diff options
author | Abseil Team <absl-team@google.com> | 2019-04-24T08·44-0700 |
---|---|---|
committer | Matt Calabrese <calabrese@x.team> | 2019-04-24T18·35-0400 |
commit | 33841c5c963aa9c3f096ef8e6c1e71624b941940 (patch) | |
tree | 8ddb90d308e40dfcb793d3db77759741b4e809cf /absl/synchronization/blocking_counter.h | |
parent | ca3f87560a0eef716195cadf66dc6b938a579ec6 (diff) |
Export of internal Abseil changes.
-- bb92c768e2271ddbebc1b1eb7e16a7b7c86a6e1c by Abseil Team <absl-team@google.com>: Automated g4 rollback of changelist 244998488. *** Reason for rollback *** I'm seeing test failures, rolling this back. *** Original change description *** BEGIN_PUBLIC The default constructor for optional<T> is filling dummy_ with zeros (see https://godbolt.org/z/IVea7X for a reduced example), which has a performance impact for large Ts. This comes from the gcc6 bugfix that made dummy as big as T. Because constexpr constructors are required to initialize all members of a struct, we cannot prevent this in a standard-compliant way as soon as dummy has any members (note that clang will happily accept adding a `constexpr dummy_type() {}` constructor... *** PiperOrigin-RevId: 245004716 -- 6e3ee35af50ffbee604c22300f3260ebc5f6cf52 by Abseil Team <absl-team@google.com>: The default constructor for optional<T> is filling dummy_ with zeros (see https://godbolt.org/z/IVea7X for a reduced example), which has a performance impact for large Ts. This comes from the gcc6 bugfix that made dummy as big as T. Because constexpr constructors are required to initialize all members of a struct, we cannot prevent this in a standard-compliant way as soon as dummy has any members (note that clang will happily accept adding a `constexpr dummy_type() {}` constructor to dummy_type to prevent zero-initialization, but this is UB AFAICT). This all stems from the fact that we're constructing an object by using placement new on dummy_. The solution I'm using here is to do the placement new on the actual data_. This creates a new issue in when T is volatile, because we can no longer use `&data_` to do the placement new. The solution I'm using here is to make data_ a non-const and non-volatile T, and only provide fully possibly qualified access through `reference()` accessors. I think this correctly prevents UB. PiperOrigin-RevId: 244998488 -- 4f52e64c4cf6aef8df6360007bcc53d8b00db2b4 by Abseil Team <absl-team@google.com>: Increase SYMBOL_BUF_SIZE from 2KB to 3KB. PiperOrigin-RevId: 244954529 GitOrigin-RevId: bb92c768e2271ddbebc1b1eb7e16a7b7c86a6e1c Change-Id: Iaed9a027064a9ecd194c5c146169c683b77f12ef
Diffstat (limited to 'absl/synchronization/blocking_counter.h')
0 files changed, 0 insertions, 0 deletions