

Closes #51346.
Closes #47374.

@SingleAccretion
Enhance the FloatOvfToInt2 test to exercise VN paths
04f6198 
It is not very effective right now because the "bad" combinations of
types are morphed into helpers too early, but it will be helpful when
we enable folding outside of the importer for those cases too.
@SingleAccretion
Enhance the FloatOvfToInt2 test to cover importer
c8a2270 
It now fails on Windows x86, where many of the
previous similar failures were observed.
@dotnet-issue-labeler dotnet-issue-labeler bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Jun 6, 2021
@SingleAccretion SingleAccretion changed the title Disable folding for implementation-defined casts Disable folding of implementation-defined casts Jun 6, 2021
@SingleAccretion SingleAccretion force-pushed the Disable-Folding-For-Implementation-Defined-Casts branch from d47687b to d13b563 Compare June 6, 2021 16:51
@SingleAccretion SingleAccretion mentioned this pull request Jun 7, 2021
Closed
@SingleAccretion SingleAccretion marked this pull request as ready for review June 7, 2021 20:19
@tannergooding
Copy link
Member

CC. @dotnet/jit-contrib, community PR

@JulieLeeMSFT JulieLeeMSFT added this to the 6.0.0 milestone Jul 6, 2021
@JulieLeeMSFT
Copy link
Member

@kunalspathak PTAL for community PR.

kunalspathak
kunalspathak reviewed Jul 6, 2021
View reviewed changes
Copy link
Member

@kunalspathak kunalspathak left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Added some questions. Could you share some of the sample diffs that you see in the test (just to understand how this works underlying). Also, did you check asmdiffs on benchmarks/libraries?

src/coreclr/jit/valuenum.cpp Show resolved Hide resolved
src/coreclr/jit/valuenum.cpp Show resolved Hide resolved
@SingleAccretion
Copy link
Contributor Author

Also, did you check asmdiffs on benchmarks/libraries?

Yep, the PR's header is for results of running SPMI over all available collections, so, no diffs outside tests.

Could you share some of the sample diffs that you see in the test (just to understand how this works underlying).

Will recollect the diffs and share the samples.

@SingleAccretion
Copy link
Contributor Author

Case 1: ILGEN_0x44f8f663:Method_0x2e75f3b4, 76 (133.33% of base):

 G_M64081_IG01:
        sub      rsp, 40
-
+       vzeroupper
+
 G_M64081_IG02:
        movsx    rcx, cl
        xor      ecx, 5
@@ -25,22 +26,41 @@ G_M64081_IG02:        ; gcrefRegs=00000000 {}, byrefRegs=00000000 {}, byref, isz
        add      rcx, rax
        mov      rax, rcx
        shr      rax, 32
-       jne      SHORT G_M64081_IG03
+       jne      SHORT G_M64081_IG04
        xor      ecx, ecx
        call     CORINFO_HELP_OVERFLOW
        ; gcr arg pop 0
-       xor      ecx, ecx
-       call     CORINFO_HELP_OVERFLOW
+       vmovsd   xmm0, qword ptr [reloc @RWD00]
+       call     CORINFO_HELP_DBL2ULNG
        ; gcr arg pop 0
-       int3
-                                               ;; bbWeight=1    PerfScore 5.75
-G_M64081_IG03:        ; gcrefRegs=00000000 {}, byrefRegs=00000000 {}, byref
+       vxorps   xmm0, xmm0
+       vcvtsi2ss  xmm0, rax
+       vcvtss2sd xmm0, xmm0
+       call     CORINFO_HELP_DBL2ULNG
+       ; gcr arg pop 0
+       mov      rdx, 0xD1FFAB1E
+       imul     rax, rdx
+       mov      rdx, 0xD1FFAB1E
+       add      rax, rdx
+       jo       SHORT G_M64081_IG04
+       cmp      rax, 0xD1FFAB1E
+       jg       SHORT G_M64081_IG04
+       cmp      rax, 0xD1FFAB1E
+       jl       SHORT G_M64081_IG04
+                                               ;; bbWeight=1    PerfScore 23.83
+G_M64081_IG03:        ; , epilog, nogc, extend
+       add      rsp, 40
+       ret
+                                               ;; bbWeight=1    PerfScore 1.25
+G_M64081_IG04:        ; gcVars=0000000000000000 {}, gcrefRegs=00000000 {}, byrefRegs=00000000 {}, gcvars, byref
        call     CORINFO_HELP_OVERFLOW
        ; gcr arg pop 0
        int3
                                                ;; bbWeight=0    PerfScore 0.00
+RWD00          dq      4D95800D5C7F3082h       ; 5.66058935e+65

Here the early importer folding is blocked, due to the new checks in gtFoldExprConst:

     [ 1] 100 (0x064) nop
     [ 1] 101 (0x065) ldc.r8 5.6605893542014790e+65
     [ 2] 110 (0x06e) conv.u
-Folding long operator with constant nodes into a constant:
-               [000025] ------------              *  CAST      long <- ulong <- double
-               [000024] ------------              \--*  CNS_DBL   double 5.6605893542014790e+65
-Bashed to long constant:
-               [000025] ------------              *  CNS_INT   long   0
-
     [ 2] 111 (0x06f) conv.r4
-Folding fp operator with constant nodes into a fp constant:
-               [000026] ------------              *  CAST      float <- long
-               [000025] ------------              \--*  CNS_INT   long   0
-Bashed to fp constant:
-               [000026] ------------              *  CNS_DBL   float  0.0000000000000000
-
     [ 2] 112 (0x070) conv.u8
-Folding long operator with constant nodes into a constant:
-               [000027] ------------              *  CAST      long <- ulong <- float
-               [000026] ------------              \--*  CNS_DBL   float  0.0000000000000000
-Bashed to long constant:
-               [000027] ------------              *  CNS_INT   long   0
-
     [ 2] 113 (0x071) mul
     [ 1] 114 (0x072) ldc.r4 2281166340096.0000
     [ 2] 119 (0x077) ldc.r8 -1.4862075407321284e+46
@@ -475,7 +457,10 @@ STMT00009 (IL   ???...  ???)
                [000095] ---X--------                 \--*  ADD_ovfl  long
                [000028] ------------                    +--*  MUL       long
                [000018] ------------                    |  +--*  CNS_INT   long   -0xf925aa1447c2972
-               [000027] ------------                    |  \--*  CNS_INT   long   0
+               [000027] ------------                    |  \--*  CAST      long <- ulong <- float
+               [000026] ------------                    |     \--*  CAST      float <- long
+               [000025] ------------                    |        \--*  CAST      long <- ulong <- double
+               [000024] ------------                    |           \--*  CNS_DBL   double 5.6605893542014790e+65
                [000094] ---X--------                    \--*  SUB_ovfl  long

Case 2: ILGEN_1806854430:main():int, 42 (700.00% of base):

-       mov      eax, 100
-                                               ;; bbWeight=1    PerfScore 0.25
+       vcvttss2si rax, qword ptr [reloc @RWD00]
+       cmp      rax, 0xD1FFAB1E
+       setl     dl
+       movzx    rdx, dl
+       movsxd   rdx, edx
+       cmp      rdx, rax
+       setg     al
+       movzx    rax, al
+       cmp      eax, 0xD1FFAB1E
+       seta     al
+       movzx    rax, al
+       add      eax, 100
+                                               ;; bbWeight=1    PerfScore 12.00
 G_M58196_IG03:        ; , epilog, nogc, extend
        ret
                                                ;; bbWeight=1    PerfScore 1.00
+RWD00          dd      63E756A3h               ; 8.53488e+21

This time, local assertion propagation does not result in gtFoldExprConst folding the expressions in post-order morph:

@@ -518,56 +512,35 @@ Assertion prop in BB01:
 Constant Assertion: V00 == 8.53488e+21 index=#01, mask=0000000000000001
                [000020] ------------              *  CNS_DBL   float  8.5348814290707441e+21

-Folding long operator with constant nodes into a constant:
-               [000021] ------------              *  CAST      long <- float
-               [000020] -----+------              \--*  CNS_DBL   float  8.5348814290707441e+21
-Bashed to long constant:
-               [000021] ------------              *  CNS_INT   long   -0x8000000000000000
-
-Folding operator with constant nodes into a constant:
-               [000023] ------------              *  GT        int
-               [000022] -----+------              +--*  CNS_INT   long   0x45E23077
-               [000021] -----+------              \--*  CNS_INT   long   -0x8000000000000000
-Bashed to int constant:
-               [000023] ------------              *  CNS_INT   int    1
-
-Folding long operator with constant nodes into a constant:
-               [000025] ------------              *  CAST      long <- int
-               [000023] -----+------              \--*  CNS_INT   int    1
-Bashed to long constant:
-               [000025] ------------              *  CNS_INT   long   1

Case 3: ConvTest:test1():byte, 16 (44.44% of base):

 G_M30804_IG01:        ; gcrefRegs=00000000 {}, byrefRegs=00000000 {}, byref, nogc <-- Prolog IG
-       sub      rsp, 40
-                                               ;; bbWeight=1    PerfScore 0.25
+       push     rsi
+       sub      rsp, 32
+                                               ;; bbWeight=1    PerfScore 1.25
 G_M30804_IG02:        ; gcrefRegs=00000000 {}, byrefRegs=00000000 {}, byref
        mov      rcx, 0xD1FFAB1E
        mov      rcx, gword ptr [rcx]
@@ -21,17 +22,22 @@ G_M30804_IG02:        ; gcrefRegs=00000000 {}, byrefRegs=00000000 {}, byref
        call     hackishModuleName:hackishMethodName()
        ; gcrRegs -[rcx]
        ; gcr arg pop 0
-       xor      ecx, ecx
+       vcvttsd2si ecx, dword ptr [reloc @RWD00]
+       movsx    rsi, cl
+       mov      ecx, esi
        call     hackishModuleName:hackishMethodName()
        ; gcr arg pop 0
-       xor      eax, eax
-                                               ;; bbWeight=1    PerfScore 4.75
+       movsx    rax, sil
+                                               ;; bbWeight=1    PerfScore 12.00
 G_M30804_IG03:        ; , epilog, nogc, extend
-       add      rsp, 40
+       add      rsp, 32
+       pop      rsi
        ret
-                                               ;; bbWeight=1    PerfScore 1.25
+                                               ;; bbWeight=1    PerfScore 1.75
+RWD00          dq      FFF8000000000000h       ;    -nan(ind)

Here, VN block takes hold:

 N001 [000002]   CNS_DBL   -nan(ind) => $140 {DblCns[-nan(ind)]}
     VNForCastOper(int) is $41
-N002 [000013]   CAST      => $42 {IntCns 0x80000000}
-    VNForCastOper(byte) is $43
-N003 [000003]   CAST      => $44 {IntCns 0}
-N004 [000004]   LCL_VAR   V01 tmp1         d:1 => $44 {IntCns 0}
-N005 [000005]   ASG       => $44 {IntCns 0}
+N002 [000013]   CAST      => $180 {Cast($140, $41)}
+    VNForCastOper(byte) is $42
+N003 [000003]   CAST      => $181 {Cast($180, $42)}
+N004 [000004]   LCL_VAR   V01 tmp1         d:1 => $181 {Cast($180, $42)}
+N005 [000005]   ASG       => $181 {Cast($180, $42)}

 ***** BB01, STMT00001(after)
-N005 ( 10, 12) [000005] -A------R---              *  ASG       int    $44
-N004 (  1,  1) [000004] D------N----              +--*  LCL_VAR   int    V01 tmp1         d:1 $44
-N003 ( 10, 12) [000003] ------------              \--*  CAST      int <- byte <- int $44
-N002 (  9, 10) [000013] ------------                 \--*  CAST      int <- double $42
+N005 ( 10, 12) [000005] -A------R---              *  ASG       int    $181
+N004 (  1,  1) [000004] D------N----              +--*  LCL_VAR   int    V01 tmp1         d:1 $181
+N003 ( 10, 12) [000003] ------------              \--*  CAST      int <- byte <- int $181
+N002 (  9, 10) [000013] ------------                 \--*  CAST      int <- double $180
 N001 (  3,  4) [000002] ------------                    \--*  CNS_DBL   double -nan(ind) $140

This means no constant propagation in assertion prop:

-After constant propagation on [000003]:
-STMT00001 (IL 0x00A...  ???)
-N005 ( 10, 12) [000005] -A------R---              *  ASG       int    $44
-N004 (  1,  1) [000004] D------N----              +--*  LCL_VAR   int    V01 tmp1         d:1 $44
-               [000016] ------------              \--*  CNS_INT   int    0 $44
-optVNAssertionPropCurStmt morphed tree:
-N003 (  1,  3) [000005] -A------R---              *  ASG       int    $44
-N002 (  1,  1) [000004] D------N----              +--*  LCL_VAR   int    V01 tmp1         d:1 $44
-N001 (  1,  1) [000016] ------------              \--*  CNS_INT   int    0 $44
-
-After constant propagation on [000007]:
-STMT00002 (IL   ???...  ???)
-N003 ( 15,  7) [000008] --CXG-------              *  CALL      void   hackishModuleName.hackishMethodName $VN.Void
-               [000017] ------------ arg0 in rcx  \--*  CNS_INT   int    0 $44
-ReMorphing args for 8.CALL:
-argSlots=1, preallocatedArgCount=4, nextSlotNum=4, nextSlotByteOffset=32, outgoingArgSpaceSize=32
-ArgTable for 8.CALL after fgMorphArgs:
-fgArgTabEntry[arg 0 17.CNS_INT int (By ref), 1 reg: rcx, byteAlignment=8, lateArgInx=0, processed]
-
-optVNAssertionPropCurStmt morphed tree:
-N003 ( 15,  7) [000008] --CXG-------              *  CALL      void   hackishModuleName.hackishMethodName $VN.Void
-N002 (  1,  1) [000017] ------------ arg0 in rcx  \--*  CNS_INT   int    0 $44
-
-After constant propagation on [000015]:
-STMT00003 (IL 0x01A...  ???)
-N003 (  3,  4) [000009] ------------              *  RETURN    int    $182
-               [000018] ------------              \--*  CNS_INT   int    0 $44
-optVNAssertionPropCurStmt morphed tree:
-N002 (  2,  2) [000009] ------------              *  RETURN    int    $182
-N001 (  1,  1) [000018] ------------              \--*  CNS_INT   int    0 $44

Case 4: ConvTest:test4():long, -3 (-5.77% of base):

-; Lcl frame size = 40
+; Lcl frame size = 32

 G_M60593_IG01:        ; gcrefRegs=00000000 {}, byrefRegs=00000000 {}, byref, nogc <-- Prolog IG
-       sub      rsp, 40
-                                               ;; bbWeight=1    PerfScore 0.25
+       push     rsi
+       sub      rsp, 32
+                                               ;; bbWeight=1    PerfScore 1.25
 G_M60593_IG02:        ; gcrefRegs=00000000 {}, byrefRegs=00000000 {}, byref
        mov      rcx, 0xD1FFAB1E
        mov      rcx, gword ptr [rcx]
@@ -21,17 +22,21 @@ G_M60593_IG02:        ; gcrefRegs=00000000 {}, byrefRegs=00000000 {}, byref
        call     hackishModuleName:hackishMethodName()
        ; gcrRegs -[rcx]
        ; gcr arg pop 0
-       mov      rcx, 0xD1FFAB1E
+       vcvttsd2si rsi, qword ptr [reloc @RWD00]
+       mov      rcx, rsi
        call     hackishModuleName:hackishMethodName()
        ; gcr arg pop 0
-       mov      rax, 0xD1FFAB1E
-                                               ;; bbWeight=1    PerfScore 4.75
+       mov      rax, rsi
+                                               ;; bbWeight=1    PerfScore 11.75
 G_M60593_IG03:        ; , epilog, nogc, extend
-       add      rsp, 40
+       add      rsp, 32
+       pop      rsi
        ret
-                                               ;; bbWeight=1    PerfScore 1.25
+                                               ;; bbWeight=1    PerfScore 1.75
+RWD00          dq      FFF8000000000000h       ;    -nan(ind)
+

-; Total bytes of code 52, prolog size 4, PerfScore 11.45, instruction count 9, allocated bytes for code 52 (MethodHash=1dad134e) for method ConvTest:test4():long
+; Total bytes of code 49, prolog size 5, PerfScore 19.65, instruction count 12, allocated bytes for code 49 (MethodHash=1dad134e) for method ConvTest:test4():long

This is another case of no folding in VN:

-N002 [000003]   CAST      => $180 {LngCns:  0x8000000000000000}
-N003 [000004]   LCL_VAR   V01 tmp1         d:1 => $180 {LngCns:  0x8000000000000000}
-N004 [000005]   ASG       => $180 {LngCns:  0x8000000000000000}
+N002 [000003]   CAST      => $180 {Cast($140, $41)}
+N003 [000004]   LCL_VAR   V01 tmp1         d:1 => $180 {Cast($140, $41)}
+N004 [000005]   ASG       => $180 {Cast($140, $41)}

kunalspathak
kunalspathak approved these changes Jul 6, 2021
View reviewed changes
Copy link
Member

@kunalspathak kunalspathak left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM. Thanks for your contribution!

@SingleAccretion
Move the tests back to Pri1
f48608f 
Where they originally were.
@kunalspathak
Copy link
Member

@SingleAccretion - do you mind taking a look into failures to make sure they are not related to your changes?

@SingleAccretion
Copy link
Contributor Author

SingleAccretion commented Jul 7, 2021 •
edited
Loading

do you mind taking a look into failures to make sure they are not related to your changes?

@kunalspathak yes of course. Unfortunately the CI is not in the greatest of shapes right now :(.

  1. Build Android x64 Release AllSubsets_Mono_RuntimeTests - log, looks like an infrastructure problem:
[21:26:28] fail: Cannot find a device with app=package:net.dot.JIT_CodeGenBringUpTests, please check that a device is attached
XHarness exit code: 85 (ADB_DEVICE_ENUMERATION_FAILURE)
Test Harness Exitcode is : 85
  1. Build tvOS arm64 Release AllSubsets_Mono - restore failed:
.dotnet/sdk/6.0.100-preview.4.21255.9/NuGet.RestoreEx.targets(19,5): error : (NETCORE_ENGINEERING_TELEMETRY=Restore) Failed to download package 'Microsoft.Private.Intellisense.5.0.0-preview-20201009.2' from 'https://pkgs.dev.azure.com/dnceng/9ee6d478-d288-47f7-aacc-f6e6d082ae6d/_packaging/c9f8ac11-6bd8-4926-8306-f075241547f7/nuget/v3/flat2/microsoft.private.intellisense/5.0.0-preview-20201009.2/microsoft.private.intellisense.5.0.0-preview-20201009.2.nupkg'.
Response status code does not indicate success: 503 (Service Unavailable - TF10216: Azure DevOps services are currently unavailable. Try again later. Activity Id: 84ef9393-5dd8-434a-80b7-f2eb40580efb (DevOps Activity ID: 84EF9393-5DD8-434A-80B7-F2EB40580EFB)).
  1. CoreCLR Product Build OSX arm64 checked - log, another restore failure:
2021-07-06T20:40:41.9426000Z /Users/runner/work/1/s/.dotnet/sdk/6.0.100-preview.4.21255.9/NuGet.RestoreEx.targets(19,5): error : Failed to download package 'System.Linq.4.1.0-rc2-24027' from 'https://pkgs.dev.azure.com/dnceng/9ee6d478-d288-47f7-aacc-f6e6d082ae6d/_packaging/45bacae2-5efb-47c8-91e5-8ec20c22b4f8/nuget/v3/flat2/system.linq/4.1.0-rc2-24027/system.linq.4.1.0-rc2-24027.nupkg'. [/Users/runner/work/1/s/src/tests/Common/external/external.csproj] [/Users/runner/work/1/s/src/tests/build.proj]
  1. Libraries Test Run checked coreclr Linux_musl arm Release - log, this time a real test failure:
  Discovering: System.Text.Encoding.Tests (method display = ClassAndMethod, method display options = None)
  Discovered:  System.Text.Encoding.Tests (found 338 of 341 test cases)
  Starting:    System.Text.Encoding.Tests (parallel test collections = on, max threads = 4)
    System.Text.Tests.TranscodingStreamTests.Read_Span(bufferLength: 1) [FAIL]
      Assert.Equal() Failure
                                       ↓ (pos 4839671)
      Expected: ···lmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrst···
      Actual:   ···lmnopqrstuvwxyzabcdeFghijklmnopqrstuvwxyzabcdefghijklmnopqrst···
                                       ↑ (pos 4839671)

This looks like #47444. Seeing as this is a rare failure, maybe we can extract some useful information from this run...?

  1. Libraries Test Run release coreclr Linux x64 Debug - log, another test failure:
System.Net.Http.Functional.Tests.SocketsHttpHandler_HttpClientHandler_ConnectionPooling_Test.Http2_SmallConnectionTimeout_SubsequentRequestUsesDifferentConnection(timeoutPropertyName: "PooledConnectionLifetime") [FAIL]
      System.IO.IOException : Unable to write data to the transport connection: Broken pipe.

  1. runtime-dev-innerloop, build #20210706.55, runtime-dev-innerloop, Build OSX x64 release Runtime_Debug, runtime-staging (Build iOSSimulator x64 Release AllSubsets_Mono) - more restore failures.

  2. runtime-staging, many failures - log1, log2, log3 (there are others), not enough space on the device:

Standard Error:
adb: failed to install /datadisks/disk1/work/A1BB08F0/w/B3F70977/e/System.DirectoryServices.Protocols.Tests.apk: 
Exception occurred while executing:
android.os.ParcelableException: java.io.IOException: Requested internal only, but not enough space

  1. runtime-staging, an interpreter test failure - log

  2. runtime-staging, many System.Drawing.Imaging failures (this is Mono on Windows), PNSEs:

    System.Drawing.Imaging.Tests.MetafileTests.Ctor_InvalidEmfTypeIII_ThrowsArgumentException(emfType: 2) [FAIL]
      Assert.Throws() Failure
      Expected: typeof(System.ArgumentException)
      Actual:   typeof(System.TypeInitializationException): The type initializer for 'System.Drawing.DrawingCom' threw an exception.
      ---- System.TypeInitializationException : The type initializer for 'System.Drawing.DrawingCom' threw an exception.

A rich assortment to be sure...

"Real" failures seem to be HTTP one, and the interpreter one, though I find it hard to see how they could be caused by this change (note as well that it passed CI on first run, and no code changes have been committed since).

@kunalspathak kunalspathak merged commit 527e566 into dotnet:main Jul 7, 2021
@SingleAccretion SingleAccretion deleted the Disable-Folding-For-Implementation-Defined-Casts branch July 8, 2021 21:21
@SingleAccretion SingleAccretion mentioned this pull request Aug 5, 2021
Closed
@ghost ghost locked as resolved and limited conversation to collaborators Aug 7, 2021
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Reviewers

@kunalspathak kunalspathak kunalspathak approved these changes

Assignees

@SingleAccretion SingleAccretion

Labels
area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI
Projects
None yet
Milestone
6.0.0
Development

Successfully merging this pull request may close these issues.

System.Linq.Expressions tests are broken in CI System.Linq.Expressions tests failing on some PRs
4 participants
@SingleAccretion @tannergooding @JulieLeeMSFT @kunalspathak