JIT: Incorrect optimization of x & -x on x86

[jakobbotsch]

Description

The following program gives an invalid result when optimizing on x86.

Reproduction Steps

// Generated by Fuzzlyn v1.5 on 2022-03-16 11:07:55
// Run on X86 Windows
// Seed: 13083630089409025722
// Reduced from 309.3 KiB to 0.2 KiB in 00:06:28
// Debug: Outputs 1
// Release: Outputs 4294967297
public class Program
{
    public static long s_61 = -1;
    public static void Main()
    {
        s_61 &= -s_61;
        System.Console.WriteLine(s_61);
    }
}

Expected behavior

Same result with and without optimizations.

Actual behavior

Different result.

Regression?

No response

Known Workarounds

No response

Configuration

No response

Other information

Probably some interaction between long decomposition and the new optimization added in #66193? cc @Wraith2 @dotnet/jit-contrib

[ghost]

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

Issue Details

---

Description

The following program gives an invalid result when optimizing on x86.

Reproduction Steps

// Generated by Fuzzlyn v1.5 on 2022-03-16 11:07:55
// Run on X86 Windows
// Seed: 13083630089409025722
// Reduced from 309.3 KiB to 0.2 KiB in 00:06:28
// Debug: Outputs 1
// Release: Outputs 4294967297
public class Program
{
    public static long s_61 = -1;
    public static void Main()
    {
        s_61 &= -s_61;
        System.Console.WriteLine(s_61);
    }
}

Expected behavior

Same result with and without optimizations.

Actual behavior

Different result.

Regression?

No response

Known Workarounds

No response

Configuration

No response

Other information

Probably some interaction between long decomposition and the new optimization added in #66193? cc @Wraith2 @dotnet/jit-contrib

Author:	jakobbotsch
Assignees:	-
Labels:	area-CodeGen-coreclr, untriaged
Milestone:	-

[deeprobin]

https://godbolt.org/z/bfEx84ovj
Godbolt does not show me any blsi optimization at all.

[jakobbotsch]

https://godbolt.org/z/bfEx84ovj Godbolt does not show me any blsi optimization at all.

It's a new optimization added in #66193. It's not part of .NET 6.

[deeprobin]

https://godbolt.org/z/bfEx84ovj Godbolt does not show me any blsi optimization at all.

It's a new optimization added in #66193. It's not part of .NET 6.

As far as I know Godbolt always uses the daily builds, even if it says .NET 6.

Referring to C# Community Discord discussion with @EgorBo

[EgorBo]

@deeprobin I think it actually doesn't use daily builds, it uses some fixed version of .NET 7.0 daily build, probably a few month old

[hez2010]

it uses some fixed version of .NET 7.0 daily build, probably a few month old

No. Godbolt is currently using tag v6.0.1 as runtime, so it's .NET 6.

You can confirm the version here: https://github.com/compiler-explorer/infra/blob/main/bin/yaml/dotnet.yaml#L11

[hez2010]

https://godbolt.org/z/bfEx84ovj Godbolt does not show me any blsi optimization at all.

Note that your compile options are invalid. Godbolt only supports below compiler options/switches for dotnet languages:

• --targetos: operating system, available options are windows, linux and osx. Default value is based on current platform, on complier explorer it's linux.
• --targetarch: architecture, available options are arm, armel, arm64, x86 and x64. Default value is x64.
• --instruction-set: instruction set for compilation. Default value is sse2.
  • The allowable values for the --instruction-set option are described in the table below. Each architecture has a different set of valid instruction sets, and multiple instruction sets may be specified by separating the instructions sets by a ,. For example avx2,bmi,lzcnt
  • Available instruction sets for each architecture:
    • arm: N/A
    • armel: N/A
    • arm64: base, neon, aes, crc, dotprod, rdma, sha1, sha2, lse
    • x86: base, sse, sse2, sse3, ssse3, sse4.1, sse4.2, avx, avx2, aes, bmi, bmi2, fma, lzcnt, pclmul, popcnt, avxvnni
    • x64: base, sse, sse2, sse3, ssse3, sse4.1, sse4.2, avx, avx2, aes, bmi, bmi2, fma, lzcnt, pclmul, popcnt, avxvnni
• --singlemethodtypename: Type name for single method compilation
• --singlemethodname: Name for single method compilation
• --singlemethodindex: Index for single method compilation
• --singlemethodgenericarg: Generic arguments for single method compilation
• --codegenopt, --codegen-options: Code generation options passed to JIT
• -O, --optimize: Enable optimizations (default)
• --Od, --optimize-disabled: Disable optimizations
• --Os, --optimize-size: Optimize code for size
• --Ot, --optimize-time: Optimize code for speed

[Wraith2]

I'll take a look.

[Wraith2]

By x86 do you mean then 32 bit build specifically? Because I can't get that to build locally. Or are we just talking x86 as in not arm.

[jakobbotsch]

By x86 do you mean then 32 bit build specifically? Because I can't get that to build locally. Or are we just talking x86 as in not arm.

I mean the 32-bit build specifically.

[jakobbotsch]

I can't get that to build locally.

What errors are you getting? The standard build commands (passing -arch x86) work fine for me (including build.cmd -vs coreclr.sln -arch x86).