Introduction to memory order consume

Introduction to memory order
consume
2015
issue.hsu@gmail.com

Outline
• Quick recap of acquire and release semantics
• The purpose of consume semantics
• Today’s compiler support
2

Quick Recap of Acquire and Release
Semantics
3

Memory Order
• In the C++11 standard atomic library, most functions accept a
memory_order argument
• Both consume and acquire serve the same purpose
– To help pass non-atomic information safely between threads
• Like acquire operations, a consume operation must be combined with
a release operation in another thread
4
enum memory_order {
memory_order_relaxed,
memory_order_consume,
memory_order_acquire,
memory_order_release,
memory_order_acq_rel,
memory_order_seq_cst
};

Example of Acquire and Release
• Declare two shared variables
• The main thread sits in a loop, repeatedly attempting the following
sequence of read operations
• Another asynchronous task running in another thread try to do a write
operation
5
atomic<int> Guard(0);
int Payload = 0;
for(…)
{
….
g = Guard.load(memory_order_acquire);
if (g != 0)
p = Payload;
…
}
Payload = 42;
Guard.store(1, memory_order_release);

Example of Acquire and Release
• Once the asynchronous task writes to Guard, the main thread reads it
– It means that the write-release synchronized-with the read-acquire
– We are guaranteed that p will equal 42, no matter what platform we run this
example on
• We’ve used acquire and release semantics to pass a simple non-
atomic integer Payload between threads
6

The Cost of Acquire Semantics
7
if (g != 0)
p = Payload;
strong memory model weakly-ordered CPU

The Purpose of Consume Semantics
8

Data Dependency
• The PowerPC and ARM are weakly-ordered CPUs, but in fact, there
are some cases where they do enforce memory ordering at the
machine instruction level without the need for explicit memory barrier
instructions
– These processors always preserve memory ordering between data-dependent
instructions
• When multiple instructions are data-dependent on each other, we call
it a data dependency chain
– In the following PowerPC listing, there are two independent data dependency
chains
9

Data Dependency
• Consume semantics are designed to exploit the data dependency
ordering
• At the source code level, a dependency chain is a sequence of
expressions whose evaluations all carry-a-dependency to each
another
– Carries-a-dependency is defined in §1.10.9 of the C++11 standard
– It mainly says that one evaluation carries-a-dependency to another if the value of
the first is used as an operand of the second
10

Example of Consume and Release
• Declare two shared variables
• The main thread sits in a loop, repeatedly attempting the following
sequence of read operations
• Another asynchronous task running in another thread try to do a write
operation
11
atomic<int> Guard(0);
int Payload = 0;
for(…)
{
…
if (g != 0)
p = Payload;
…
}
Payload = 42;
Guard.store(1, memory_order_release);
atomic<int*> Guard(nullptr);
int Payload = 0;
Payload = 42;
Guard.store(&Payload, memory_order_release);
for(…)
{
…
g = Guard.load(memory_order_consume);
if (g != nullptr)
p = *g;
…
}

• This time, we don’t have a synchronizes-with relationship anywhere.
What we have this time is called a dependency-ordered-before
relationship
• In any dependency-ordered-before relationship, there’s a
dependency chain starting at the consume operation, and all memory
operations performed before the write-release are guaranteed to be
visible to that chain.
12

13

Current Compiler Status
• Those assembly code listings just showed you for PowerPC and
ARMv7 were fabricated
– Sorry, but GCC 4.8.3 and Clang 4.6 don’t actually generate that machine code for
consume operations
• Current versions of GCC and Clang/LLVM use the heavy strategy, all
the time
– As a result, if you compile memory_order_consume for PowerPC or ARMv7 using
today’s compilers, you’ll end up with unnecessary memory barrier instructions
15

Efficient Compiler Strategy in GCC
• GCC 4.9.2 actually has an efficient compiler strategy in its
implementation of memory_order_consume, as described in this
GCC bug report
– Only available in GCC 4.9.2 AARCH64 target
16

• In this example, we are admittedly abusing C++11’s definition of carry-a-dependency
by using f in an expression that cancels it out (f - f). Nonetheless, we are still
technically playing by the standard’s current rules, and thus, its ordering guarantees
should still apply
Example That Illustrates the Compiler Bug
17
int read()
{
int f = Guard.load(std::memory_order_consume); // load-consume
if (f != 0)
return Payload[f - f]; // plain load from Payload[f - f]
return 0;
}
int write()
{
Payload[0] = 42; // plain store to Payload[0]
Guard.store(1, std::memory_order_release); // store-release
}
#include <atomic>
std::atomic<int> Guard(0);
int Payload[1] = { 0xbadf00d };
$ aarch64-linux-g++ -std=c++11 -O2 -S consumetest.cpp

A Patch for This Bug
• Andrew Macleod posted a patch for this issue in the bug report. His
patch adds the following lines near the end of the get_memmodel
function in gcc/builtins.c
• After patching
– $ aarch64-linux-g++ -std=c++11 -O2 -S consumetest.cpp
18
/* Workaround for Bugzilla 59448. GCC doesn't track consume properly, so
be conservative and promote consume to acquire. */
if (val == MEMMODEL_CONSUME)
val = MEMMODEL_ACQUIRE;

This Bug Doesn’t Happen on PowerPC
• Interestingly, if you compile the same example for PowerPC, there is
no bug. This is using the same GCC version 4.9.2 without Andrew’s
patch applied
– $ powerpc-linux-g++ -std=c++11 -O2 -S consumetest.cpp
19
if (model == MEMMODEL_RELAXED
|| model == MEMMODEL_CONSUME
|| model == MEMMODEL_RELEASE)
return "ldr<atomic_sfx>t%<w>0, %1";
else
return "ldar<atomic_sfx>t%<w>0, %1";
switch (model)
{
case MEMMODEL_RELAXED:
break;
case MEMMODEL_CONSUME:
case MEMMODEL_ACQUIRE:
case MEMMODEL_SEQ_CST:
emit_insn (gen_loadsync_<mode> (operands[0]));
break;
gcc-4.9.2/gcc/config/rs6000/sync.mdgcc-4.9.2/gcc/config/aarch64/atomics.md

The Uncertain Future of memory order
consume
• The C++ standard committee is wondering what to do with
memory_order_consume in future revisions of C++
• The author’s opinion is that the definition of carries-a-dependency
should be narrowed to require that different return values from a load-
consume result in different behavior for any dependent statements
that are executed
– Using f - f as a dependency is nonsense, and narrowing the definition would free
the compiler from having to support such nonsense “dependencies” if it chooses
to implement the efficient strategy
– This idea was first proposed by Torvald Riegel in the Linux Kernel Mailing List and
is captured among various alternatives described in Paul McKenney’s proposal
N4036
20
int my_array[MY_ARRAY_SIZE];
i = atomic_load_explicit(gi, memory_order_consume);
r1 = my_array[i];

References
• The Purpose of memory_order_consume in C++11
– http://preshing.com/20140709/the-purpose-of-memory_order_consume-in-cpp11/
• Fixing GCC's Implementation of memory_order_consume
– http://preshing.com/20141124/fixing-gccs-implementation-of-
memory_order_consume/
• http://en.cppreference.com/w/cpp/atomic/memory_order
• Bug 59448 - Code generation doesn't respect C11 address-dependency
– https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59448
• N4036: Towards Implementation and Use of memory order consume
– https://isocpp.org/files/papers/n4036.pdf
• Demo program
– https://github.com/preshing/ConsumeDemo
22

Introduction to memory order consume

More Related Content

What's hot (20)

Similar to Introduction to memory order consume (20)

More from Yi-Hsiu Hsu (7)

Recently uploaded (20)

Introduction to memory order consume