Fix CompareInfo.IsPrefix/IsSuffix with CompareOptions.IgnoreSymbols not ignoring leading/trailing symbols

[HarnageaGabriel]

Fixes #118521

Root cause

ComplexStartsWith/ComplexEndsWith in pal_collation.c use ICU string search to locate the requested affix, then call CanIgnoreAllCollationElements to check whether any skipped text before/after the match is entirely ignorable under the active CompareOptions.

CanIgnoreAllCollationElements walked the raw collation elements returned by ucol_next (via ucol_openElements) and required each one to equal UCOL_IGNORABLE (0). Raw collation-element iteration does not apply the collator's UCOL_ALTERNATE_HANDLING setting — so when CompareOptions.IgnoreSymbols sets UCOL_ALTERNATE_HANDLING = UCOL_SHIFTED on the collator, a symbol character that should now be ignorable (e.g. ') still produces a non-zero raw collation element. This made CanIgnoreAllCollationElements incorrectly report the skipped text as "not ignorable," so:

CultureInfo.InvariantCulture.CompareInfo.IsPrefix("''Tests", "Tests", CompareOptions.IgnoreSymbols)

returned false instead of true on ICU (non-Windows NLS, non-Apple-hybrid) platforms — those other implementations already used a different code path that isn't affected by this quirk, which is why the existing test file already had a platform-conditional branch documenting this exact case as a known ICU-backend bug pointing at this issue.

I verified this against a live ICU instance: raw ucol_next elements for symbol characters made ignorable via IgnoreSymbols remain non-zero, while ucol_strcoll against those same characters correctly honors the collator's strength/alternate-handling and reports equality with an empty string.

Fix

CanIgnoreAllCollationElements now compares the skipped substring against an empty string using ucol_strcoll (the same high-level comparison entry point already used elsewhere in this file, e.g. GlobalizationNative_CompareString), instead of manually iterating and masking raw collation elements. This honors the collator's configured strength, alternate handling (IgnoreSymbols), and locale tailoring uniformly, rather than re-implementing that logic against the low-level iterator API. A non-null empty UChar buffer is passed (rather than NULL) to avoid a documented old-ICU null-input issue that this file already works around elsewhere (ICU-9396).

This helper is shared by both ComplexStartsWith and ComplexEndsWith, so the fix resolves the symmetric IsSuffix case as well (e.g. "Tests''".EndsWith("Tests", CompareOptions.IgnoreSymbols)), for which I added a regression test.

Changes

• src/native/libs/System.Globalization.Native/pal_collation.c: replaced the raw collation-element loop in CanIgnoreAllCollationElements with an ucol_strcoll-against-empty-string check.
• src/libraries/System.Runtime/tests/System.Globalization.Tests/CompareInfo/CompareInfoTests.IsPrefix.cs: removed the platform-conditional "known ICU bug" branch for ''Tests/Tests and set the expected result to true (matched length 7) unconditionally.
• src/libraries/System.Runtime/tests/System.Globalization.Tests/CompareInfo/CompareInfoTests.IsSuffix.cs: added the symmetric Tests''/Tests regression case.

Validation

I don't have a full native ICU build/toolchain available in this environment, so I wasn't able to run the managed test suite against a compiled libSystem.Globalization.Native. I did directly probe ICU's C collation API (usearch_first/usearch_last, raw ucol_next vs ucol_strcoll) to confirm the root cause and the fix's behavior character-by-character for the affected cases. The change is narrowly scoped to the single shared helper function and mirrors an existing, already-used API call pattern in the same file.

[azure-pipelines]

Azure Pipelines:
Successfully started running 3 pipeline(s).
13 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/area-system-globalization
See info in area-owners.md if you want to be subscribed.

[HarnageaGabriel]

@dotnet-policy-service agree

[HarnageaGabriel]

@dotnet-policy-service agree

[tarekgh]

@EgorBot -windows_x64 -linux_x64 --filter "StringSearch.*"

[tarekgh]

@EgorBot -windows_x64 -linux_x64 --filter "System.Globalization.Tests.StringSearch.*"

[tarekgh]

This change is a nice improvement because it also removes an existing ICU-vs-NLS divergence, not just the IgnoreSymbols case in the linked issue.

Since CanIgnoreAllCollationElements now honors the collator's configured strength and alternate handling, the leftover-text check also becomes correct for IgnoreNonSpace when a leading/trailing nonspacing mark is skipped. NLS already returned true for those; ICU was returning false.

Comparing the current runtime (ICU, pre-fix) against NLS on the same inputs with InvariantCulture:

Case	ICU (pre-fix)	NLS (reference)	ICU (after fix)
IsPrefix("''Tests", "Tests", IgnoreSymbols)	False	True	True
IsSuffix("Tests''", "Tests", IgnoreSymbols)	False	True	True
IsPrefix("\u0308abc", "abc", IgnoreNonSpace)	False	True	True
IsSuffix("abc\u0308", "abc", IgnoreNonSpace)	True	True	True

So the fix moves ICU toward the NLS behavior in all of these cases. The IgnoreNonSpace leading-mark case isn't currently covered by the IsPrefix/IsSuffix test data, so nothing regresses, but it might be worth adding a regression row for it to lock in the behavior.

I also checked that the change stays precise and doesn't over-ignore: a combining mark is still not ignorable under IgnoreSymbols, and a real letter is never ignorable under IgnoreNonSpace.

[tarekgh]

/ba-g the failures are not related

[tarekgh]

The benchmarks not showing any regression either. We should be good to go.

[tarekgh]

Thanks @HarnageaGabriel for providing the fix.