Fix bad folding

[SingleAccretion]

GT_CNS_INT of TYP_INT on 64 hosts has an implicit contract: the value returned by IconValue() must "fit" into 32 bit
signed integer, i. e. static_cast<int>(IconValue()) == IconValue(). gtFoldExprConst was failing to uphold this contract when the target was 32 bit and the host was 64 bit.

Fix this by always truncating before calling SetIconValue().

I also added a test that reproduces the bad behavior, but to be effective it needs to be CG'ed on a 64 bit machine for a 32 bit target. I am not sure how to make sure that happens...

This change is not zero-diff. There were ~10 cases of either bad codegen or missed folding in the SPMI collections (all available for win-64 -> win-x86 cross-compilation), though none in the product code.

Found while looking at #53983.

[sandreenko]

Sorry for a late response, the change looks good, but I would like to repro the issue.

I also added a test that reproduces the bad behavior, but to be effective it needs to be CG'ed on a 64 bit machine for a 32 bit target. I am not sure how to make sure that happens...

It should be pretty straightforward with this instruction: https://github.com/dotnet/runtime/blob/main/docs/workflow/debugging/coreclr/debugging-crossgen2.md

you run your test with set RunCrossgen2=1 it will give you a command like dotnet crossgen2.dll *.rsp, now change in the rsp file target from x64 to x86 and run with --jitpath path to clrjit_windows_x86_x64.dll and it should produce a native image compiled on x64 host for x86 target. I should have time this week to check that it reproduces the issue that you described if you don't do this first.

[SingleAccretion]

@sandreenko Hmm, I suppose I was not really clear with the wording. What I meant is how to make sure the test is run the right way in CI, locally it can be reproduced easily with the usual AltJit setup.

That said, I have now checked that it does indeed reproduce with CG2 too.

[sandreenko]

LGTM

[SingleAccretion]

Will rerun the CI one more time...

[SingleAccretion]

Will rerun the CI one more time...

Well, that's not so many failures, so, progress :).

1) Build windows x64 Release SingleFile - mysterious build failures:

error NU1105: (NETCORE_ENGINEERING_TELEMETRY=Restore) Unable to find project information for 'D:\workspace\_work\1\s\src\libraries\System.Runtime.CompilerServices.Unsafe\src\System.Runtime.CompilerServices.Unsafe.ilproj'. If you are using Visual Studio, this may be because the project is unloaded or not part of the current solution so run a restore from the command-line. Otherwise, the project file may be invalid or missing targets required for restore.

2) CoreCLR Pri0 Runtime Tests Run OSX arm64 checked:

• log1, log2 - tests succeeded, but the work item failed. Unclear why.
• log3 - truncated log, likewise unclear what's up.

3) Installer Build and Test coreclr windows_x86 Debug - download failure, chalking up to #55449.

D:\workspace\_work\1\s\.dotnet\sdk\6.0.100-preview.4.21255.9\NuGet.targets(131,5): error : Failed to retrieve information about 'System.Spatial' from remote source 'https://pkgs.dev.azure.com/dnceng/9ee6d478-d288-47f7-aacc-f6e6d082ae6d/_packaging/45bacae2-5efb-47c8-91e5-8ec20c22b4f8/nuget/v3/flat2/system.spatial/index.json'. [D:\workspace\_work\1\s\src\installer\tests\Assets\TestProjects\ResourceLookup\ResourceLookup.csproj]

[sandreenko]

I saw these failures in other PRs, the infra is in a bad state, azure is trying to patch it up, but for now we have to merge on red.