GNU inline assembly optimisation

gcc, icc, inline-assembly, optimization, x86-64

Solution

I've tried your example on Linux like this (making it "evil" by forcing a stack ref/loc for `test` through using `&test` in the `printf`:):

#include <stdio.h>
int main(int argc, char **argv)
{
    unsigned long test = 0;
    unsigned long bsr;
// bit test and set 39th bit
    asm ("btsq\t%1, %0 " : "+rm" (test) : "rJ" (39) );
// bit scan reverse (get most significant bit id)
    asm ("bsrq\t%1, %0" : "=r" (bsr) : "rm" (test) );
    printf("test = %lu, bsr = %d, &test = %p\n", test, bsr, &test);
    return 0;
}

and compiled it with various versions of `gcc -O3` ... to the following results:

code generated                                                     gcc version
================================================================================
  400630:       48 83 ec 18             sub    $0x18,%rsp          4.7.2,
  400634:       31 c0                   xor    %eax,%eax           4.6.2,
  400636:       bf 50 07 40 00          mov    $0x400750,%edi      4.4.6
  40063b:       48 8d 4c 24 08          lea    0x8(%rsp),%rcx
  400640:       48 0f ba e8 27          bts    $0x27,%rax
  400645:       48 89 44 24 08          mov    %rax,0x8(%rsp)
  40064a:       48 89 c6                mov    %rax,%rsi
  40064d:       48 0f bd d0             bsr    %rax,%rdx
  400651:       31 c0                   xor    %eax,%eax
  400653:       e8 68 fe ff ff          callq  4004c0 
[ ... ]
---------------------------------------------------------------------------------
  4004f0:       48 83 ec 18             sub    $0x18,%rsp          4.1
  4004f4:       31 c0                   xor    %eax,%eax
  4004f6:       bf 28 06 40 00          mov    $0x400628,%edi
  4004fb:       48 8d 4c 24 10          lea    0x10(%rsp),%rcx
  400500:       48 c7 44 24 10 00 00 00 00      movq   $0x0,0x10(%rsp)
  400509:       48 0f ba e8 27          bts    $0x27,%rax
  40050e:       48 89 44 24 10          mov    %rax,0x10(%rsp)
  400513:       48 89 c6                mov    %rax,%rsi
  400516:       48 0f bd d0             bsr    %rax,%rdx
  40051a:       31 c0                   xor    %eax,%eax
  40051c:       e8 c7 fe ff ff          callq  4003e8 
[ ... ]
---------------------------------------------------------------------------------
  400500:       48 83 ec 08             sub    $0x8,%rsp           3.4.5
  400504:       bf 30 06 40 00          mov    $0x400630,%edi
  400509:       31 c0                   xor    %eax,%eax
  40050b:       48 c7 04 24 00 00 00 00         movq   $0x0,(%rsp)
  400513:       48 89 e1                mov    %rsp,%rcx
  400516:       48 0f ba 2c 24 27       btsq   $0x27,(%rsp)
  40051c:       48 8b 34 24             mov    (%rsp),%rsi
  400520:       48 0f bd 14 24          bsr    (%rsp),%rdx
  400525:       e8 fe fe ff ff          callq  400428 
[ ... ]
---------------------------------------------------------------------------------
  4004e0:       48 83 ec 08             sub    $0x8,%rsp           3.2.3
  4004e4:       bf 10 06 40 00          mov    $0x400610,%edi
  4004e9:       31 c0                   xor    %eax,%eax
  4004eb:       48 c7 04 24 00 00 00 00         movq   $0x0,(%rsp)
  4004f3:       48 0f ba 2c 24 27       btsq   $0x27,(%rsp)
  4004f9:       48 8b 34 24             mov    (%rsp),%rsi
  4004fd:       48 89 e1                mov    %rsp,%rcx
  400500:       48 0f bd 14 24          bsr    (%rsp),%rdx
  400505:       e8 ee fe ff ff          callq  4003f8 
[ ... ]

and while there's a significant difference in the created code (including whether the `bsr` acceesses `test` as register or memory), none of the tested revs recreate the assembly that you've shown. I'd suspect a bug in the 4.2.x version you used on MacOSX, but then I don't have either your testcase nor that specific compiler version available.

Edit: The code above is obviously different in the sense that it forces `test` into the stack; if that is not done, then all "plain" gcc versions I've tested do a direct pair `bts $39, %rsi` / `bsr %rsi, %rdx`.

I have found, though, that `clang` creates different code there:

 140:   50                      push   %rax
 141:   48 c7 04 24 00 00 00 00         movq   $0x0,(%rsp)
 149:   31 f6                   xor    %esi,%esi
 14b:   48 0f ba ee 27          bts    $0x27,%rsi
 150:   48 89 34 24             mov    %rsi,(%rsp)
 154:   48 0f bd d6             bsr    %rsi,%rdx
 158:   bf 00 00 00 00          mov    $0x0,%edi
 15d:   30 c0                   xor    %al,%al
 15f:   e8 00 00 00 00          callq  printf@plt>

so the difference seems to be indeed between the code generators of clang/llvm and "gcc proper".

Problem

I am trying to write a small library for highly optimised x86-64 bit operation code and am fiddling with inline asm. While testing this particular case has caught my attention: ``` unsigned long test = 0; unsigned long bsr; // bit test and set 39th bit __asm__ ("btsq\t%1, %0 " : "+rm" (test) : "rJ" (39) ); // bit scan reverse (get most significant bit id) __asm__ ("bsrq\t%1, %0" : "=r" (bsr) : "rm" (test) ); printf("test = %lu, bsr = %d\n", test, bsr); ``` compiles and runs fine in both gcc and icc, but when I inspect the assembly I get differences `gcc -S -fverbose-asm -std=gnu99 -O3` ``` movq $0, -8(%rbp) ## InlineAsm Start btsq $39, -8(%rbp) ## InlineAsm End movq -8(%rbp), %rax movq %rax, -16(%rbp) ## InlineAsm Start bsrq -16(%rbp), %rdx ## InlineAsm End movq -8(%rbp), %rsi leaq L_.str(%rip), %rdi xorb %al, %al callq _printf ``` I am wondering why so complicated? I am writing high performance code in which the number of instructions is critical. I am especially wondering why gcc makes a copy of my variable `test` before passing it to the second inline asm? Same code compiled with icc gives far better results: ``` xorl %esi, %esi # test = 0 movl $.L_2__STRING.0, %edi # has something to do with printf orl $32832, (%rsp) # part of function initiation xorl %eax, %eax # has something to do with printf ldmxcsr (%rsp) # part of function initiation btsq $39, %rsi #106.0 bsrq %rsi, %rdx #109.0 call printf #111.2 ``` despite the fact that gcc decides to keep my variables on the stack rather then in registers, what I do not understand is why make a copy of `test` before passing it to the second asm? If I put `test` in as an input/output variable in the second asm `__asm__ ("bsrq\t%1, %0" : "=r" (bsr) , "+rm" (test) );` then those lines disappear. ``` movq $0, -8(%rbp) ## InlineAsm Start btsq $39, -8(%rbp) ## InlineAsm End ## InlineAsm Start bsrq -8(%rbp), %rdx ## InlineAsm End movq -8(%rbp), %rsi leaq L_.str(%rip), %rdi xorb %al, %al callq _printf ``` Is this gcc screwed up optimisation or am I missing some vital compiler switches? I do have icc for my production system, but if I decide to distribute the source code at some point then it will have to be able to compile with gcc too. compilers used: gcc version 4.2.1 (Based on Apple Inc. build 5658) (LLVM build 2336.1.00) icc Version 12.0.2

Original source