GCC does not optimize out functions without side effects with asm statements inside loop even if return velue is ignored

[Saldyrkine, Mikhail]

g++ (GCC) 6.3.1 20170216 (Red Hat 6.3.1-3)

In the below case compile_test_asm_inside_loop invokes test_asm_inside_loop and 
ignores results.
The call into test_asm_inside_loop is expected to be eliminated since return 
value is not used and there is no side effect
The call elimination works fine without asm and without loop
It does not work with asm inside loop

TEST CODE

#include <iostream>
#include <assert.h>

using namespace std;
constexpr static size_t capacity = 1024;
uint64_t objects[capacity];

// THE FUNCTION IS ELIMINATED BY COMPILER IF OUTPUT IS NOT USED
inline uint64_t test_noloop(uint64_t idx) {
    uint64_t result;
    asm ("movq (%1,%2,8), %0" : "=r" (result) : "r" (objects), "r" (idx) );
    if( result > 128 )
        return result;
    return 0;
}

// THE FUNCTION IS ELIMINATED BY COMPILER IF OUTPUT IS NOT USED
inline uint64_t test_noasm(uint64_t idx) {
    while(true)
    {
        if( objects[idx] > 128 )
            return objects[idx];
        ++idx;
    }
}

// THE FUNCTION IS KEEPT EVEN WHEN IF RESULT IS NOT USED - ASM INSIDE LOOP 
CAUSING THE ISSUE
inline uint64_t test_asm_inside_loop(uint64_t idx) {
    while(true)
    {
        uint64_t result;
        asm ("movq (%1,%2,8), %0" : "=r" (result) : "r" (objects), "r" (idx) );
        if( result > 128 )
            return result;
        ++idx;
    }
}

void init() {
    srand(time(nullptr));
    for( size_t i = 0; i < capacity - 1; ++i )
        objects[i] = random() % 256;
    objects[capacity-1] = 255;
}

// TETS THAT test_noasm AND test_asm_inside_loop PRODUCE SAME RESULT
void sanity_test() {
    for( size_t i = 0; i < capacity; ++i ) {
        assert( test_noasm(i) == test_asm_inside_loop(i));
    }
}

void compile_test_noasm() {
    test_noasm(0);
}

void compile_test_noloop() {
    test_noloop(0);
}

void compile_test_asm_inside_loop() {
    test_asm_inside_loop(0);
}

int main( int argc, char* argv[] ) {
    init();
    sanity_test();
    compile_test_noasm();
    compile_test_noloop();
    compile_test_asm_inside_loop();
}

COMPILATION AND DISASSEMBLER RESULTS:

/opt/rh/devtoolset-6//root/bin/g++  -O3 -funroll-loops  loop_optimization.cpp; 
gdb -batch -ex "file a.out" -ex "disas compile_test_noasm" -ex "disas 
compile_test_noloop" -ex "disas compile_test_asm_inside_loop"
Dump of assembler code for function _Z18compile_test_noasmv:
   0x0000000000400970 <+0>:     repz retq
End of assembler dump.
Dump of assembler code for function _Z19compile_test_noloopv:
   0x0000000000400980 <+0>:     repz retq
End of assembler dump.
Dump of assembler code for function _Z28compile_test_asm_inside_loopv:
   0x0000000000400990 <+0>:     xor    %edx,%edx
   0x0000000000400992 <+2>:     mov    $0x601080,%ecx
   0x0000000000400997 <+7>:     xor    %eax,%eax
   0x0000000000400999 <+9>:     mov    (%rcx,%rdx,8),%rsi
   0x000000000040099d <+13>:    cmp    $0x80,%rsi
   0x00000000004009a4 <+20>:    ja     0x4009c1 
<_Z28compile_test_asm_inside_loopv+49>
   0x00000000004009a6 <+22>:    nopw   %cs:0x0(%rax,%rax,1)
   0x00000000004009b0 <+32>:    add    $0x1,%rax
   0x00000000004009b4 <+36>:    mov    (%rcx,%rax,8),%rdi
   0x00000000004009b8 <+40>:    cmp    $0x80,%rdi
   0x00000000004009bf <+47>:    jbe    0x4009b0 
<_Z28compile_test_asm_inside_loopv+32>
   0x00000000004009c1 <+49>:    repz retq
End of assembler dump.