Fixup CodeFragmentHeap block allocation reporting

[hoyosjs]

#122274 fixed up some cases where block allocations produced code, but try to do a read of uncommitted memory. This change just makes it such that block records have the right size for range identification.

After this the settings provide the following expectations:

   DOTNET_PerfMapEnabled
     0 (default)  Disabled
     1            Perfmap text file + jitdump
     2            Jitdump only
     3            Perfmap text file only

   DOTNET_PerfMapShowOptimizationTiers
     0                 No tier info in symbol names
     1 (default)  Show optimization tiers in symbol names

   DOTNET_PerfMapStubGranularity
     Controls how stubs are reported. Bit 0 controls naming,
     bit 1 controls individual allocation site reporting.

     0 (default)  Grouped names, block-level only
                  Name format: "stub Type<Owner>"
                  Block stubs: perfmap yes, jitdump yes (no disassembly)
                  IndividualWithinBlock: not reported

     1            Unique names, block-level only
                  Name format: "stub<N> Type<Owner>"
                  Block stubs: perfmap yes, jitdump yes (no disassembly)
                  IndividualWithinBlock: not reported

     2            Grouped names, individual alloc sites
                  Name format: "stub Type<Owner>"
                  Block stubs: not reported
                  IndividualWithinBlock: perfmap yes, jitdump yes

     3            Unique names, individual alloc sites
                  Name format: "stub<N> Type<Owner>"
                  Block stubs: not reported
                  IndividualWithinBlock: perfmap yes, jitdump yes

     Individual stubs (e.g. write barriers) are always reported
     with full disassembly regardless of granularity setting.

[Copilot]

Pull request overview

This PR fixes malformed Linux jitdump JIT_CODE_LOAD records emitted for block-level stub allocations when DOTNET_PerfMapStubGranularity is 0 or 1, by avoiding emitting jitdump records that would require unreadable/uninitialized code bytes. This preserves perfmap text output while preventing jitdump parser corruption.

Changes:

• Remove the reportCodeBlock flag from PAL_PerfJitDump_LogMethod and always write code_size bytes when a jitdump record is emitted.
• Skip emitting jitdump records for block-level stub allocations (still emit perfmap text entries).
• Adjust CodeFragmentHeap stub-block reporting timing for newly allocated blocks.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
src/coreclr/vm/perfmap.cpp	Stops emitting jitdump entries for block stub allocations; updates jitdump logging call sites for methods/stubs to new signature.
src/coreclr/vm/codeman.cpp	Moves ReportStubBlock call to after size bookkeeping for new fragment blocks.
src/coreclr/pal/src/misc/perfjitdump.cpp	Simplifies jitdump record emission: removes reportCodeBlock and always appends code_size bytes.
src/coreclr/pal/inc/pal.h	Updates the PAL API signature for PAL_PerfJitDump_LogMethod.

[dotnet-policy-service]

Tagging subscribers to this area: @steveisok, @tommcdon, @dotnet/dotnet-diag
See info in area-owners.md if you want to be subscribed.

[noahfalk]

fyi @tommcdon who worked on the previous iteration of fixing this

[hoyosjs]

I got a bit more context about the intention of the settings from David. There's a few more bugs lurking here. Going to try to get one more round of fixes.

[davidwrighton]

Yeah, the intention is to actually report these regions, so they can be visible in a Perf trace. I'm not sure not emitting a jitdump record is the right thing here.

[Copilot]

Pull request overview

Adjusts CoreCLR perfmap/jitdump stub reporting so stub-block entries better reflect the committed/usable ranges and unique stub names are generated safely under concurrency, aligning with the stub granularity expectations described in the PR.

Changes:

• Make stub unique-name generation use an atomic increment instead of a plain ++.
• Move ReportStubBlock to after block splitting so the reported size matches the portion retained/used.
• Update the s_StubsMapped member intent/comment to reflect its role as a naming counter.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
src/coreclr/vm/perfmap.h	Changes stub-name counter declaration/comment.
src/coreclr/vm/perfmap.cpp	Switches unique stub naming to InterlockedIncrement.
src/coreclr/vm/codeman.cpp	Reports stub-block allocation after adjusting dwSize to avoid over-reporting the range.

[tommcdon]

Could we end up double reporting blocks if we don't call AllocCodeFragmentBlock?

[hoyosjs]

you can end up grabbing it from the best fit path - that but since that got aligned down and returned during the if block right above this, that would cound as a separate block and should be reported separately