Related
When I run this little assembly program on my Ryzen 9 3900X:
_start: xor rax, rax
xor rcx, rcx
loop0: add rax, 1
mov rdx, rax
and rdx, 1
add rcx, rdx
cmp rcx, 1000000000
jne loop0
It completes in 450 ms if all the instructions between loop0 and up to and including the jne, are contained entirely in one cacheline. That is, if:
round((address of loop0)/64) == round((address of end of jne-instruction)/64)
However, if the above condition does not hold, the loop takes 900 ms instead.
I've made a repo with the code https://github.com/avl/strange_performance_repro .
Why is the inner loop much slower in some specific cases?
Edit: Removed a claim with a conclusion from a mistake in testing.
Your issue lies in the variable cost of the jne instruction.
First of all, to understand the impact of the effect, we need to analyze the full loop itself. The architecture of the Ryzen 9 3900X is Zen2. We can retrieve information about this on the AMD website or also WikiChip.
This architecture have 4 ALU and 3 AGU. This roughly means it can execute up to 4 instructions like add/and/cmp per cycle.
Here is the cost of each instruction of the loop (based on the Agner Fog instruction table for Zen1):
# Reciprocal throughput
loop0: add rax, 1 # 0.25
mov rdx, rax # 0.2
and rdx, 1 # 0.25
add rcx, rdx # 0.25
cmp rcx, 1000000000 # 0.25 | Fused 0.5-2 (2 if jumping)
jne loop0 # 0.5-2 |
As you can see, the 4 first computing instructions of the loop can be executed in ~1 cycle. The 2 last instructions can be merged by your processor into a faster one.
Your main issue is that this last jne instruction could be quite slow compared to the rest of the loop. So you are very likely to measure only the overhead of this instruction. From this point, things start to be complicated.
Engineer and researcher worked hard the last decades to reduce the cost of such instructions at (almost) any price. Nowadays, processors (like the Ryzen 9 3900X) use out-of-order instruction scheduling to execute the dependent instructions required by the jne instruction as soon as possible. Most processors can also predict the address of the next instruction to execute after the jne and fetch new instructions (eg. the one of the next loop iteration) while the other instruction of the current loop iteration are being executed.
Performing the fetch as soon as possible is important to prevent any stall in the execution pipeline of the processor (especially in your case).
From the AMD document "Software Optimization Guide for AMD Family 17h Models 30h and Greater Processors", we can read:
2.8.3 Loop Alignment:
For the processor loop alignment is not usually a significant issue. However, for hot loops, some further knowledge of trade-offs can be helpful. Since the processor can read an aligned 64-byte fetch block every cycle, aligning the end of the loop to the last byte of a 64-byte cache line is the best thing to do, if possible.
2.8.1.1 Next Address Logic
The next-address logic determines addresses for instruction fetch. [...]. When branches are identified, the next-address logic is redirected by the branch target and branch direction prediction hardware to generate a non-sequential fetch block address. The processor facilities that are designed to predict the next instruction to be executed following a branch are detailed in the following sections.
Thus, performing a conditional branching to instructions located in another cache line introduces an additional latency overhead due to a fetch of the Op Cache (an instruction cache faster than the L1) not required if the whole loop fit in 1 cache line. Indeed, if a loop is crossing a cache line, 2 line-cache fetches are required, which takes no less than 2 cycles. If the whole loop is fitting in a cache-line, only one 1 line-cache fetch is needed, which take only 1 cycle. As a result, since your loop iterations are very fast, paying 1 additional cycle introduces a significant slowdown. But how much?
You say that the program takes about 450 ms.
Since the Ryzen 9 3900X turbo frequency is 4.6 GHz and your loop is executed 2e9 times, the number cycles per loop iteration is 1.035. Note that this is better than what we can expected before (I guess this processor is able to rename registers, ignore the mov instruction, execute the jne instruction in parallel in only 1 cycle while other instructions of the loop are perfectly pipelined; which is astounding). This also shows that paying an additional fetch overhead of 1 cycle will double the number of cycles needed to execute each loop iteration and so the overall execution time.
If you do not want to pay this overhead, consider unrolling your loop to significantly reduce the number of conditional branches and non-sequential fetches.
This problem can occur on other architectures such as on Intel Skylake. Indeed, the same loop on a i5-9600KF takes 0.70s with loop alignment and 0.90s without (also due to the additional 1 cycle fetch latency). With a 8x unrolling, the result is 0.53s (whatever the alignment).
I've written some code for profiling small functions. At the high level it:
Sets the thread affinity to only one core and the thread priority to maximum.
Computes statistics from doing the following 100 times:
Estimate the latency of a function that does nothing.
Estimate the latency of the test function.
Subtract the first from the second to remove the cost of doing function-call overhead, thereby roughly getting the cost of the test function's contents.
To estimate the latency of a function, it:
Invalidates caches (this is difficult to actually do in user-mode, but I allocate and write a buffer the size of the L3 to memory, which should maybe help).
Yields the thread, so that the profile loop has as-few-as-possible context switches.
Gets the current time from a std::chrono::high_resolution_clock (which seems to compile to system_clock, but).
Runs the profile loop 100,000,000 times, calling the tested function within.
Gets the current time from a std::chrono::high_resolution_clock and subtracts to get latency.
Because at this level, individual instructions matter, at all points we have to write very careful code to ensure that the compiler doesn't elide, inline, cache, or treat-differently the functions. I have manually validated the generated assembly in various test cases, including the one which I present below.
I am getting extremely low (sub-nanosecond) latencies reported in some cases. I have tried everything I can think of to account for this, but cannot find an error.
I am looking for an explanation accounting for this behavior. Why are my profiled functions taking so little time?
Let's take the example of computing a square root for float.
The function signature is float(*)(float), and the empty function is trivial:
empty_function(float):
ret
Let's compute the square root by using the sqrtss instruction, and by the multiplication-by-reciprocal-square-root hack. I.e., the tested functions are:
sqrt_sseinstr(float):
sqrtss xmm0, xmm0
ret
sqrt_rcpsseinstr(float):
movaps xmm1, xmm0
rsqrtss xmm1, xmm0
mulss xmm0, xmm1
ret
Here's the profile loop. Again, this same code is called with the empty function and with the test functions:
double profile(float):
...
mov rbp,rdi
push rbx
mov ebx, 0x5f5e100
call 1c20 <invalidate_caches()>
call 1110 <sched_yield()>
call 1050 <std::chrono::high_resolution_clock::now()>
mov r12, rax
xchg ax, ax
15b0:
movss xmm0,DWORD PTR [rip+0xba4]
call rbp
sub rbx, 0x1
jne 15b0 <double profile(float)+0x20>
call 1050 <std::chrono::high_resolution_clock::now()>
...
The timing result for sqrt_sseinstr(float) on my Intel 990X is 3.60±0.13 nanoseconds. At this processor's rated 3.46 GHz, that works out to be 12.45±0.44 cycles. This seems pretty spot-on, given that the docs say the latency of sqrtss is around 13 cycles (it's not listed for this processor's Nehalem architecture, but it seems likely to also be around 13 cycles).
the timing result for sqrt_rcpsseinstr(float) is stranger: 0.01±0.07 nanoseconds (or 0.02±0.24 cycles). This is flatly implausible unless another effect is going on.
I thought perhaps the processor is able to hide the latency of the tested function somewhat or perfectly because the tested function uses different instruction ports (i.e. superscalarity is hiding something)? I tried to analyze this by hand, but didn't get very far because I didn't really know what I was doing.
(Note: I cleaned up some of the assembly notation for your convenience. An unedited objdump of the whole program, which includes several other variants, is here, and I am temporarily hosting the binary here (x86-64 SSE2+, Linux).)
The question, again: Why are some profiled functions producing implausibly small values? If it is a higher-order effect, explain?
The problem is with the basic approach of subtracting out the "latency"1 of an empty function, as described:
Estimate the latency of a function that does nothing.
Estimate the latency of the test function.
Subtract the first from the second to remove the cost of doing function-call overhead, thereby roughly getting the cost of the test
function's contents.
The built-in assumption is that the cost of calling a function is X, and if the latency of the work done in the function is Y, then the total cost will be something like X + Y.
This isn't generally true for any two blocks of work and especially isn't true when when one of them is "calling a function". A more sophisticated view would be that the total time would be somewhere between min(X, Y) and X + Y - but even this is often wrong depending on the details. Still, it's enough of a refinement to explain what is going on here: the cost of the function is not additive with the work being doing in the function: they happen in parallel.
The cost of an empty function call is something like 4 to 5 cycles on modern Intel, probably bottlenecked on the front-end throughput for the two taken branches, and possibly by branch and return predictor latency.
However, when you add additional work to an empty function, it generally won't compete for the same resources, and its execution instructions won't depend on the "output" of the call (i.e., the work will form a separate dependency chain), except perhaps in rare cases where the stack pointer is manipulated and the stack engine doesn't remove the dependency.
So essentially the function will take the greater of the time needed for the function call mechanics, or the actual work done by the function. This approximation isn't exact, because some types of work may actually add to the overhead of the function call (e.g., if there are enough instructions for the front end to get through before getting to the ret, the total time may increase on top of the 4-5 cycle empty function time, even if the total work is less than that) - but it's a good first order approximation.
Your first function takes enough time that the actual work dominates the execution time. The second function is much faster, however, enabling it to "hide under" the existing time taken by the call/ret mechanics.
The solution is simple: duplicate the work within the function N times, so that the work always dominates. N=10 or N=50 or something like that is fine. You have to decide whether you want to test latency, in which case the output of one copy of the work should feed into the next, or throughput, in which case it shouldn't.
On other hand, if you actually want to test the cost of the function call + work, e.g., because that's how you'll be using it in real life, it is likely the results you have gotten is already close to correct: stuff really can be "incrementally free" when it hides behind a function call.
1 I'm putting "latency" in quotes here because it isn't clear whether we should be talking about the latency of call/ret or the throughput. call and ret don't have any explicit outputs (and ret has no input), so it doesn't participate in a classic register-based dependency chain - but it might make sense to think of latency if you consider other hidden architectural components like the instruction pointer. In either case latency of throughput mostly points down to the same thing because all call and ret on a thread operate on the same state, so it doesn't make sense to have say "independent" vs "dependent" call chains.
Your benchmarking approach is fundamentally wrong, and your "careful code" is bogus.
First, emptying the cache is bogus. Not only will it quickly be repopulated with the required data, but also the examples you have posted have very little memory interaction (only cache access by call/ret and a load we'll get to.
Second, yielding before the benchmarking loop is bogus. You iterate 100000000 times, which even on a reasonably fast modern processor will take longer than typical scheduling clock interrupts on a stock operating system. If, on the other hand, you disable scheduling clock interrupts, then yielding before the benchmark doesn't do anything.
Now that the useless incidental complexity is out of the way, about the fundamental misunderstanding of modern CPUs:
You expect loop_time_gross/loop_count to be the time spent in each loop iteration. This is wrong. Modern CPUs do not execute instructions one after the other, sequentially. Modern CPUs pipeline, predict branches, execute multiple instructions in parallel, and (reasonably fast CPUs) out of order.
So after the first handful of iterations of the benchmarking loop, all branches are perfectly predicted for the next almost 100000000 iterations. This enables the CPU to speculate. Effectively, the conditional branch in the benchmarking loop goes away, as does most of the cost of the indirect call. Effectively, the CPU can unroll the loop:
movss xmm0, number
movaps xmm1, xmm0
rsqrtss xmm1, xmm0
mulss xmm0, xmm1
movss xmm0, number
movaps xmm1, xmm0
rsqrtss xmm1, xmm0
mulss xmm0, xmm1
movss xmm0, number
movaps xmm1, xmm0
rsqrtss xmm1, xmm0
mulss xmm0, xmm1
...
or, for the other loop
movss xmm0, number
sqrtss xmm0, xmm0
movss xmm0, number
sqrtss xmm0, xmm0
movss xmm0, number
sqrtss xmm0, xmm0
...
Notable, the load of number is always the same (thus quickly cached), and it overwrites the just computed value, breaking the dependency chain.
To be fair, the
call rbp
sub rbx, 0x1
jne 15b0 <double profile(float)+0x20>
are still executed, but the only resources they take from the floating-point code are decode/micro-op cache and execution ports. Notably, while the integer loop code has a dependency chain (ensuring a minimum execution time), the floating-point code does not carry a dependency on it. Furthermore, the floating-point code consists of many mutually totally independent short dependency chains.
Where you expect the CPU to execute instructions sequentially, the CPU can instead execute them in parallel.
A small look at https://agner.org/optimize/instruction_tables.pdf reveals why this parallel execution doesn't work for sqrtss on Nehalem:
instruction: SQRTSS/PS
latency: 7-18
reciprocal throughput: 7-18
i.e., the instruction cannot be pipelined and only runs on one execution port.
In contrast, for movaps, rsqrtss, mulss:
instruction: MOVAPS/D
latency: 1
reciprocal throughput: 1
instruction: RSQRTSS
latency: 3
reciprocal throughput: 2
instruction: MULSS
latency: 4
reciprocal throughput: 1
the maximum reciprocal throughput of the dependency chain is 2, so you can expect the code to finish executing one dependency chain every 2 cycles in the steady state. At this point, the execution time of the floating-point part of the benchmarking loop is less than or equal to the loop overhead and overlapped with it, so your naive approach to subtract the loop overhead leads to nonsensical results.
If you wanted to do this properly, you would ensure that separate loop iterations are dependent on each other, for example by changing your benchmarking loop to
float x = INITIAL_VALUE;
for (i = 0; i < 100000000; i++)
x = benchmarked_function(x);
Obviously you will not benchmark the same input this way, unless INITIAL_VALUE is a fixed point of benchmarked_function(). However, you can arrange for it to be a fixed point of an expanded function by computing float diff = INITIAL_VALUE - benchmarked_function(INITIAL_VALUE); and then making the loop
float x = INITIAL_VALUE;
for (i = 0; i < 100000000; i++)
x = diff + benchmarked_function(x);
with relatively minor overhead, though you should then ensure that floating-point errors do not accumulate to significantly change the value passed to benchmarked_function().
I use this code to test the impact of dependency in a loop iteration on IvyBridge:
global _start
_start:
mov rcx, 1000000000
.for_loop:
inc rax ; uop A
inc rax ; uop B
dec rcx ; uop C
jnz .for_loop
xor rdi, rdi
mov rax, 60 ; _exit(0)
syscall
Since dec and jnz will be macro-fused to a single uop, there are 3 uops in my loop, they are labeled in the comments.
uop B depends on uop A, so I think the execution would be like this:
A C
B A C ; the previous B and current A can be in the same cycle
B A C
...
B A C
B
Therefore the loop can be executed 1 cycle per iter.
However, the perf tool shows:
2,009,704,779 cycles
1,008,054,984 stalled-cycles-frontend # 50.16% frontend cycles idl
So it's 2 cycle per iter, and there are 50% frontend cycle idle.
What caused the frontend 50% idle? Why the hypothetical execution diagram can't be realized?
B and A form a loop-carried dependency chain. A in the next iteration can't run until it has the result of B in the previous.
Any given B can never run in the same cycle as an A: what input would the later one use, if the earlier one hasn't produced a result yet?
This chain is 2 cycles long (per iteration), because the latency of inc is 1 cycle. This creates a latency bottleneck in the back-end that out-of-order execution can't hide. (Except for very low iteration counts where it can overlap it with code after the loop).
Just like if you fully unrolled a huge chain of times 102400 inc eax, there's no instruction-level parallelism for the CPU to find between a chain of instructions that each depend on the previous.
The macro-fused dec rcx/jnz uop is independent of the RAX chain, and is a shorter chain (only 1 cycle per iteration, being only 1 dec&branch uop with 1c latency). So it can run in parallel with B or A uops.
See my answer on another question for more about the concept of instruction-level parallelism and dependency chains, and how CPUs exploit that parallelism to run instruction in parallel when they're independent.
Agner Fog's microarch PDF shows this with examples in an early chapter: Chapter 2: Out-of-order execution (All processors except P1,
PMMX).
If you started a new 2-cycle dep chain every iteration, it would run as you expect. A new chain forking off every iteration would expose instruction-level parallelism for the CPU to keep A and B from different iterations in flight at the same time.
.for_loop:
xor eax,eax ; dependency-breaking for RAX
inc rax ; uop A
inc rax ; uop B
dec rcx ; uop C
jnz .for_loop
Sandybridge-family handles xor-zeroing without an execution unit, so this is still only 3 unfused-domain uops in the loop, so IvyBridge has enough ALU execution ports to run all 3 in a single cycle. This also maxes out the front-end at 4 fused-domain uops per clock.
Or if you changed A to start a new dep chain in RAX with any instruction that unconditionally overwrites RAX without depending on the result of the inc, you'd be fine.
lea rax, [rdx + rdx] ; no dependency on B from last iter
inc rax ; uop B
Except for a couple instructions with an unfortunate output dependency: Why does breaking the "output dependency" of LZCNT matter?
popcnt rax, rdx ; false dependency on RAX, 3 cycle latency
inc rax ; uop B
On Intel CPUs, only popcnt, and lzcnt/tzcnt have an output dependency for no reason. It's because they use the same execution unit as bsf/bsr, which leave the destination unmodified if the input is zero, on Intel and AMD CPUs. Intel still only documents it on paper as undefined if the input is zero for BSF/BSR, but they build hardware that implements stronger guarantees. (AMD does even document this BSF/BSR behaviour.) Anyway, so Intel's BSF/BSR are like CMOV, and need the destination as an input in case the source reg is 0. popcnt, (and lzcnt/tzcnt on pre-Skylake) suffer from this, too.
If you made the loop more than 5 fused-domain uops, SnB/IvB could issue it at best 1 per 2 cycles from the front-end. Haswell and later "unroll" in the loop buffer or something so a 5 uop loop can run at ~1.25 c per iteration, but SnB/IvB don't. Is performance reduced when executing loops whose uop count is not a multiple of processor width?
The front-end issue/rename stage is 4 fused-domain uops wide in Intel CPUs since Core 2.
I'm looking for a type of a formula / way to measure how fast an instruction is, or more specific to give a "score" each of the instruction by CPU cycles.
Let's take the follow assembly program for an example,
nop
mov eax,dword ptr [rbp+34h]
inc eax
mov dword ptr [rbp+34h],eax
and the following Intel Skylake information:
mov r,m : Throughput=0.5 Latency=2
mov m,r
: Throughput=1 Latency=2
nop : Throughput=0.25 Latency=non
inc : Throughput=0.25 Latency=1
I know that the order of the instructions in the program are matter in here but
I'm looking to create something general that not need to be "accurate to the single cycle"
any one have any idea how can I do that?
There is no formula you can apply; you have to measure.
The same instruction on different versions of the same uarch family can have different performance. e.g. mulps:
Sandybridge 1c / 5c throughput/latency.
HSW 0.5 / 5. BDW 0.5 / 3 (faster multiply path in the FMA unit? FMA is still 5c).
SKL 0.5 / 4 (lower latency FMA, too). SKL runs addps on the FMA unit as well, dropping the dedicated FP multiply unit so add latency is higher, but throughput is higher.
There's no way you could predict any of this without measuring, or knowing some microarchitectural details. We expect FP math ops won't be single-cycle latency, because they're much more complicated than integer ops. (So if they were single cycle, the clock speed is set too low for integer ops.)
You measure by repeating the instruction many times in an unrolled loop. Or fully unrolled with no looping, but then you defeat the uop-cache and can get front-end bottlenecks. (e.g. for decoding 10-byte mov r64, imm64)
https://uops.info/ has already automated this testing for every form of every (unprivileged) instruction, and you can even click on any table entry to see what test loops they used. e.g. Skylake xchg r32, eax latency testing (https://uops.info/html-lat/SKL/XCHG_R32_EAX-Measurements.html) from each input operand to each output. (2 cycle latency from EAX -> R8D, but 1 cycle latency from R8D -> EAX.) So we can guess that the 3 uops include copying EAX to an internal temporary, but moving directly from the other operand to EAX.
https://uops.info/ is the current best source of test data; when it and Agner's tables disagree, my own measurements and/or other sources have always confirmed uops.info's testing was accurate. And they don't try to make up a latency number for 2 halves of a round-trip like movd xmm0,eax and back, they show you the range of possible latencies assuming the rest of the chain was the minimum plausible.
Agner Fog creates his instruction tables (which you appear to be reading) by timing large non-looping blocks of code that repeat an instruction. https://agner.org/optimize/. The intro section of his instruction-tables explains briefly how he measures, and his microarch guide explains more details of how different x86 microarchitectures work internally. Unfortunately there are occasional typos or copy/paste errors in his hand-edited tables.
http://instlatx64.atw.hu/ also has results of experimental measurements. I think they use a similar technique of a large block of the same instruction repeated, maybe small enough to fit in the uop cache. But they don't use perf counters to measure what execution port each instruction needs, so their throughput numbers don't help you figure out which instructions compete with which other instructions.
These latter two sources have been around for longer than uops.info, and cover some older CPUs, especially older AMD.
To measure latency yourself, you make the output of each instruction an input for the next.
mov ecx, 10000000
inc_latency:
inc eax
inc eax
inc eax
inc eax
inc eax
inc eax
sub ecx,1 ; avoid partial-flag false dep for P4
jnz inc_latency ; dec or sub/jnz macro-fuses into 1 uop on Intel SnB-family
This dependency chain of 7 inc instructions will bottleneck the loop at 1 iteration per 7 * inc_latency cycles. Using perf counters for core clock cycles (not RDTSC cycles), you can easily measure the time for all the iterations to 1 part in 10k, and with more care probably even more precisely than that. The repeat count of 10000000 hides start/stop overhead of whatever timing you use.
I normally put a loop like this in a Linux static executable that just makes a sys_exit(0) system call directly (with a syscall) instruction, and time the whole executable with perf stat ./testloop to get time and a cycle count. (See Can x86's MOV really be "free"? Why can't I reproduce this at all? for an example).
Another example is Understanding the impact of lfence on a loop with two long dependency chains, for increasing lengths, with the added complication of using lfence to drain the out-of-order execution window for two dep chains.
To measure throughput, you use separate registers, and/or include an xor-zeroing occasionally to break dep chains and let out-of-order exec overlap things. Don't forget to also use perf counters to see which ports it can run on, so you can tell which other instructions it will compete with. (e.g. FMA (p01) and shuffles (p5) don't compete at all for back-end resources on Haswell/Skylake, only for front-end throughput.) Don't forget to measure front-end uop counts, too: some instructions decode to multiply uops.
How many different dependency chains do we need to avoid a bottleneck? Well we know the latency (measure it first), and we know the max possible throughput (number of execution ports, or front-end throughput.)
For example, if FP multiply had 0.25c throughput (4 per clock), we could keep 20 in flight at once on Haswell (5c latency). That's more than we have registers, so we could just use all 16 and discover that in fact the throughput is only 0.5c. But if it had turned out that 16 registers was a bottleneck, we could add xorps xmm0,xmm0 occasionally and let out-of-order execution overlap some blocks.
More is normally better; having just barely enough to hide latency can slow down with imperfect scheduling. If we wanted to go nuts measuring inc, we'd do this:
mov ecx, 10000000
inc_latency:
%rep 10 ;; source-level repeat of a block, no runtime branching
inc eax
inc ebx
; not ecx, we're using it as a loop counter
inc edx
inc esi
inc edi
inc ebp
inc r8d
inc r9d
inc r10d
inc r11d
inc r12d
inc r13d
inc r14d
inc r15d
%endrep
sub ecx,1 ; break partial-flag false dep for P4
jnz inc_latency ; dec/jnz macro-fuses into 1 uop on Intel SnB-family
If we were worried about partial-flag false dependencies or flag-merging effects, we might experiment with mixing in an xor eax,eax somewhere to let OoO exec overlap more than just when sub wrote all flags. (See INC instruction vs ADD 1: Does it matter?)
There's a similar problem for measuring throughput and latency of shl r32, cl on Sandybridge-family: the flag dependency chain isn't normally relevant for a computation, but putting shl back-to-back creates a dependency through FLAGS as well as through the register. (Or for throughput, there isn't even a register dep).
I posted about this on Agner Fog's blog: https://www.agner.org/optimize/blog/read.php?i=415#860. I mixed shl edx,cl in with four add edx,1 instructions, to see what incremental slowdown adding one more instruction had, where the FLAGS dependency was a non-issue. On SKL, it only slows down by an extra 1.23 cycles on average, so the true latency cost of that shl was only ~1.23 cycles, not 2. (It's not a whole number or just 1 because of resource conflicts to run the flag-merging uops of the shl, I guess. BMI2 shlx edx, edx, ecx would be exactly 1c because it's only a single uop.)
Related: for static performance analysis of whole blocks of code (containing different instructions), see What considerations go into predicting latency for operations on modern superscalar processors and how can I calculate them by hand?. (It's using the word "latency" for the end-to-end latency of a whole computation, but actually asking about things small enough for OoO exec to overlap different parts, so instruction latency and throughput both matter.)
The Latency=2 numbers for load/store appear to be from Agner Fog's instruction tables (https://agner.org/optimize/). They unfortunately aren't accurate for a chain of mov rax, [rax]. You'll find that's 4c
latency if you measure it by putting that in a loop.
Agner splits up load/store latency into something that makes the total store/reload latency come out correct, but for some reason he doesn't make the load part equal to the L1d load-use latency when it comes from cache instead of the store buffer. (But also note that if the load feeds an ALU instruction instead of another load, the latency is 5c. So the simple addressing-mode fast-path only helps for pure pointer-chasing.)
I was playing with the code in this answer, slightly modifying it:
BITS 64
GLOBAL _start
SECTION .text
_start:
mov ecx, 1000000
.loop:
;T is a symbol defined with the CLI (-DT=...)
TIMES T imul eax, eax
lfence
TIMES T imul edx, edx
dec ecx
jnz .loop
mov eax, 60 ;sys_exit
xor edi, edi
syscall
Without the lfence I the results I get are consistent with the static analysis in that answer.
When I introduce a single lfence I'd expect the CPU to execute the imul edx, edx sequence of the k-th iteration in parallel with the imul eax, eax sequence of the next (k+1-th) iteration.
Something like this (calling A the imul eax, eax sequence and D the imul edx, edx one):
|
| A
| D A
| D A
| D A
| ...
| D A
| D
|
V time
Taking more or less the same number of cycles but for one unpaired parallel execution.
When I measure the number of cycles, for the original and modified version, with taskset -c 2 ocperf.py stat -r 5 -e cycles:u '-x ' ./main-$T for T in the range below I get
T Cycles:u Cycles:u Delta
lfence no lfence
10 42047564 30039060 12008504
15 58561018 45058832 13502186
20 75096403 60078056 15018347
25 91397069 75116661 16280408
30 108032041 90103844 17928197
35 124663013 105155678 19507335
40 140145764 120146110 19999654
45 156721111 135158434 21562677
50 172001996 150181473 21820523
55 191229173 165196260 26032913
60 221881438 180170249 41711189
65 250983063 195306576 55676487
70 281102683 210255704 70846979
75 312319626 225314892 87004734
80 339836648 240320162 99516486
85 372344426 255358484 116985942
90 401630332 270320076 131310256
95 431465386 285955731 145509655
100 460786274 305050719 155735555
How can the values of Cycles:u lfence be explained?
I would have expected them to be similar to those of Cycles:u no lfence since a single lfence should prevent only the first iteration from being executed in parallel for the two blocks.
I don't think it's due to the lfence overhead as I believe that should be constant for all Ts.
I'd like to fix what's wrong with my forma mentis when dealing with the static analysis of code.
Supporting repository with source files.
I think you're measuring accurately, and the explanation is microarchitectural, not any kind of measurement error.
I think your results for mid to low T support the conclusion that lfence stops the front-end from even issuing past the lfence until all earlier instructions retire, rather than having all the uops from both chains already issued and just waiting for lfence to flip a switch and let multiplies from each chain start to dispatch on alternating cycles.
(port1 would get edx,eax,empty,edx,eax,empty,... for Skylake's 3c latency / 1c throughput multiplier right away, if lfence didn't block the front-end, and overhead wouldn't scale with T.)
You're losing imul throughput when only uops from the first chain are in the scheduler because the front-end hasn't chewed through the imul edx,edx and loop branch yet. And for the same number of cycles at the end of the window when the pipeline is mostly drained and only uops from the 2nd chain are left.
The overhead delta looks linear up to about T=60. I didn't run the numbers, but the slope up to there looks reasonable for T * 0.25 clocks to issue the first chain vs. 3c-latency execution bottleneck. i.e. delta growing maybe 1/12th as fast as total no-lfence cycles.
So (given the lfence overhead I measured below), with T<60:
no_lfence cycles/iter ~= 3T # OoO exec finds all the parallelism
lfence cycles/iter ~= 3T + T/4 + 9.3 # lfence constant + front-end delay
delta ~= T/4 + 9.3
#Margaret reports that T/4 is a better fit than 2*T / 4, but I would have expected T/4 at both the start and end, for a total of 2T/4 slope of the delta.
After about T=60, delta grows much more quickly (but still linearly), with a slope about equal to the total no-lfence cycles, thus about 3c per T. I think at that point, the scheduler (Reservation Station) size is limiting the out-of-order window. You probably tested on a Haswell or Sandybridge/IvyBridge, (which have a 60-entry or 54-entry scheduler respectively. Skylake's is 97 entry (but not fully unified; IIRC BeeOnRope's testing showed that not all the entries could be used for any type of uop. Some were specific to load and/or store, for example.)
The RS tracks un-executed uops. Each RS entry holds 1 unfused-domain uop that's waiting for its inputs to be ready, and its execution port, before it can dispatch and leave the RS1.
After an lfence, the front-end issues at 4 per clock while the back-end executes at 1 per 3 clocks, issuing 60 uops in ~15 cycles, during which time only 5 imul instructions from the edx chain have executed. (There's no load or store micro-fusion here, so every fused-domain uop from the front-end is still only 1 unfused-domain uop in the RS2.)
For large T the RS quickly fills up, at which point the front-end can only make progress at the speed of the back-end. (For small T, we hit the next iteration's lfence before that happens, and that's what stalls the front-end). When T > RS_size, the back-end can't see any of the uops from the eax imul chain until enough back-end progress through the edx chain has made room in the RS. At that point, one imul from each chain can dispatch every 3 cycles, instead of just the 1st or 2nd chain.
Remember from the first section that time spent just after lfence only executing the first chain = time just before lfence executing only the second chain. That applies here, too.
We get some of this effect even with no lfence, for T > RS_size, but there's opportunity for overlap on both sides of a long chain. The ROB is at least twice the size of the RS, so the out-of-order window when not stalled by lfence should be able to keep both chains in flight constantly even when T is somewhat larger than the scheduler capacity. (Remember that uops leave the RS as soon as they've executed. I'm not sure if that means they have to finish executing and forward their result, or merely start executing, but that's a minor difference here for short ALU instructions. Once they're done, only the ROB is holding onto them until they retire, in program order.)
The ROB and register-file shouldn't be limiting the out-of-order window size (http://blog.stuffedcow.net/2013/05/measuring-rob-capacity/) in this hypothetical situation, or in your real situation. They should both be plenty big.
Blocking the front-end is an implementation detail of lfence on Intel's uarches. The manual only says that later instructions can't execute. That wording would allow the front-end to issue/rename them all into the scheduler (Reservation Station) and ROB while lfence is still waiting, as long as none are dispatched to an execution unit.
So a weaker lfence would maybe have flat overhead up to T=RS_size, then the same slope as you see now for T>60. (And the constant part of the overhead might be lower.)
Note that guarantees about speculative execution of conditional/indirect branches after lfence apply to execution, not (as far as I know) to code-fetch. Merely triggering code-fetch is not (AFAIK) useful to a Spectre or Meltdown attack. Possibly a timing side-channel to detect how it decodes could tell you something about the fetched code...
I think AMD's LFENCE is at least as strong on actual AMD CPUs, when the relevant MSR is enabled. (Is LFENCE serializing on AMD processors?).
Extra lfence overhead:
Your results are interesting, but it doesn't surprise me at all that there's significant constant overhead from lfence itself (for small T), as well as the component that scales with T.
Remember that lfence doesn't allow later instructions to start until earlier instructions have retired. This is probably at least a couple cycles / pipeline-stages later than when their results are ready for bypass-fowarding to other execution units (i.e. the normal latency).
So for small T, it's definitely significant that you add extra latency into the chain by requiring the result to not only be ready, but also written back to the register file.
It probably takes an extra cycle or so for lfence to allow the issue/rename stage to start operating again after detecting retirement of the last instruction before it. The issue/rename process takes multiple stages (cycles), and maybe lfence blocks at the start of this, instead of in the very last step before uops are added into the OoO part of the core.
Even back-to-back lfence itself has 4 cycle throughput on SnB-family, according to Agner Fog's testing. Agner Fog reports 2 fused-domain uops (no unfused), but on Skylake I measure it at 6 fused-domain (still no unfused) if I only have 1 lfence. But with more lfence back-to-back, it's fewer uops! Down to ~2 uops per lfence with many back-to-back, which is how Agner measures.
lfence/dec/jnz (a tight loop with no work) runs at 1 iteration per ~10 cycles on SKL, so that might give us an idea of the real extra latency that lfence adds to the dep chains even without the front-end and RS-full bottlenecks.
Measuring lfence overhead with only one dep chain, OoO exec being irrelevant:
.loop:
;mfence ; mfence here: ~62.3c (with no lfence)
lfence ; lfence here: ~39.3c
times 10 imul eax,eax ; with no lfence: 30.0c
; lfence ; lfence here: ~39.6c
dec ecx
jnz .loop
Without lfence, runs at the expected 30.0c per iter. With lfence, runs at ~39.3c per iter, so lfence effectively added ~9.3c of "extra latency" to the critical path dep chain. (And 6 extra fused-domain uops).
With lfence after the imul chain, right before the loop-branch, it's slightly slower. But not a whole cycle slower, so that would indicate that the front-end is issuing the loop-branch + and imul in a single issue-group after lfence allows execution to resume. That being the case, IDK why it's slower. It's not from branch misses.
Getting the behaviour you were expecting:
Interleave the chains in program order, like #BeeOnRope suggests in comments, doesn't require out-of-order execution to exploit the ILP, so it's pretty trivial:
.loop:
lfence ; at the top of the loop is the lowest-overhead place.
%rep T
imul eax,eax
imul edx,edx
%endrep
dec ecx
jnz .loop
You could put pairs of short times 8 imul chains inside a %rep to let OoO exec have an easy time.
Footnote 1: How the front-end / RS / ROB interact
My mental model is that the issue/rename/allocate stages in the front-end add new uops to both the RS and the ROB at the same time.
Uops leave the RS after executing, but stay in the ROB until in-order retirement. The ROB can be large because it's never scanned out-of-order to find the first-ready uop, only scanned in-order to check if the oldest uop(s) have finished executing and thus are ready to retire.
(I assume the ROB is physically a circular buffer with start/end indices, not a queue which actually copies uops to the right every cycle. But just think of it as a queue / list with a fixed max size, where the front-end adds uops at the front, and the retirement logic retires/commits uops from the end as long as they're fully executed, up to some per-cycle per-hyperthread retirement limit which is not usually a bottleneck. Skylake did increase it for better Hyperthreading, maybe to 8 per clock per logical thread. Perhaps retirement also means freeing physical registers which helps HT, because the ROB itself is statically partitioned when both threads are active. That's why retirement limits are per logical thread.)
Uops like nop, xor eax,eax, or lfence, which are handled in the front-end (don't need any execution units on any ports) are added only to the ROB, in an already-executed state. (A ROB entry presumably has a bit that marks it as ready to retire vs. still waiting for execution to complete. This is the state I'm talking about. For uops that did need an execution port, I assume the ROB bit is set via a completion port from the execution unit. And that the same completion-port signal frees its RS entry.)
Uops stay in the ROB from issue to retirement.
Uops stay in the RS from issue to execution. The RS can replay uops in a few cases, e.g. for the other half of a cache-line-split load, or if it was dispatched in anticipation of load data arriving, but in fact it didn't. (Cache miss or other conflicts like Weird performance effects from nearby dependent stores in a pointer-chasing loop on IvyBridge. Adding an extra load speeds it up?) Or when a load port speculates that it can bypass the AGU before starting a TLB lookup to shorten pointer-chasing latency with small offsets - Is there a penalty when base+offset is in a different page than the base?
So we know that the RS can't remove a uop right as it dispatches, because it might need to be replayed. (Can happen even to non-load uops that consume load data.) But any speculation that needs replays is short-range, not through a chain of uops, so once a result comes out the other end of an execution unit, the uop can be removed from the RS. Probably this is part of what a completion port does, along with putting the result on the bypass forwarding network.
Footnote 2: How many RS entries does a micro-fused uop take?
TL:DR: P6-family: RS is fused, SnB-family: RS is unfused.
A micro-fused uop is issued to two separate RS entries in Sandybridge-family, but only 1 ROB entry. (Assuming it isn't un-laminated before issue, see section 2.3.5 for HSW or section 2.4.2.4 for SnB of Intel's optimization manual, and Micro fusion and addressing modes. Sandybridge-family's more compact uop format can't represent indexed addressing modes in the ROB in all cases.)
The load can dispatch independently, ahead of the other operand for the ALU uop being ready. (Or for micro-fused stores, either of the store-address or store-data uops can dispatch when its input is ready, without waiting for both.)
I used the two-dep-chain method from the question to experimentally test this on Skylake (RS size = 97), with micro-fused or edi, [rdi] vs. mov+or, and another dep chain in rsi. (Full test code, NASM syntax on Godbolt)
; loop body
%rep T
%if FUSE
or edi, [rdi] ; static buffers are in the low 32 bits of address space, in non-PIE
%else
mov eax, [rdi]
or edi, eax
%endif
%endrep
%rep T
%if FUSE
or esi, [rsi]
%else
mov eax, [rsi]
or esi, eax
%endif
%endrep
Looking at uops_executed.thread (unfused-domain) per cycle (or per second which perf calculates for us), we can see a throughput number that doesn't depend on separate vs. folded loads.
With small T (T=30), all the ILP can be exploited, and we get ~0.67 uops per clock with or without micro-fusion. (I'm ignoring the small bias of 1 extra uop per loop iteration from dec/jnz. It's negligible compared to the effect we'd see if micro-fused uops only used 1 RS entry)
Remember that load+or is 2 uops, and we have 2 dep chains in flight, so this is 4/6, because or edi, [rdi] has 6 cycle latency. (Not 5, which is surprising, see below.)
At T=60, we still have about 0.66 unfused uops executed per clock for FUSE=0, and 0.64 for FUSE=1. We can still find basically all the ILP, but it's just barely starting to dip, as the two dep chains are 120 uops long (vs. a RS size of 97).
At T=120, we have 0.45 unfused uops per clock for FUSE=0, and 0.44 for FUSE=1. We're definitely past the knee here, but still finding some of the ILP.
If a micro-fused uop took only 1 RS entry, FUSE=1 T=120 should be about the same speed as FUSE=0 T=60, but that's not the case. Instead, FUSE=0 or 1 makes nearly no difference at any T. (Including larger ones like T=200: FUSE=0: 0.395 uops/clock, FUSE=1: 0.391 uops/clock). We'd have to go to very large T before we start for the time with 1 dep-chain in flight to totally dominate the time with 2 in flight, and get down to 0.33 uops / clock (2/6).
Oddity: We have such a small but still measurable difference in throughput for fused vs. unfused, with separate mov loads being faster.
Other oddities: the total uops_executed.thread is slightly lower for FUSE=0 at any given T. Like 2,418,826,591 vs. 2,419,020,155 for T=60. This difference was repeatable down to +- 60k out of 2.4G, plenty precise enough. FUSE=1 is slower in total clock cycles, but most of the difference comes from lower uops per clock, not from more uops.
Simple addressing modes like [rdi] are supposed to only have 4 cycle latency, so load + ALU should be only 5 cycle. But I measure 6 cycle latency for the load-use latency of or rdi, [rdi], or with a separate MOV-load, or with any other ALU instruction I can never get the load part to be 4c.
A complex addressing mode like [rdi + rbx + 2064] has the same latency when there's an ALU instruction in the dep chain, so it appears that Intel's 4c latency for simple addressing modes only applies when a load is forwarding to the base register of another load (with up to a +0..2047 displacement and no index).
Pointer-chasing is common enough that this is a useful optimization, but we need to think of it as a special load-load forwarding fast-path, not as a general data ready sooner for use by ALU instructions.
P6-family is different: an RS entry holds a fused-domain uop.
#Hadi found an Intel patent from 2002, where Figure 12 shows the RS in the fused domain.
Experimental testing on a Conroe (first gen Core2Duo, E6600) shows that there's a large difference between FUSE=0 and FUSE=1 for T=50. (The RS size is 32 entries).
T=50 FUSE=1: total time of 2.346G cycles (0.44IPC)
T=50 FUSE=0: total time of 3.272G cycles (0.62IPC = 0.31 load+OR per clock). (perf / ocperf.py doesn't have events for uops_executed on uarches before Nehalem or so, and I don't have oprofile installed on that machine.)
T=24 there's a negligible difference between FUSE=0 and FUSE=1, around 0.47 IPC vs 0.9 IPC (~0.45 load+OR per clock).
T=24 is still over 96 bytes of code in the loop, too big for Core 2's 64-byte (pre-decode) loop buffer, so it's not faster because of fitting in a loop buffer. Without a uop-cache, we have to be worried about the front-end, but I think we're fine because I'm exclusively using 2-byte single-uop instructions that should easily decode at 4 fused-domain uops per clock.
I'll present an analysis for the case where T = 1 for both codes (with and without lfence). You can then extend this for other values of T. You can refer to Figure 2.4 of the Intel Optimization Manual for a visual.
Because there is only a single easily predicted branch, the frontend will only stall if the backend stalled. The frontend is 4-wide in Haswell, which means up to 4 fused uops can be issued from the IDQ (instruction decode queue, which is just a queue that holds in-order fused-domain uops, also called the uop queue) to the reservation station (RS) entires of the scheduler. Each imul is decoded into a single uop that cannot be fused. The instructions dec ecx and jnz .loop get macrofused in the frontend to a single uop. One of the differences between microfusion and macrofusion is that when the scheduler dispatches a macrofused uop (that are not microfused) to the execution unit it's assigned to, it gets dispatched as a single uop. In contrast, a microfused uop needs to be split into its constituent uops, each of which must be separately dispatched to an execution unit. (However, splitting microfused uops happens on entrance to the RS, not on dispatch, see Footnote 2 in #Peter's answer). lfence is decoded into 6 uops. Recognizing microfusion only matters in the backend, and in this case, there is no microfusion in the loop.
Since the loop branch is easily predictable and since the number of iterations is relatively large, we can just assume without compromising accuracy that the allocator will always be able to allocate 4 uops per cycle. In other words, the scheduler will receive 4 uops per cycle. Since there is no micorfusion, each uop will be dispatched as a single uop.
imul can only be executed by the Slow Int execution unit (see Figure 2.4). This means that the only choice for executing the imul uops is to dispatch them to port 1. In Haswell, the Slow Int is nicely pipelined so that a single imul can be dispatched per cycle. But it takes three cycles for the result of the multiplication be available for any instruction that requires (the writeback stage is the third cycle from the dispatch stage of the pipeline). So for each dependence chain, at most one imul can be dispatched per 3 cycles.
Becausedec/jnz is predicted taken, the only execution unit that can execute it is Primary Branch on port 6.
So at any given cycle, as long as the RS has space, it will receive 4 uops. But what kind of uops? Let's examine the loop without lfence:
imul eax, eax
imul edx, edx
dec ecx/jnz .loop (macrofused)
There are two possibilities:
Two imuls from the same iteration, one imul from a neighboring iteration, and one dec/jnz from one of those two iterations.
One dec/jnz from one iteration, two imuls from the next iteration, and one dec/jnz from the same iteration.
So at the beginning of any cycle, the RS will receive at least one dec/jnz and at least one imul from each chain. At the same time, in the same cycle and from those uops that are already there in the RS, the scheduler will do one of two actions:
Dispatch the oldest dec/jnz to port 6 and dispatch the oldest imul that is ready to port 1. That's a total of 2 uops.
Because the Slow Int has a latency of 3 cycles but there are only two chains, for each cycle of 3 cycles, no imul in the RS will be ready for execution. However, there is always at least one dec/jnz in the RS. So the scheduler can dispatch that. That's a total of 1 uop.
Now we can calculate the expected number of uops in the RS, XN, at the end of any given cycle N:
XN = XN-1 + (the number of uops to be allocated in the RS at the beginning of cycle N) - (the expected number of uops that will be dispatched at the beginning of cycle N)
= XN-1 + 4 - ((0+1)*1/3 + (1+1)*2/3)
= XN-1 + 12/3 - 5/3
= XN-1 + 7/3 for all N > 0
The initial condition for the recurrence is X0 = 4. This is a simple recurrence that can be solved by unfolding XN-1.
XN = 4 + 2.3 * N for all N >= 0
The RS in Haswell has 60 entries. We can determine the first cycle in which the RS is expected to become full:
60 = 4 + 7/3 * N
N = 56/2.3 = 24.3
So at the end of cycle 24.3, the RS is expected to be full. This means that at the beginning of cycle 25.3, the RS will not be able to receive any new uops. Now the number of iterations, I, under consideration determines how you should proceed with the analysis. Since a dependency chain will require at least 3*I cycles to execute, it takes about 8.1 iterations to reach cycle 24.3. So if the number of iterations is larger than 8.1, which is the case here, you need to analyze what happens after cycle 24.3.
The scheduler dispatches instructions at the following rates every cycle (as discussed above):
1
2
2
1
2
2
1
2
.
.
But the allocator will not allocate any uops in the RS unless there are at least 4 available entries. Otherwise, it will not waste power on issuing uops at a sub-optimal throughput. However, it is only at the beginning of every 4th cycle are there at least 4 free entries in the RS. So starting from cycle 24.3, the allocator is expected to get stalled 3 out of every 4 cycles.
Another important observation for the code being analyzed is that it never happens that there are more than 4 uops that can be dispatched, which means that the average number of uops that leave their execution units per cycle is not larger than 4. At most 4 uops can be retired from the ReOrder Buffer (ROB). This means that the ROB can never be on the critical path. In other words, performance is determined by the dispatch throughput.
We can calculate the IPC (instructions per cycles) fairly easily now. The ROB entries look something like this:
imul eax, eax - N
imul edx, edx - N + 1
dec ecx/jnz .loop - M
imul eax, eax - N + 3
imul edx, edx - N + 4
dec ecx/jnz .loop - M + 1
The column to the right shows the cycles in which the instruction can be retired. Retirement happens in order and is bounded by the latency of the critical path. Here each dependency chain have the same path length and so both constitute two equal critical paths of length 3 cycles. So every 3 cycles, 4 instructions can be retired. So the IPC is 4/3 = 1.3 and the CPI is 3/4 = 0.75. This is much smaller than the theoretical optimal IPC of 4 (even without considering micro- and macro-fusion). Because retirement happens in-order, the retirement behavior will be the same.
We can check our analysis using both perf and IACA. I'll discuss perf. I've a Haswell CPU.
perf stat -r 10 -e cycles:u,instructions:u,cpu/event=0xA2,umask=0x10,name=RESOURCE_STALLS.ROB/u,cpu/event=0x0E,umask=0x1,cmask=1,inv=1,name=UOPS_ISSUED.ANY/u,cpu/event=0xA2,umask=0x4,name=RESOURCE_STALLS.RS/u ./main-1-nolfence
Performance counter stats for './main-1-nolfence' (10 runs):
30,01,556 cycles:u ( +- 0.00% )
40,00,005 instructions:u # 1.33 insns per cycle ( +- 0.00% )
0 RESOURCE_STALLS.ROB
23,42,246 UOPS_ISSUED.ANY ( +- 0.26% )
22,49,892 RESOURCE_STALLS.RS ( +- 0.00% )
0.001061681 seconds time elapsed ( +- 0.48% )
There are 1 million iterations each takes about 3 cycles. Each iteration contains 4 instructions and the IPC is 1.33.RESOURCE_STALLS.ROB shows the number of cycles in which the allocator was stalled due to a full ROB. This of course never happens. UOPS_ISSUED.ANY can be used to count the number of uops issued to the RS and the number of cycles in which the allocator was stalled (no specific reason). The first is straightforward (not shown in the perf output); 1 million * 3 = 3 million + small noise. The latter is much more interesting. It shows that about 73% of all time the allocator stalled due to a full RS, which matches our analysis. RESOURCE_STALLS.RS counts the number of cycles in which the allocator was stalled due to a full RS. This is close to UOPS_ISSUED.ANY because the allocator does not stall for any other reason (although the difference could be proportional to the number of iterations for some reason, I'll have to see the results for T>1).
The analysis of the code without lfence can be extended to determine what happens if an lfence was added between the two imuls. Let's check out the perf results first (IACA unfortunately does not support lfence):
perf stat -r 10 -e cycles:u,instructions:u,cpu/event=0xA2,umask=0x10,name=RESOURCE_STALLS.ROB/u,cpu/event=0x0E,umask=0x1,cmask=1,inv=1,name=UOPS_ISSUED.ANY/u,cpu/event=0xA2,umask=0x4,name=RESOURCE_STALLS.RS/u ./main-1-lfence
Performance counter stats for './main-1-lfence' (10 runs):
1,32,55,451 cycles:u ( +- 0.01% )
50,00,007 instructions:u # 0.38 insns per cycle ( +- 0.00% )
0 RESOURCE_STALLS.ROB
1,03,84,640 UOPS_ISSUED.ANY ( +- 0.04% )
0 RESOURCE_STALLS.RS
0.004163500 seconds time elapsed ( +- 0.41% )
Observe that the number of cycles has increased by about 10 million, or 10 cycles per iteration. The number of cycles does not tell us much. The number of retired instruction has increased by a million, which is expected. We already know that the lfence will not make instruction complete any faster, so RESOURCE_STALLS.ROB should not change. UOPS_ISSUED.ANY and RESOURCE_STALLS.RS are particularly interesting. In this output, UOPS_ISSUED.ANY counts cycles, not uops. The number of uops can also be counted (using cpu/event=0x0E,umask=0x1,name=UOPS_ISSUED.ANY/u instead of cpu/event=0x0E,umask=0x1,cmask=1,inv=1,name=UOPS_ISSUED.ANY/u) and has increased by 6 uops per iteration (no fusion). This means that an lfence that was placed between two imuls was decoded into 6 uops. The one million dollar question is now what these uops do and how they move around in the pipe.
RESOURCE_STALLS.RS is zero. What does that mean? This indicates that the allocator, when it sees an lfence in the IDQ, it stops allocating until all current uops in the ROB retire. In other words, the allocator will not allocate entries in the RS past an lfence until the lfence retires. Since the loop body contains only 3 other uops, the 60-entry RS will never be full. In fact, it will be always almost empty.
The IDQ in reality is not a single simple queue. It consists of multiple hardware structures that can operate in parallel. The number of uops an lfence requires depends on the exact design of the IDQ. The allocator, which also consists of many different hardware structures, when it see there is an lfence uops at the front of any of the structures of the IDQ, it suspends allocation from that structure until the ROB is empty. So different uops are usd with different hardware structures.
UOPS_ISSUED.ANY shows that the allocator is not issuing any uops for about 9-10 cycles per iteration. What is happening here? Well, one of the uses of lfence is that it can tell us how much time it takes to retire an instruction and allocate the next instruction. The following assembly code can be used to do that:
TIMES T lfence
The performance event counters will not work well for small values of T. For sufficiently large T, and by measuring UOPS_ISSUED.ANY, we can determine that it takes about 4 cycles to retire each lfence. That's because UOPS_ISSUED.ANY will be incremented about 4 times every 5 cycles. So after every 4 cycles, the allocator issues another lfence (it doesn't stall), then it waits for another 4 cycles, and so on. That said, instructions that produce results may require 1 or few more cycle to retire depending on the instruction. IACA always assume that it takes 5 cycles to retire an instruction.
Our loop looks like this:
imul eax, eax
lfence
imul edx, edx
dec ecx
jnz .loop
At any cycle at the lfence boundary, the ROB will contain the following instructions starting from the top of the ROB (the oldest instruction):
imul edx, edx - N
dec ecx/jnz .loop - N
imul eax, eax - N+1
Where N denotes the cycle number at which the corresponding instruction was dispatched. The last instruction that is going to complete (reach the writeback stage) is imul eax, eax. and this happens at cycle N+4. The allocator stall cycle count will be incremented during cycles, N+1, N+2, N+3, and N+4. However it will about 5 more cycles until imul eax, eax retires. In addition, after it retires, the allocator needs to clean up the lfence uops from the IDQ and allocate the next group of instructions before they can be dispatched in the next cycle. The perf output tells us that it takes about 13 cycles per iteration and that the allocator stalls (because of the lfence) for 10 out of these 13 cycles.
The graph from the question shows only the number of cycles for up to T=100. However, there is another (final) knee at this point. So it would be better to plot the cycles for up to T=120 to see the full pattern.