; Comparison of register allocator between Cemu 1.11.0 and 1.11.1

; Old register allocator:
; Not global, looks at blocks of uninterrupted instructions individually (uninterrupted means no branches to or from the instructions)

; New register allocator:
; Global, looks at the whole function. Can carry registers across branches and loops

; PPC function

mflr      r0
stw       r0, 4(r1)
stwu      r1, -8(r1)
lis       r12, ((byte_1043EB90+0x10000)@h)
lbz       r0, byte_1043EB90@l(r12)
cmpwi     r0, 0
bne       loc_2FA73C4
bl        sub_2FA72D4

loc_2FA73C4:
lbz       r0, 0(r3)
cmpwi     r0, 0
li        r11, -1
beq       loc_2FA73FC
lis       r12, dword_10512FA8@h
addi      r12, r12, dword_10512FA8@l

loc_2FA73DC:
xor       r9, r11, r0
lbzu      r0, 1(r3)
clrlslwi  r10, r9, 24,2
srwi      r8, r11, 8
lwzx      r11, r10, r12
cmpwi     r0, 0
xor       r11, r8, r11
bne       loc_2FA73DC

loc_2FA73FC:
not       r3, r11
lwz       r0, 0xC(r1)
mtlr      r0
addi      r1, r1, 8
blr

____________________________________________________
; Code generated by Cemu 1.11.0 (partially annotated)

mov         eax,dword ptr [rsp+2BCh]            ; Move PPC LR into eax
mov         ebx,dword ptr [rsp+8]               ; Move PPC R1 into ebx
sub         dword ptr [rsp+2B4h],7              ; decrement ppc CTR
mov         edx,eax  
mov         r14,rdx  
movbe       dword ptr [r13+rbx+4],r14d  
mov         r14,rbx  
movbe       dword ptr [r13+rbx-8],r14d  
add         ebx,0FFFFFFF8h  
mov         ebp,10440000h  
movzx       edx,byte ptr [r13+rbp-1470h]  
cmp         edx,0  
sete        byte ptr [rsp+28Eh]  
mov         dword ptr [rsp+4],edx  
mov         dword ptr [rsp+8],ebx  
mov         dword ptr [rsp+34h],ebp  
jne         000000001DE4596E  

sub         dword ptr [rsp+2B4h],1  
mov         dword ptr [rsp+2BCh],1FA73C4h       ; update LR for function call
mov         edx,1FA72D4h  
jmp         qword ptr [r15+23F4E5A8h]           ; call function (BL)

loc_000000001DE4596E:
mov         esi,dword ptr [rsp+10h]  
sub         dword ptr [rsp+2B4h],4  
movzx       edx,byte ptr [r13+rsi]  
cmp         edx,0  
sete        byte ptr [rsp+28Eh]  
mov         edi,0FFFFFFFFh  
mov         dword ptr [rsp+4],edx  
mov         dword ptr [rsp+30h],edi  
je          after_loop
  
sub         dword ptr [rsp+2B4h],2  
mov         ebp,10510000h  
add         ebp,2FA8h  
mov         dword ptr [rsp+34h],ebp  

loc_loop:
mov         edx,dword ptr [rsp+4]               ; load ppc registers
mov         ebp,dword ptr [rsp+34h]  
mov         esi,dword ptr [rsp+10h]  
mov         edi,dword ptr [rsp+30h]  
sub         dword ptr [rsp+2B4h],8  
mov         r8d,edi                             ; xor       r9, r11, r0
xor         r8d,edx  
add         esi,1                               ; lbzu      r0, 1(r3)
movzx       edx,byte ptr [r13+rsi]  
mov         r9d,r8d                             ; clrlslwi  r10, r9, 24,2
rol         r9d,2
and         r9d,3FCh
mov         r10,rdi                             ; srwi      r8, r11, 8
shr         r10d,8  
add         r9d,ebp                             ; lwzx      r11, r10, r12
movbe       edi,dword ptr [r13+r9]  
sub         r9d,ebp  
xor         edi,r10d                            ; xor       r11, r8, r11 (reordered to avoid modifing rflags set by CMP)
cmp         edx,0                               ; cmpwi     r0, 0
sete        byte ptr [rsp+28Eh]                 ; update PPC CR0 equal-bit
mov         dword ptr [rsp+4],edx               ; store modified ppc registers
mov         dword ptr [rsp+10h],esi  
mov         dword ptr [rsp+30h],edi  
mov         dword ptr [rsp+28h],r8d  
mov         dword ptr [rsp+2Ch],r9d  
mov         dword ptr [rsp+24h],r10d  
jne         loc_loop                            ; bne

after_loop:
mov         ebx,dword ptr [rsp+8]  
mov         edi,dword ptr [rsp+30h]  
sub         dword ptr [rsp+2B4h],5  
mov         esi,edi  
not         esi  
movbe       edx,dword ptr [r13+rbx+0Ch]  
mov         eax,edx  
add         ebx,8  
mov         dword ptr [rsp+2BCh],eax  
mov         dword ptr [rsp+4],edx  
mov         dword ptr [rsp+8],ebx  
mov         dword ptr [rsp+10h],esi  
mov         edx,dword ptr [rsp+2BCh]  
jmp         qword ptr [r15+rdx*2+20000000h]     ; blr


____________________________________________________
; Code generated by Cemu 1.11.1

sub         dword ptr [rsp+2B4h],7  
mov         r10d,dword ptr [rsp+2BCh]  
mov         r9d,r10d  
mov         r8d,dword ptr [rsp+8]  
movbe       dword ptr [r13+r8+4],r9d  
mov         r14,r8  
movbe       dword ptr [r13+r8-8],r14d  
add         r8d,0FFFFFFF8h  
mov         edi,10440000h  
movzx       r9d,byte ptr [r13+rdi-1470h]  
cmp         r9d,0  
sete        byte ptr [rsp+28Eh]  
mov         dword ptr [rsp+4],r9d  
mov         dword ptr [rsp+34h],edi  
mov         dword ptr [rsp+8],r8d  
jne         000000001D7BFD74  

sub         dword ptr [rsp+2B4h],1  
mov         dword ptr [rsp+2BCh],1FA73C4h       ; update LR for function call
mov         edx,1FA72D4h  
jmp         qword ptr [r15+23F4E5A8h]           ; call function (BL)

mov         edi,dword ptr [rsp+34h]  
mov         r8d,dword ptr [rsp+8]  
mov         r9d,dword ptr [rsp+4]  
mov         r10d,dword ptr [rsp+2BCh]  

loc_000000001D7BFD74:
sub         dword ptr [rsp+2B4h],4  
mov         esi,dword ptr [rsp+10h]  
movzx       r9d,byte ptr [r13+rsi]  
cmp         r9d,0  
sete        byte ptr [rsp+28Eh]  
mov         ebp,0FFFFFFFFh  
mov         dword ptr [rsp+30h],ebp  
mov         dword ptr [rsp+4],r9d  
je          after_loop  

sub         dword ptr [rsp+2B4h],2  
mov         edi,10510000h                       
add         edi,2FA8h  
mov         eax,dword ptr [rsp+24h]             ; ppc register loads moved out of hot path
mov         edx,dword ptr [rsp+2Ch]  
mov         ebx,dword ptr [rsp+28h]  

loc_loop:
sub         dword ptr [rsp+2B4h],8  
mov         ebx,ebp                             ; xor       r9, r11, r0
xor         ebx,r9d  
add         esi,1                               ; lbzu      r0, 1(r3)
movzx       r9d,byte ptr [r13+rsi]  
mov         edx,ebx                             ; clrlslwi  r10, r9, 24,2
rol         edx,2  
and         edx,3FCh  
mov         rax,rbp                             ; srwi      r8, r11, 8
shr         eax,8
add         edx,edi                             ; lwzx      r11, r10, r12
movbe       ebp,dword ptr [r13+rdx]  
sub         edx,edi  
xor         ebp,eax                             ; xor       r11, r8, r11 (reordered)
cmp         r9d,0                               ; cmpwi     r0, 0
sete        byte ptr [rsp+28Eh]                 ; update PPC CR0 equal-bit
jne         loc_loop                            ; bne

mov         dword ptr [rsp+24h],eax             ; ppc register stores moved out of hot path
mov         dword ptr [rsp+2Ch],edx  
mov         dword ptr [rsp+28h],ebx  
mov         dword ptr [rsp+30h],ebp  
mov         dword ptr [rsp+4],r9d  
mov         dword ptr [rsp+34h],edi  

after_loop:
sub         dword ptr [rsp+2B4h],5  
mov         eax,dword ptr [rsp+30h]  
mov         esi,eax  
not         esi  
mov         dword ptr [rsp+10h],esi  
movbe       eax,dword ptr [r13+r8+0Ch]  
mov         r10d,eax  
mov         dword ptr [rsp+2BCh],r10d  
mov         dword ptr [rsp+4],eax  
add         r8d,8  
mov         dword ptr [rsp+8],r8d  
mov         edx,dword ptr [rsp+2BCh]  
jmp         qword ptr [r15+rdx*2+20000000h]     ; blr