This associates debug information with the assertion sites, allowing clearer stack-traces for assertion failures and better accounting of the performance overhead of assertions.
PiperOrigin-RevId: 910863016
Change-Id: Ic615a9eeb8ad4b3f8c035074df4ff8347190e5c7
This will let us deprecate the declarations without triggering warnings in Abseil itself.
PiperOrigin-RevId: 894202105
Change-Id: I57bb2a1647be1fedf9b724a07042fd0f564ce074
absl::variant, and related types
The corresponding headers are removed from cc files, but kept in
headers to prevent breakages from transitive dependencies.
PiperOrigin-RevId: 872421685
Change-Id: I867d4c3f7c9e422289c63816d44719b0530fb0a6
It is called with `(nullptr, 0)` by `absl::Cord` constructor from
`absl::string_view(nullptr, 0)`.
The implementation works for `(nullptr, 0)`.
PiperOrigin-RevId: 802213857
Change-Id: Ibb9b9dae620262ebee8aa3347c4177e7a3da52c0
When a container or a string is mixed, size is only supplementary data.
We are still mixing size to avoid hash expansion to be a suffix of one another.
Mixing is done by addition `size + Seed()`. `+ Seed()` is needed to make an empty string to change the hash state. We assume that `Seed()` is already loaded to some register. Addition would modify low bits that will be spread with data mixing later.
We considered the following optimization. Mix the size at the beginning in order to improve the dependency graph. Mixing would happen in parallel with reading strings data. It's not feasible because absl::Hash API requires that hash expansions can't be suffixes of each other.
```
name old CYCLES/op new CYCLES/op delta
BM_latency_AbslHash_String3 27.2 ± 0% 21.5 ± 0% -21.08% (p=0.000 n=52+52)
BM_latency_AbslHash_String5 28.1 ± 8% 22.3 ±11% -20.70% (p=0.000 n=57+57)
BM_latency_AbslHash_String9 27.7 ± 9% 22.2 ±17% -19.77% (p=0.000 n=56+57)
BM_latency_AbslHash_String17 25.9 ± 5% 21.2 ±23% -18.02% (p=0.000 n=57+54)
BM_latency_AbslHash_String33 28.0 ± 5% 22.5 ± 4% -19.69% (p=0.000 n=54+54)
BM_latency_AbslHash_String65 37.9 ± 8% 32.2 ±10% -15.19% (p=0.000 n=52+53)
BM_latency_AbslHash_String257 57.1 ± 8% 52.0 ±10% -8.88% (p=0.000 n=52+54)
```
PiperOrigin-RevId: 750151406
Change-Id: I2245bad4906960d9236bea671738a218a85eb1af
An `operator->` that doesn't return a record type doesn't really serve a useful
purpose, and it is not required by the iterator concepts. I haven't been able
to find any usage in non-test code, so I'm removing it.
Even though the `pointer` type alias is now not being used anywhere, I'm keeping
it because otherwise some standard libraries don't accept `CharIterator` when
passed to `std::string::apend()`, for example.
PiperOrigin-RevId: 659832167
Change-Id: I5de183a01009bfd71c8477f2023973075be8d860
This is portable because cord already has `operator<` etc., which will be unaffected. This just allows C++ >= 20 users to explicitly call `operator<=>`.
PiperOrigin-RevId: 646951415
Change-Id: I1432e224bd5dc09b99d56a1d27e95078463adf45
These were already on the declarations, but it seems like the definitions need them as well.
PiperOrigin-RevId: 621868985
Change-Id: I2595aae506924f8b1dad3b07efe1660ae0e4488e
`MaybeRemoveEmptyCrcNode()`. Otherwise appending a `CordBuffer` an empty Cord
with a CRC node crashes (`RemoveCrcNode()` which increases the refcount of a
nullptr child).
Cosmetics: in `Cord::InlineRep::AppendArray()`, return early for empty `src`
before removing the empty CRC node.
PiperOrigin-RevId: 619107278
Change-Id: I4f1bc6b75c662f4678c61e3ef310e8597d62e2e1
We have no intention to use it instead of the CordRepBtree implementation, so cleanup up and remove all code and references.
PiperOrigin-RevId: 563803813
Change-Id: I95a67318d0f722f3eb7ecdcc7b6c87e28f2e26dd
There's no change to the existing kTotal or kFairShare accounting modes.
PiperOrigin-RevId: 544378591
Change-Id: I7b4ae55cd93d631194e59a9cd0ff07f47611219e
The current implementation correctly copies up to 15 bytes of data, but the nullify code clears up to 16 bytes, which was likely motivated by the assumption that the length indicator is always the last byte.
This changes limits the nullify to 15 bytes as well removing this layout specific assumption from cord.h, making future platform specific internal layout changes easier to land.
PiperOrigin-RevId: 477997741
Change-Id: Idcdfeca2a005139f97eafcc77111542d90b817af
This removes layout specific details from InlineData from cord.h, making future platform specific internal layout changes easier to land.
PiperOrigin-RevId: 477869206
Change-Id: I1d417af47d7f04e34a98ba7b93ae591ece8f9151
This makes future platform specific internal layout changes easier to land.
PiperOrigin-RevId: 477843948
Change-Id: I4ca4fdea5e965261c029d08319aba0290721c227
The Cord::GetCustomAppendBuffer() method provides the same functionality as Cord::GetAppendBuffer(), except that callers can specify a custom block size/limit if the method returns a newly allocated buffer.
In other words: Cord::GetAppendBuffer() defaults to CordBuffer::CreateWithDefaultLimit(), Cord::GetCustomAppendBuffer() defaults to CordBuffer::CreateWithCustomLimit().
PiperOrigin-RevId: 461231989
Change-Id: I5c03f31139d9b068feee1bea76d59e1c5e30ef07
absl::CordBuffer holds data for eventual inclusion within an existing
absl::Cord. CordBuffers are useful for building large Cords that may
require custom allocation of its associated memory, a pattern that is
common in zero-copy APIs.
PiperOrigin-RevId: 453212229
Change-Id: I6a8adc3a8d206691cb1b0001a9161e5080dd1c5f
--
d3b99682554d339c42556680f4d65f83226005e2 by Martijn Vels <mvels@google.com>:
Inline CycleClock code and remove branch for x86 CycleClockSource function
This CL removes the relaxed load for x86 as there is no acquire price to pay on x86. It inlines the UnscaledCycleClock::Now() which is a single RTDSC op for x86, and likewise inlines CycleClock::Now() for x86. The inlining should mostly have secondary benefits such as reducing spills on outlined calls.
LTO may eventually hoist these functions inline for the hotspots, but it doesn't hurt to default inline this for all builds and let the compiler decide on the first pass.
The perlab benchmark is noisy for the plain BM_Now, but the other benchmarks and the run on my local machine are clear.
------------- Local Benchy Benchmark
name old cpu/op new cpu/op delta
BM_Now 3.41ns ± 1% 2.30ns ± 2% -32.52% (p=0.000 n=50+50)
BM_NowWithRegisterPresure 4.96ns ± 2% 4.19ns ± 2% -15.57% (p=0.000 n=56+55)
BM_NowWithCallback 3.30ns ± 2% 1.91ns ± 2% -42.00% (p=0.000 n=47+60)
------------- Perflab Benchy Benchmark
name old cpu/op new cpu/op delta
BM_Now 8.20ns ±13% 4.32ns ±83% ~ (p=0.413 n=4+5)
BM_NowWithRegisterPresure 7.91ns ± 1% 3.68ns ± 2% -53.45% (p=0.029 n=4+4)
BM_NowWithCallback 2.66ns ±13% 1.58ns ± 0% -40.51% (p=0.008 n=5+5)
PiperOrigin-RevId: 434474766
Change-Id: I991d987ae9233e50f09606c874055cf4c5a56300
--
b38330686a0af176a2679163e4d2fa1b90e2f667 by Laramie Leavitt <lar@google.com>:
Style, comment, and test updates
* Remove a redundant assert in uniform_real_distribution.
* Update comment in internal/generate_real.h
* Style updates to uniform_real_distribution_test
mainly replacing TypeParam with real_type, using aliases for some limits, etc.
* Add a few more minor tests.
PiperOrigin-RevId: 433902174
Change-Id: Id75be8e24be2fb8f6aea05feec13e3ef320a7254
--
ab2da6047ff7f5dae3add3779fcddf73b03feabf by Abseil Team <absl-team@google.com>:
Remove declaration of method whose definition was previously removed.
PiperOrigin-RevId: 433507828
Change-Id: I0130b689813125250f7de2664e767e181f676c89
--
df0c87f4ec2c010691931c1bef9d26470a6e63a2 by Derek Mauro <dmauro@google.com>:
Internal change
PiperOrigin-RevId: 433289136
Change-Id: Iba157dc83ed99dafd17a2223d2504e49f8afbb9e
--
7445fa312f2995772900eda82467325b3401a17d by Martijn Vels <mvels@google.com>:
Optimize CordReader logic now that CONCAT is removed
This CL cleans up various helper functions and logic remaining from previous complex CONCAT logic that is no longer needed, simplifying the CordReader logic.
PiperOrigin-RevId: 433208748
Change-Id: I5f7b1883573c44e7c6f8af12c3cddbd197cb134d
GitOrigin-RevId: d3b99682554d339c42556680f4d65f83226005e2
--
ed829ac612f090375427c3488827c6e74deb2e3f by Derek Mauro <dmauro@google.com>:
Update latest GCC/Clang Linux tests to Bazel 5.0.0 and CMake 3.22.2
PiperOrigin-RevId: 429369775
--
76952303c4d942288c4e7657ffb5893cec54a132 by Martijn Vels <mvels@google.com>:
Optimize Cord::ChunkIterator now that CordRepConcat is removed
PiperOrigin-RevId: 429321455
--
dcd0d287793649aba9b98268c5783e449a34749f by Martijn Vels <mvels@google.com>:
Add IsDataEdge() and DataEdgeValue() helper functions.
This moves repetitive logic accessing data edges into its own header, and more strongly defines the notion of what a data edge is, enforcing the internal invariants. This will also be incorporated in optimized Cord iteration logic once CordRepConcat is totally removed from the Cord code.
PiperOrigin-RevId: 429307248
--
6a0903962155988085bf8656743fda9c4cdcba6c by Abseil Team <absl-team@google.com>:
Make it clear that the probability function given for the zipf distribution is unnormalized, i.e., sum(p(x) for x = 0..k) != 100%.
Quoting Section 7 of the paper cited in the comments, where this formula comes from (emphasis mine): "We will consider the two parameter generalization as defined in Dagpunar [1988] with the *unnormalized* probability function ..."
PiperOrigin-RevId: 429068258
--
3899ff6d444ba755148bc521a6ee031d9e9d4485 by Abseil Team <absl-team@google.com>:
Internal Changes
PiperOrigin-RevId: 428644856
--
319de702d2b537cbb76c4c71277ae89b349b162e by Benjamin Barenblat <bbaren@google.com>:
Support symbolization on PA-RISC
Null out supervisor bits in PA-RISC addresses before symbolizing, and
handle function descriptor tables correctly.
Change symbolize_test.cc to use 32-bit aligned addresses, allowing that
test to pass on PA-RISC.
PiperOrigin-RevId: 428590564
GitOrigin-RevId: ed829ac612f090375427c3488827c6e74deb2e3f
Change-Id: Ie01ff3b9365fd45e5a55f858038552679f3180d3
