Memory order consume usage in C11

consume is cheaper than acquire. All CPUs (except DEC Alpha AXP’s famously weak memory model1) do it for free, unlike acquire. (Except on x86 and SPARC-TSO, where the hardware has acq/rel memory ordering without extra barriers or special instructions.)

On ARM/AArch64/PowerPC/MIPS/etc weakly-ordered ISAs, consume and relaxed are the only orderings that don’t require any extra barriers, just ordinary cheap load instructions. i.e. all asm load instructions are (at least) consume loads, except on Alpha. acquire requires LoadStore and LoadLoad ordering, which is a cheaper barrier instruction than a full-barrier for seq_cst, but still more expensive than nothing.

mo_consume is like acquire only for loads with a data dependency on the consume load. e.g. float *array = atomic_ld(&shared, mo_consume);, then access to any array[i] is safe if the producer stored the buffer and then used a mo_release store to write the pointer to the shared variable. But independent loads/stores don’t have to wait for the consume load to complete, and can happen before it even if they appear later in program order. So consume only orders the bare minimum, not affecting other loads or stores.


(It’s basically free to implement support for consume semantics in hardware for most CPU designs, because OoO exec can’t break true dependencies, and a load has a data dependency on the pointer, so loading a pointer and then dereferencing it inherently orders those 2 loads just by the nature of causality. Unless CPUs do value-prediction or something crazy.
Value prediction is like branch prediction, but guess what value is going to be loaded instead of which way a branch is going to go.

Alpha had to do some crazy stuff to make CPUs that could actually load data from before the pointer value was truly loaded, when the stores were done in order with sufficient barriers.

Unlike for stores, where the store buffer can introduce reordering between store execution and commit to L1d cache, loads become “visible” by taking data from L1d cache when they execute, not when the retire + eventually commit. So ordering 2 loads wrt. each other really does just mean executing those 2 loads in order. With a data dependency of one on the other, causality requires that on CPUs without value prediction, and on most architectures the ISA rules do specifically require that. So you don’t have to use a barrier between loading + using a pointer in asm, e.g. for traversing a linked list.)

See also Dependent loads reordering in CPU


But current compilers just give up and strengthen consume to acquire

… instead of trying to map C dependencies to asm data dependencies (without accidentally breaking having only a control dependency that branch prediction + speculative execution could bypass). Apparently it’s a hard problem for compilers to keep track of it and make it safe.

It’s non-trivial to map C to asm, because if the dependency is only in the form of a conditional branch, the asm rules don’t apply. So it’s hard to define C rules for mo_consume propagating dependencies only in ways that line up with what does “carry a dependency” in terms of asm ISA rules.

So yes, you’re correct that consume can be safely replaced with acquire, but you’re totally missing the point.


ISAs with weak memory-ordering rules do have rules about which instructions carry a dependency. So even an instruction like ARM eor r0,r0 which unconditionally zeroes r0 is architecturally required to still carry a data dependency on the old value, unlike x86 where the xor eax,eax idiom is specially recognized as dependency-breaking2.

See also http://preshing.com/20140709/the-purpose-of-memory_order_consume-in-cpp11/

I also mentioned mo_consume in an answer on Atomic operations, std::atomic<> and ordering of writes.


Footnote 1: The few Alpha models that actually could in theory “violate causality” didn’t do value-prediction, there was a different mechanism with their banked cache. I think I’ve seen a more detailed explanation of how it was possible, but Linus’s comments about how rare it actually was are interesting.

Linus Torvalds (Linux lead developer), in a RealWorldTech forum thread

I wonder, did you see non-causality on Alpha by yourself or just in the manual?

I never saw it myself, and I don’t think any of the models I ever had
access to actually did it. Which actually made the (slow) RMB
instruction extra annoying, because it was just pure downside.

Even on CPU’s that actually could re-order the loads, it was
apparently basically impossible to hit in practice. Which is actually
pretty nasty. It result in “oops, I forgot a barrier, but everything
worked fine for a decade, with three odd reports of ‘that can’t
happen’ bugs from the field” kinds of things. Figuring out what’s
going on is just painful as hell.

Which models actually had it? And how exactly they got here?

I think it was the 21264, and I have this dim memory of it being due
to a partitioned cache: even if the originating CPU did two writes in
order (with a wmb in between), the reading CPU might end up having the
first write delayed (because the cache partition that it went into was
busy with other updates), and would read the second write first. If
that second write was the address to the first one, it could then
follow that pointer, and without a read barrier to synchronize the
cache partitions, it could see the old stale value.

But note the “dim memory”. I may have confused it with something else.
I haven’t actually used an alpha in closer to two decades by now. You
can get very similar effects from value prediction, but I don’t think
any alpha microarchitecture ever did that.

Anyway, there definitely were versions of the alpha that could do
this, and it wasn’t just purely theoretical.

(RMB = Read Memory Barrier asm instruction, and/or the name of Linux kernel function rmb() that wraps whatever inline asm is necessary to make that happen. e.g. on x86, just a barrier to compile-time reordering, asm("":::"memory"). I think modern Linux manages to avoid an acquire barrier when only a data dependency is needed, unlike C11/C++11, but I forget. Linux is only portable to a few compilers, and those compilers do take care to support what Linux depends on, so they have an easier time than the ISO C11 standard in cooking up something that works in practice on real ISAs.)

See also https://lkml.org/lkml/2012/2/1/521 re: Linux’s smp_read_barrier_depends() which is necessary in Linux only because of Alpha. (But a reply from Hans Boehm points out that “compilers can, and sometimes do, remove dependencies“, which is why C11 memory_order_consume support needs to be so elaborate to avoid risk of breakage. Thus smp_read_barrier_depends is potentially brittle.)


Footnote 2: x86 orders all loads whether they carry a data dependency on the pointer or not, so it doesn’t need to preserve “false” dependencies, and with a variable-length instruction set it actually saves code size to xor eax,eax (2 bytes) instead mov eax,0 (5 bytes).

So xor reg,reg became the standard idiom since early 8086 days, and now it’s recognized and actually handled like mov, with no dependency on the old value or RAX. (And in fact more efficiently than mov reg,0 beyond just code-size: What is the best way to set a register to zero in x86 assembly: xor, mov or and?)

But this is impossible for ARM or most other weakly ordered ISAs, like I said they’re literally not allowed to do this.

ldr r3, [something]       ; load r3 = mem
eor r0, r3,r3             ; r0 = r3^r3 = 0
ldr r4, [r1, r0]          ; load r4 = mem[r1+r0].  Ordered after the other load

is required to inject a dependency on r0 and order the load of r4 after the load of r3, even though the load address r1+r0 is always just r1 because r3^r3 = 0. But only that load, not all other later loads; it’s not an acquire barrier or an acquire load.

Leave a Comment