Floating operations. Is there loop-carried dependency? - optimization

Question

The following code presents QuickSort. I am trying to make it faster. Please help.
global sort
sort:
push ebp
mov ebp, esp
push edi
push esi
push ebx
sub esp, 28
mov eax, dword [ebp+8]
mov esi, dword [ebp+12]
mov dword [ebp-32], 1
lea edi, [eax-1]
sal eax, 2
lea edx, [eax+18]
movss xmm1, dword [esi-4+eax]
mov dword [ebp-28], edi
and edx, -16
sub esp, edx
lea edx, [esp+3]
shr edx, 2
lea ebx, [0+edx*4]
mov dword [4+edx*4], edi
mov dword [0+edx*4], 0
xor edi, edi
mov dword [ebp-36], ebx
.L14:
mov eax, dword [ebp-28]
lea edx, [edi-1]
cmp eax, edi
lea ebx, [esi+eax*4]
jle .L15
lea eax, [esi+edi*4]
.L18:
movss xmm0, dword [eax]
comiss xmm1, xmm0
jb .L16
add edx, 1
lea ecx, [esi+edx*4]
movss xmm2, dword [ecx]
movss dword [ecx], xmm0
movss dword [eax], xmm2
.L16:
add eax, 4
cmp eax, ebx
jne .L18
movss xmm1, dword [ebx]
.L15:
lea eax, [esi+4+edx*4]
cmp edx, edi
movss xmm0, dword [eax]
movss dword [eax], xmm1
movss dword [ebx], xmm0
jg .L19
sub dword [ebp-32], 2
lea edi, [edx+2]
cmp edi, dword [ebp-28]
jl .L25
.L21:
mov eax, dword [ebp-32]
test eax, eax
js .L11
mov edi, dword [ebp-36]
lea eax, [edi+eax*4]
mov edi, dword [eax]
mov ecx, edi
mov dword [ebp-28], edi
mov edi, dword [eax-4]
movss xmm1, dword [esi+ecx*4]
jmp .L14
.L19:
mov edi, dword [ebp-32]
mov eax, dword [ebp-36]
mov dword [eax+edi*4], edx
lea edi, [edx+2]
cmp edi, dword [ebp-28]
jge .L21
.L25:
mov ebx, dword [ebp-32]
mov ecx, dword [ebp-36]
movaps xmm1, xmm0
lea eax, [ecx+ebx*4]
mov ecx, dword [ebp-28]
mov dword [eax+4], edi
mov dword [eax+8], ecx
mov eax, ebx
add eax, 2
mov dword [ebp-32], eax
jns .L14
.L11:
lea esp, [ebp-12]
pop ebx
pop esi
pop edi
pop ebp
ret