You can't fool the optimizer

    julia> function f(n)
             total = 0
             for x in 1:n
               total += Base.donotdelete(x)
             end
             return total
           end

I remember reading (although I can't find it now) a great analysis of all the optimizations that Javascript compilers _can't_ do because of the existence of the "eval" instruction.

For that you could at least argue that if the libc's strlen is faster than strcmp, that improves performance if the programmer expects the function to be usually called with a short input.

That said, changing it to `if (strlen(s) == 5) return 0;` it still doesn't get optimized (https://godbolt.org/z/7feWWjhfo), even though the entire function is completely equivalent to just `return 0;`.

  bool is_divisible_by_6(int x) {
      return x % 2 == 0 && x % 3 == 0;
  }

  bool is_divisible_by_6_optimal(int x) {
      return x % 6 == 0;
  }

   int foo(char const *s) {
     if (s[0] == 'h' && s[1] == 'e' && s[2] == 'l' && s[3] == 'l')
        return 1;
     return 0;
   }

I feel like it is unfair to blame the compiler when you've explicitly asked for `/O1`. If you change this to `/O2` or `/Ox` then MSVC will optimize this into a constant 5, proving that it does "know" that strlen will return 5 in this case.

Nit: some might think the reason this doesn't work is because the shift would "move" the sign bit, but actually arithmetic shifting instructions exist for this exact purpose. The reason they are not enough is because shifting provides the wrong kind of division rounding for negative numbers. This can however be fixed up by adding 1 if the number is negative (this can be done with an additional logical shift for moving the sign bit to the rightmost position and an addition).

You've misrepresented the situation. Turn up the optimiser to `/O2` and MSVC returns 5 directly, too.

> This function returns 1 if s is "hello". 0 otherwise. I've added a pointless strlen(). It seems like no compiler is clever enough to remove it.

It's funny how sometimes operating at a higher level of abstraction allows the compiler to optimise the code better: https://godbolt.org/z/EYP5764Mv

In this, the string literal "hello" is lowered not merely into a static string, but a handful of integral immediates that are directly inline in the assembly, no label-dereferencing required, and the 'is equal to "hello"' test is cast as the result of some sign extends and a bitwise-xor.

Of course, one could argue that std::string_view::size() is statically available, but then my counter-argument is that C's zero-terminated strings are a massive pessimisation (which is why the compiler couldn't 'see' what we humans can), and should always be avoided.

Like with people in general, it depends on what compiler/interpreter we're talking about, I'll freely grant that clang is smarter than me, but CPython for sure isn't. :)

More generally, canonicalization goes very far, but no farther than language semantics allows. Not even the notorious "sufficiently smart compiler" with infinite time can figure out what you don't tell it.

pareto principle like always, dont need the best but good enough

not every company is google level anyway

I agree that most people are not writing hand-tuned avr8 assembly. Most people aren't attempting to do DSP on 8-bit AVRs either.

You can mostly not think about super low level integer manipulation stuff though.

I work with Cuda kernels a lot for computer vision. I am able to consistently and significantly improve on the performance of research code without any fancy tricks, just with good software engineering practices.

By organising variables into structs, improving naming, using helper functions, etc... the previously impenetrable code becomes so much clearer and the obvious optimisations reveal themselves.

Not to say there aren't certain tricks / patterns / gotchas / low level hardware realities to keep in mind, of course.

...I don't know... for instance the MSVC compiler creates this output for the last two 'non-trivial' functions with '/Ox':

  add w8,w1,w0
  cmp w0,#0
  cseleq w0,w1,w8

A better mindset is "don't trust the compiler for code that's actually performance sensitive".

You shouldn't validate each line of compiler output, but at least for the 'hot areas' in the code base that definitely pays off, because sometimes compilers do really weird shit for no good reason (often because of 'interference' between unrelated optimizer passes) - and often you don't need to dig deep to stumble over weird output like in the example above.

Though when your data is behind pointers, it's very easy to write code that the compiler can no longer figure out how to optimize.

This is true, but it also means "the compiler IS made for someone median smart, that now knows the machine".

It works great for basic, simple, common code, and for code that is made with care for data structures.

A total mess of code is another story.

P.D: is similar to the query optimizers, that neither can outrun a terrible made schema and queries

The optimizations that tend to have the most impact involve changes to the algorithm or data layout. Most compilers won’t do things like add a hash table to make a lookup O(1) or rearrange an array of structures to be a structure of arrays for better data locality. Coding with an eye for these optimizations is still a very good use of your time.

That mindset will last until the first or second time you walk through the compiler's assembly-language output. GCC and LLVM can find some amazing optimizations, but they also consistently miss very obvious optimization opportunities, often even on utterly trivial code.

That isn't to say that it's worthwhile to hand-optimize all your code. But the compiler is very much not smarter than you, and if you do need to squeeze performance out of the processor, you can do a lot better than the compiler does. A good first step is to look at the stupid shit the compiler has spat out and massage your source code until the shit stinks less.

There are limits on provable equivalence in the first place. things like equality saturation also try and do a better job of formalizing equivalence based rewrites.

    return n ? (1u  + popcount(n & n - 1u)) : 0u;

    unsigned add(unsigned x, unsigned y) {
        unsigned a, b;
        do {
            a = x & y;   /* every position where addition will generate a carry */
            b = x ^ y;   /* the addition, with no carries */
            x = a << 1;  /* the carries */
            y = b;
        /* if there were any carries, repeat the loop */
        } while (a);
        return b;
    }

It's a little more difficult to see that the loop will necessarily terminate.

New a values come from a bitwise & of x and y. New x values come from a left shift of a. This means that, if x ends in some number of zeroes, the next value of a will also end in at least that many zeroes, and the next value of x will end in an additional zero (because of the left shift). Eventually a will end in as many zeroes as there are bits in a, and the loop will terminate.

This is known as "code sinking," and most optimizers are capable of doing this. Except keep in mind that a) the profitability of doing so is not always clear [1] and b) the compiler is a lot more fastidious about corner-case behavior than you are, so it might conclude that it's not in fact safe to sink the operation when you think it is safe to do so.

[1] If the operation to sink is x = y + z, you now may need to keep the values of y and z around longer to compute the addition, increasing register pressure and potentially hurting performance as a result.

Did you try using bitwise and (&), or a local for the struct? The short-circuiting behaviour of the logical means that if `s->a != a`, `s->b` must not be dereferenced, so the compiler cannot turn this into a SIMD mask operation, because it behaves differently.

Generally compilers are pretty smart these days, and I find that more often than not if they miss an "obvious" optimization it's because there's a cornercase where it behaves differently from the code I wrote.

On higher optimization levels, many passes occur multiple times. However, as far as I know, compilers don't repeatedly run passes until they've reached an optimum. Instead, they run a fixed series of passes. I don't know why, maybe someone can chime in.

https://www.open-std.org/jtc1/sc22/wg14/www/docs/n1256.pdf

6.5.13, semantics

If you mean fooling the compiler by the source code obfuscation, it won't – by the time the first optimisation pass arrives, the source had already been transformed into an abstract syntax tree and the source code obfuscation becomes irrelevant.

Multiple optimiser passes do take place, but they are bounded in time – it is not an accepted expectation that the optimiser will spend a – theoretically – indefinite amount of time trying to arrive at the most perfect instruction sequence.

There was a GNU project a long time ago, «superoptimiser», which, given a sequence of instructions, would spend a very long time trying to optimise it into oblivion. The project was more of an academic exercise, and it has been long abandoned since.

I expect because the former helps more in optimising real-world code than the latter. It’s not worth the LLVM developer's time to make the compiler better for programs that it won’t see in practice.

It’s not as if the compiler did nothing with that code, though. It replaced the multiplication by a left shift and removed the branch.

The trick is that it’s doing both the add and the left shift in parallel then selecting which to use based on a compare of the two values with csel.

(To see this, rather than reading the code sequentially, think of every instruction as being issued at the same time until you hit an instruction that needs a destination register from an earlier instruction)

The add is stored in W9 but only read if the two arguments are unequal.

If the compare succeeds and the lsl retires before the add, the add is never read, so nothing stalls waiting for it and the answer can be returned while the add is still in flight. The result of the add would then be quietly discarded assuming it ever started (maybe there’s some magic where it doesn’t even happen at all?).

It’s not clear to me that this is power efficient, or that on many real cpus there’s a latency difference to exploit between add and lsl, so it may not be faster than just unconditionally doing the addition.

That said, it is definitely faster than the code as it was written which if translated to asm verbatim stalls on the compare before executing either the add or the left shift.

Basically the idea for most compilers is to do a series of transforms which incrementally improve the program (or at least make it worse in understood and reversible ways). To do this transform you need the optimizer to do the (not always trivial) proof that the 2*x is equivalent to x+y, do the replacement, do the gvn to duplicate the adds and finally do the branch elimination. Each of these steps is however totally separate from one another and the first one doesn't trigger since as far as it's concerned a shift left is faster than an add so why should it do the replacement.

This is all even more complicated since what representation is faster can depend on the target.

Compilers are essentially massive towers of heuristics for which patterns to apply for optimization. We don't throw a general SMT solver at your code because that takes way too long to compile; instead, we look at examples of actual code and make reasonable efforts to improve code.

In the case of the incrementing in a loop, there is a general analysis called Scalar Evolution that recasts expressions as an affine expression of canonical loop iteration variables (i.e., f(x), where x is 0 on the first loop iteration, 1 on the second, etc.). In the loop `while (x--) y++;`, the x variable [at the end of each loop iteration] can be rewritten as x = x₀ + -1*i, while the y variable is y = y₀ + 1*i. The loop trip count can be solved to an exact count, so we can replace the use of y outside the loop with y = y₀ + 1*trip count = y₀ + x, and then the loop itself is dead and can be deleted. These are all optimizations that happen to be quite useful in other contexts, so it's able to easily recognize this form of loop.

In the example you give, the compiler has to recognize the equivalence of two values conditional on control flow. The problem is that this problem really starts to run into the "the time needed to optimize this isn't worth the gain you get in the end." Note that there are a lot of cases where you have conditional joins (these are "phis" in SSA optimizer parlance), most of which aren't meaningfully simplifiable, so you're cutting off the analysis for all but the simplest cases. At a guess, the simplification is looking for all of the input values to be of the same form, but 2 * x (which will actually be canonicalized to x << 1) is not the same form as x + y, so it's not going to see if the condition being used to choose between the same values would be sufficient to make some operation return the same value. There are representations that make this problem much easier (egraphs), but these are not the dominant form for optimizers at present.

If you declare them as static, it eliminates the functions and the calls completely: https://aoco.compiler-explorer.com/z/soPqe7eYx

I'm sure it could also perform definition merging like you suggest but I can't think of a way of triggering it at the moment without also triggering their complete elision.

It can't do that because the program might load a dynamic library that depends on the function (it's perfectly OK for a `.so` to depend on a function from the main executable, for example).

That's one of the reasons why a very cheap optimization is to always use `static` for functions when you can. You're telling the compiler that the function doesn't need to be visible outside the current compilation unit, so the compiler is free to even inline it completely and never produce an actual callable function, if appropriate.

A quick google suggests it's called "identical comdat folding" https://devblogs.microsoft.com/oldnewthing/20161024-00/?p=94...

> Perform Identical Code Folding for functions (-fipa-icf-functions), read-only variables (-fipa-icf-variables), or both (-fipa-icf). The optimization reduces code size and may disturb unwind stacks by replacing a function by an equivalent one with a different name. The optimization works more effectively with link-time optimization enabled.

In addition, the Gold linker supports a similar feature via `--icf={safe,all}`:

> Identical Code Folding. '--icf=safe' Folds ctors, dtors and functions whose pointers are definitely not taken

    void go_forward(Closure *clo, Closure *cont, Closure *forward) {
        GC_CHECK(clo, cont, forward);
        ((Fun0)(forward->fun))(forward, cont);
    }

    void go_left(Closure *clo, Closure *cont, Closure *left, Closure *right) {
        GC_CHECK(clo, cont, left, right);
        ((Fun0)(left->fun))(left, cont);
    }

    void go_right(Closure *clo, Closure *cont, Closure *left, Closure *right) {
        GC_CHECK(clo, cont, left, right);
        ((Fun0)(right->fun))(right, cont);
    }

    GcInfo gc_info[] = {
        { .fun = (GenericFun)&go_forward, .envc = 0, .argc = 1 },
        { .fun = (GenericFun)&go_left, .envc = 0, .argc = 2 },
        { .fun = (GenericFun)&go_right, .envc = 0, .argc = 2 },
    };

† Monomorphization takes Parametrically Polymorphic functions, ie functions which are strongly typed but those types are parameters at compile time, and it emits distinct machine code for each needed variation of the function, so e.g. add(a, b) maybe gets compiled to produce add_integer(a, b) and add_float(a, b) and add_matrix(a, b) even though we only wrote one function, and then code which calls add(a, b) with matrices, is at compile time emitted as calling add_matrix(a, b), because the compiler knew it needs that version. In C++ the number of parameters is also potentially allowed to vary between callers so add_matrix(a, b, c, d) might exist too, this feature is not yet available in Rust.

If you look at GAS manual https://ftp.gnu.org/old-gnu/Manuals/gas-2.9.1/html_chapter/a... almost every other architecture has architecture specific syntax notes, in many cases for something as trivial comments. If they couldn't even decide on single symbols for comments, there is no hope for everything else.

ARM isn't the only architecture where GAS uses similar syntax as developers of corresponding CPU arch. They are not doing the same for X86 due to historical choices inherited from Unix software ecosystem and thus AT&T. If you play around on Godbolt with compilers for different architectures it seems like x86 and use AT&T syntax is the exception, there are a few other which use similar syntax but it's a minority.

Why not use same syntax for all architectures? I don't really know all the historical reasoning but I have a few guesses and each arch probably has it's own historic baggage. Being consistent with manufacturer docs and rest of ecosystem has the obvious benefits for the ones who need to read it. Assembly is architecture specific by definition so being consistent across different architectures has little value. GAS is consistent with GCC output. Did GCC added support for some architectures early with the with help of manufacturers assembler and only later in GAS? A lot of custom syntax quirks which don't easily fit into Intel/AT&T model and are related to various addressing modes used by different architectures. For example ARM has register postincrement/preincrement and the 0 cost shifts, arm doesn't have the subregister acess like x86 (RAX/EAX/AX/AH/AL) and non word access is more or less limited to load/store instructions unlike x86 where it can show up in more places. You would need to invent quite a few extensions for AT&T syntax for it to be used by all the non x86 architectures, or you could just use the syntax made by developer of architecture.