tdm gcc 5.1 slower than 4.7

Question

Im using tdm gcc compilers to compile my winapi projects, when i test some simple mandelbrot sse code 9maybe with other projests its like the same but i not tested everything), 5.1 generates larger executable 330kB against 270kB (though i not recompiled everything, only hot loop module and linked with already compiled (in 4.7) ones) and also noticalby slower code 23.5 ms /frame agianst 20 ms in 4.7

It is scary.. What can i do with it? (the rest of settings etc is the same i only rename the compiler folder from one to anither

Im using c-like code but compile in c++ mode

Does maybe someone know how to resolve it? (by resolve i mean make 5.1 build working at least as fast as 4.7, also would prefer the executable be smaller)

//edit

ps i made yet quick test

as i may compile loop module in 51 and link all in 47

compile loop 47 link 47: size 270k speed 20 ms

compile loop 51 link 51: size 330k speed 23.5 ms

compile loop 47 link 51: size 330k speed 20 ms

compile loop 51 link 47: size 270k speed 23.5 ms

it shows that speed drop comes by compile in 51 and size bloat comes from link 51

Answer 1

I checked the assembly and it shows some changes in generated code, though slight

4.7

__Z16mandelbrot_n_sseU8__vectorfS_i: 
    pushl        %ebp 
    movl        %esp, %ebp 
    andl        $-16, %esp 
    subl        $16, %esp 
    movl        8(%ebp), %ecx 
    movaps        %xmm0, (%esp) 
    testl        %ecx, %ecx 
    js        L12 
    xorps        %xmm0, %xmm0 
    xorl        %eax, %eax 
    movaps        %xmm0, %xmm2 
    movaps        %xmm0, %xmm4 
    jmp        L11 
    .p2align 4,,7 
L19: 
    mulps        %xmm4, %xmm2 
    addl        $1, %eax 
    subps        %xmm5, %xmm6 
    movaps        (%esp), %xmm4 
    cmpl        %eax, %ecx 
    addps        %xmm6, %xmm4 
    addps        %xmm2, %xmm2 
    addps        %xmm1, %xmm2 
    jl        L10 
L11: 
    movaps        %xmm4, %xmm6 
    movaps        %xmm2, %xmm5 
    movaps        LC5, %xmm7 
    mulps        %xmm4, %xmm6 
    mulps        %xmm2, %xmm5 
    movaps        %xmm6, %xmm3 
    addps        %xmm5, %xmm3 
    cmpltps        LC4, %xmm3 
    andps        %xmm3, %xmm7 
    movmskps        %xmm3, %edx 
    testl        %edx, %edx 
    addps        %xmm7, %xmm0 
    jne        L19 
L10: 
    cvtps2dq        %xmm0, %xmm0 
    leave 
    ret 
L12: 
    xorps        %xmm0, %xmm0 
    jmp        L10 
    .globl        __Z16mandelbrot_n_sseDv4_fS_i 

5.1

__Z16mandelbrot_n_sseDv4_fS_i: 
    pushl        %ebp 
    movl        %esp, %ebp 
    andl        $-16, %esp 
    subl        $16, %esp 
    movl        8(%ebp), %ecx 
    movaps        %xmm0, (%esp) 
    testl        %ecx, %ecx 
    js        L11 
    pxor        %xmm0, %xmm0 
    xorl        %edx, %edx 
    movaps        %xmm0, %xmm5 
    movaps        %xmm0, %xmm2 
    jmp        L10 
    .p2align 4,,10 
L18: 
    mulps        %xmm2, %xmm5 
    addl        $1, %edx 
    subps        %xmm6, %xmm4 
    cmpl        %edx, %ecx 
    addps        %xmm5, %xmm5 
    addps        (%esp), %xmm4 
    addps        %xmm1, %xmm5 
    jl        L9 
    movaps        %xmm4, %xmm2 
L10: 
    movaps        %xmm2, %xmm4 
    movaps        %xmm5, %xmm6 
    movaps        LC7, %xmm7 
    mulps        %xmm2, %xmm4 
    mulps        %xmm5, %xmm6 
    movaps        %xmm4, %xmm3 
    addps        %xmm6, %xmm3 
    cmpltps        LC6, %xmm3 
    andps        %xmm3, %xmm7 
    movmskps        %xmm3, %eax 
    testl        %eax, %eax 
    addps        %xmm7, %xmm0 
    jne        L18 
L9: 
    cvtps2dq        %xmm0, %xmm0 
    leave 
    ret 
L11: 
    pxor        %xmm0, %xmm0 
    jmp        L9 
    .section        .text.unlikely,"x" 
LCOLDE8: 
    .text 

It seems that the 5.1 version is unlucky one and it cst 15% slowdown