[all-commits] [llvm/llvm-project] e8c09d: [clang] Cache normalized constraints by expression...

[Nico Weber via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: e8c09d19de8212bfb2d5c760c43a21e27729f660
      https://github.com/llvm/llvm-project/commit/e8c09d19de8212bfb2d5c760c43a21e27729f660
  Author: Nico Weber <thakis at chromium.org>
  Date:   2026-09-29 (Tue, 29 Sep 2026)

  Changed paths:
    M clang/docs/InternalsManual.md
    M clang/include/clang/Sema/Sema.h
    M clang/lib/Sema/SemaConcept.cpp

  Log Message:
  -----------
  [clang] Cache normalized constraints by expression instead of by declaration (#226620)

Sema::NormalizationCache was keyed by the constrained declaration. The
members of a class template specialization are distinct declarations for
every specialization, but they all share the uninstantiated constraint
expressions of the member of the primary template they were instantiated
from. So the constraints of e.g. the constrained constructors of
std::optional, std::span, std::pair or of the members of range adaptors
were normalized from scratch for every single specialization of those
classes that a TU uses. Normalization is expensive: it expands the whole
concept tree below the expression and substitutes the parameter mapping
at every level.

When compiling Chromium, 48%
(chrome/browser/glic/...contents_manager.cc)
to 72% (chrome/browser/ui/views/frame/browser_view.cc) of all
NormalizationCache misses were for expressions that were normalized
before
for a different declaration, and normalization is 4.5% of all compile
time
in a sampling profile.

Instead, key the cache by (expression, pack index). Each associated
constraint is cached separately, and we form the conjunction of a
declaration's associated constraints on every call. That's cheap, and
getNormalizedAssociatedConstraints() is only called on SatisfactionCache
misses and for subsumption checks. Failures are cached as nullptr, like
before.

Build time for 120 random Chromium TUs drawn with probability
proportional to their compile time:

    CPU time: 470.0 s => 451.1 s, -4.0%
    instructions: 4948.9e9 => 4702.7e9, -5.0%
    mean max RSS: 779 MB => 710 MB, -8.9%

(To check noise level, running this with the same binary twice: +-0.04%
CPU time and +-0.01% instructions.)

Recovers some more of the time lost to #161671 and contributes to
improving issue #172266.

No behavior change on the non-error case.

(If several declarations share a constraint expression that fails to
normalize (e.g. the members of different specializations of a class
template), we now diagnose the failure once instead of once per
declaration.)



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