[all-commits] [llvm/llvm-project] ba6735: [clang] Consistently cache failed constraint norma...

[Nico Weber via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: ba67350bf808d8bef21cc2d59d7bcd0396adc29d
      https://github.com/llvm/llvm-project/commit/ba67350bf808d8bef21cc2d59d7bcd0396adc29d
  Author: Nico Weber <thakis at chromium.org>
  Date:   2026-09-28 (Mon, 28 Sep 2026)

  Changed paths:
    M clang/lib/Sema/SemaConcept.cpp
    M clang/test/CXX/temp/temp.constr/temp.constr.normal/p1.cpp

  Log Message:
  -----------
  [clang] Consistently cache failed constraint normalization (#227086)

Previously, if substituting the parameter mappings of a normalized
constraint failed, Sema::getNormalizedAssociatedConstraints() returned
nullptr, but it stored the partially substituted normal form in
NormalizationCache. So the first lookup for such a declaration failed,
but every later lookup returned the broken normal form, and subsumption
checking and the ambiguous-constraint diagnostics then continued with
it.

I believe this wasn't intentional:

- Before #161671 (e9972debc98c), normalization was a single step, and a
failure was cached as nullptr.

- #161671 added the parameter mapping substitution step. It inserted the
normal form into the cache before substituting, and returned nullptr if
the substitution then failed, leaving the non-null normal form in the
cache.

- #165352 (2984a8db804e) moved the insertion after the substitution to
not use an invalidated iterator, but kept inserting the normal form if
the substitution failed.

To fix this, cache failed substitution as nullptr, like a failed
normalization.

As a consequences we don't emit the diagnostics that were only emitted
because the second lookup continued:

- A second "'type name' declared as a pointer to a reference" error
(with its notes) for the same broken concept. For

      template<typename T> concept Foo = True<T*>;
      template<typename T> concept Bar = Foo<T&>;
      template<typename T> requires Bar<T> struct S { };
      template<typename T> requires Bar<T> && true struct S<T> { };

the error got reported once for the partial specialization, and also for
the primary template after. Now, we only have the first report.

- A "similar constraint expressions not considered equivalent" note and
its "similar constraint expression here" note, which we computed from
the broken normal form. The actual problem in that test is the broken
constraint, which is still diagnosed.

This also makes it possible to key the normalization cache by constraint
expression without changing which diagnostics are emitted for failures,
which I want to do in a follow-up.



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications