Why does adding inline assembly comments cause such radical change in GCC's generated code?

Question

So, I had this code:

constexpr unsigned N = 1000;
void f1(char* sum, char* a, char* b) {
    for(int i = 0; i < N; ++i) {
        sum[i] = a[i] + b[i];
    }
}

void f2(char* sum, char* a, char* b) {
    char* end = sum + N;
    while(sum != end) {
        *sum++ = *a++ + *b++;
    }
}

I wanted to see the code that GCC 4.7.2 would generate. So I ran g++ -march=native -O3 -masm=intel -S a.c++ -std=c++11 And got the following output:

        .file   "a.c++"
        .intel_syntax noprefix
        .text
        .p2align 4,,15
        .globl  _Z2f1PcS_S_
        .type   _Z2f1PcS_S_, @function
_Z2f1PcS_S_:
.LFB0:
        .cfi_startproc
        lea     rcx, [rdx+16]
        lea     rax, [rdi+16]
        cmp     rdi, rcx
        setae   r8b
        cmp     rdx, rax
        setae   cl
        or      cl, r8b
        je      .L5
        lea     rcx, [rsi+16]
        cmp     rdi, rcx
        setae   cl
        cmp     rsi, rax
        setae   al
        or      cl, al
        je      .L5
        xor     eax, eax
        .p2align 4,,10
        .p2align 3
.L3:
        movdqu  xmm0, XMMWORD PTR [rdx+rax]
        movdqu  xmm1, XMMWORD PTR [rsi+rax]
        paddb   xmm0, xmm1
        movdqu  XMMWORD PTR [rdi+rax], xmm0
        add     rax, 16
        cmp     rax, 992
        jne     .L3
        mov     ax, 8
        mov     r9d, 992
.L2:
        sub     eax, 1
        lea     rcx, [rdx+r9]
        add     rdi, r9
        lea     r8, [rax+1]
        add     rsi, r9
        xor     eax, eax
        .p2align 4,,10
        .p2align 3
.L4:
        movzx   edx, BYTE PTR [rcx+rax]
        add     dl, BYTE PTR [rsi+rax]
        mov     BYTE PTR [rdi+rax], dl
        add     rax, 1
        cmp     rax, r8
        jne     .L4
        rep
        ret
.L5:
        mov     eax, 1000
        xor     r9d, r9d
        jmp     .L2
        .cfi_endproc
.LFE0:
        .size   _Z2f1PcS_S_, .-_Z2f1PcS_S_
        .p2align 4,,15
        .globl  _Z2f2PcS_S_
        .type   _Z2f2PcS_S_, @function
_Z2f2PcS_S_:
.LFB1:
        .cfi_startproc
        lea     rcx, [rdx+16]
        lea     rax, [rdi+16]
        cmp     rdi, rcx
        setae   r8b
        cmp     rdx, rax
        setae   cl
        or      cl, r8b
        je      .L19
        lea     rcx, [rsi+16]
        cmp     rdi, rcx
        setae   cl
        cmp     rsi, rax
        setae   al
        or      cl, al
        je      .L19
        xor     eax, eax
        .p2align 4,,10
        .p2align 3
.L17:
        movdqu  xmm0, XMMWORD PTR [rdx+rax]
        movdqu  xmm1, XMMWORD PTR [rsi+rax]
        paddb   xmm0, xmm1
        movdqu  XMMWORD PTR [rdi+rax], xmm0
        add     rax, 16
        cmp     rax, 992
        jne     .L17
        add     rdi, 992
        add     rsi, 992
        add     rdx, 992
        mov     r8d, 8
.L16:
        xor     eax, eax
        .p2align 4,,10
        .p2align 3
.L18:
        movzx   ecx, BYTE PTR [rdx+rax]
        add     cl, BYTE PTR [rsi+rax]
        mov     BYTE PTR [rdi+rax], cl
        add     rax, 1
        cmp     rax, r8
        jne     .L18
        rep
        ret
.L19:
        mov     r8d, 1000
        jmp     .L16
        .cfi_endproc
.LFE1:
        .size   _Z2f2PcS_S_, .-_Z2f2PcS_S_
        .ident  "GCC: (GNU) 4.7.2"
        .section        .note.GNU-stack,"",@progbits

I suck at reading assembly, so I decided to add some markers to know where the bodies of the loops went:

constexpr unsigned N = 1000;
void f1(char* sum, char* a, char* b) {
    for(int i = 0; i < N; ++i) {
        asm("# im in ur loop");
        sum[i] = a[i] + b[i];
    }
}

void f2(char* sum, char* a, char* b) {
    char* end = sum + N;
    while(sum != end) {
        asm("# im in ur loop");
        *sum++ = *a++ + *b++;
    }
}

And GCC spat this out:

    .file   "a.c++"
    .intel_syntax noprefix
    .text
    .p2align 4,,15
    .globl  _Z2f1PcS_S_
    .type   _Z2f1PcS_S_, @function
_Z2f1PcS_S_:
.LFB0:
    .cfi_startproc
    xor eax, eax
    .p2align 4,,10
    .p2align 3
.L2:
#APP
# 4 "a.c++" 1
    # im in ur loop
# 0 "" 2
#NO_APP
    movzx   ecx, BYTE PTR [rdx+rax]
    add cl, BYTE PTR [rsi+rax]
    mov BYTE PTR [rdi+rax], cl
    add rax, 1
    cmp rax, 1000
    jne .L2
    rep
    ret
    .cfi_endproc
.LFE0:
    .size   _Z2f1PcS_S_, .-_Z2f1PcS_S_
    .p2align 4,,15
    .globl  _Z2f2PcS_S_
    .type   _Z2f2PcS_S_, @function
_Z2f2PcS_S_:
.LFB1:
    .cfi_startproc
    xor eax, eax
    .p2align 4,,10
    .p2align 3
.L6:
#APP
# 12 "a.c++" 1
    # im in ur loop
# 0 "" 2
#NO_APP
    movzx   ecx, BYTE PTR [rdx+rax]
    add cl, BYTE PTR [rsi+rax]
    mov BYTE PTR [rdi+rax], cl
    add rax, 1
    cmp rax, 1000
    jne .L6
    rep
    ret
    .cfi_endproc
.LFE1:
    .size   _Z2f2PcS_S_, .-_Z2f2PcS_S_
    .ident  "GCC: (GNU) 4.7.2"
    .section    .note.GNU-stack,"",@progbits

This is considerably shorter, and has some significant differences like the lack of SIMD instructions. I was expecting the same output, with some comments somewhere in the middle of it. Am I making some wrong assumption here? Is GCC's optimizer hindered by asm comments?

Answer 1

This answer is now modified: it was originally written with a mindset considering inline Basic Asm as a pretty strongly specified tool, but it's nothing like that in GCC. Basic Asm is weak and so the answer was edited.

Each assembly comment acts as a breakpoint.

EDIT: But a broken one, as you use Basic Asm. Inline asm (an asm statement inside a function body) without explicit clobber list is a weakly specified feature in GCC and its behavior is hard to define. It doesn't seem (I don't fully grasp its guarantees) attached to anything in particular, so while the assembly code must be run at some point if the function is run, it isn't clear when it is run for any non trivial optimization level. A breakpoint that can be reordered with neighboring instruction isn't a very useful "breakpoint". END EDIT

You could run your program in an interpreter that breaks at each comment and prints out the state of every variable (using debug information). These points must exist so that you observe the environment (state of registers and memory).

Without the comment, no observation point exists, and the loop is compiled as a single mathematical function taking an environment and producing a modified environment.

You want to know the answer of a meaningless question: you want to know how each instruction (or maybe block, or maybe range of instruction) is compiled, but no single isolated instruction (or block) is compiled; the whole stuff is compiled as a whole.

A better question would be:

Hello GCC. Why do you believe this asm output is implementing the source code? Please explain step by step, with every assumption.

But then you wouldn't want to read a proof longer than the asm output, written in term of GCC internal representation.

Answer 2

I don't agree with the "gcc doesn't understand what is in the asm() block". For example, gcc can deal quite well with optimising parameters, and even re-arranging asm() blocks such that it intermingles with the generated C code. This is why, if you look at inline assembler in for example the Linux kernel, it is nearly always prefixed with __volatile__ to ensure that the compiler "doesn't move the code around". I have had gcc move my "rdtsc" around, which made my measurements of the time it took to do certain thing.

As documented, gcc treats certain types of asm() blocks as "special", and thus doesn't optimise the code either side of the block.

That's not to say that gcc won't, sometimes, get confused by inline assembler blocks, or simply decide to give up on some particular optimisation because it can't follow the consequences of the assembler code, etc, etc. More importantly, it can often get confused by missing clobber tags - so if you have some instruction like cpuid that changes the value of EAX-EDX, it but you wrote the code so that it only uses EAX, the compiler may store things in EBX, ECX and EDX, and then your code acts very strange when these registers are overwritten... If you are lucky, it crashes immediately - then it's easy to figure out what goes on. But if you are unlucky, it crashes way down the line... Another tricky one is the divide instruction that give a second result in edx. If you don't care about the modulo, it's easy to forget that EDX was changed.

Answer 3

The interactions with optimisations are explained about halfway down the "Assembler Instructions with C Expression Operands" page in the documentation.

GCC doesn't try to understand any of the actual assembly inside the asm; the only thing it knows about the content is what you (optionally) tell it in the output and input operand specification and the register clobber list.

In particular, note:

An asm instruction without any output operands will be treated identically to a volatile asm instruction.

and

The volatile keyword indicates that the instruction has important side-effects [...]

So the presence of the asm inside your loop has inhibited a vectorisation optimisation, because GCC assumes it has side effects.

Answer 4

Note that gcc vectorized the code, splitting the loop body into two parts, the first processing 16 items at a time, and the second doing the remainder later.

As Ira commented, the compiler doesn't parse the asm block, so it does not know that it's just a comment. Even if it did, it has no way of knowing what you intended. The optmized loops have the body doubled, should it put your asm in each? Would you like it that it isn't executed 1000 times? It doesn't know, so it goes the safe route and falls back to the simple single loop.