I'm starting to read LLVM docs and IR documentation.
In common architectures, an asm cmp instruction "result" value is -at least- 3 bits long, let's say the first bit is the SIGN flag, the second bit is the CARRY flag and the third bit is the ZERO flag.
Question 1)
Why the IR icmp instruction result value is only i1? (you can choose only one flag)
Why doesn't IR define, let's call it a icmp2 instruction returning an i3 having SIGN,CARRY and ZERO flags?
This i3 value can be acted upon with a switch instruction, or maybe a specific br2 instruction, like:
%result = cmp2 i32 %a, i32 %b
br2 i3 %result onzero label %EQUAL, onsign label %A_LT_B
#here %a GT %b
Question 2)
Does this make sense? Could this br2 instruction help create new optimizations? i.e. remove all jmps? it is necessary or the performance gains are negligible?
The reason I'm asking this -besides not being an expert in LLVM- is because in my first tests I was expecting some kind of optimization to be made by LLVM in order to avoid making the comparison twice and also avoid all branches by using asm conditional-move instructions.
My Tests:
I've compiled with clang-LLVM this:
#include <stdlib.h>
#include <inttypes.h>
typedef int32_t i32;
i32 compare (i32 a, i32 b){
// return (a - b) & 1;
if (a>b) return 1;
if (a<b) return -1;
return 0;
}
int main(int argc, char** args){
i32 n,i;
i32 a,b,avg;
srand(0); //fixed seed
for (i=0;i<500;i++){
for (n=0;n<1e6;n++){
a=rand();
b=rand();
avg+=compare(a,b);
}
}
return avg;
}
Output asm is: ...
mov r15d, -1
...
.LBB1_2: # Parent Loop BB1_1 Depth=1
# => This Inner Loop Header: Depth=2
call rand
mov r12d, eax
call rand
mov ecx, 1
cmp r12d, eax
jg .LBB1_4
# BB#3: # in Loop: Header=BB1_2 Depth=2
mov ecx, 0
cmovl ecx, r15d
.LBB1_4: # %compare.exit
# in Loop: Header=BB1_2 Depth=2
add ebx, ecx
...
I expected (all jmps removed in the inner loop):
mov r15d, -1
mov r13d, 1 # HAND CODED
call rand
mov r12d, eax
call rand
xor ecx,ecx # HAND CODED
cmp r12d, eax
cmovl ecx, r15d # HAND CODED
cmovg ecx, r13d # HAND CODED
add ebx, ecx
Performance difference (1s) seems to be negligible (on a VM under VirtualBox):
- LLVM generated asm: 12.53s
- hancoded asm: 11.53s
- diff: 1s, in 500 millions iterations
Question 3)
Are my performance measures correct? Here's the makefile and the full hancoded.compare.s
makefile:
CC=clang -mllvm --x86-asm-syntax=intel
all:
$(CC) -S -O3 compare.c
$(CC) compare.s -o compare.test
$(CC) handcoded.compare.s -o handcoded.compare.test
echo `time ./compare.test`
echo `time ./handcoded.compare.test`
echo `time ./compare.test`
echo `time ./handcoded.compare.test`
hand coded (fixed) asm:
.text
.file "handcoded.compare.c"
.globl compare
.align 16, 0x90
.type compare,@function
compare: # @compare
.cfi_startproc
# BB#0:
mov eax, 1
cmp edi, esi
jg .LBB0_2
# BB#1:
xor ecx, ecx
cmp edi, esi
mov eax, -1
cmovge eax, ecx
.LBB0_2:
ret
.Ltmp0:
.size compare, .Ltmp0-compare
.cfi_endproc
.globl main
.align 16, 0x90
.type main,@function
main: # @main
.cfi_startproc
# BB#0:
push rbp
.Ltmp1:
.cfi_def_cfa_offset 16
push r15
.Ltmp2:
.cfi_def_cfa_offset 24
push r14
.Ltmp3:
.cfi_def_cfa_offset 32
push r12
.Ltmp4:
.cfi_def_cfa_offset 40
push rbx
.Ltmp5:
.cfi_def_cfa_offset 48
.Ltmp6:
.cfi_offset rbx, -48
.Ltmp7:
.cfi_offset r12, -40
.Ltmp8:
.cfi_offset r14, -32
.Ltmp9:
.cfi_offset r15, -24
.Ltmp10:
.cfi_offset rbp, -16
xor r14d, r14d
xor edi, edi
call srand
mov r15d, -1
mov r13d, 1 # HAND CODED
# implicit-def: EBX
.align 16, 0x90
.LBB1_1: # %.preheader
# =>This Loop Header: Depth=1
# Child Loop BB1_2 Depth 2
mov ebp, 1000000
.align 16, 0x90
.LBB1_2: # Parent Loop BB1_1 Depth=1
# => This Inner Loop Header: Depth=2
call rand
mov r12d, eax
call rand
xor ecx,ecx #hand coded
cmp r12d, eax
cmovl ecx, r15d #hand coded
cmovg ecx, r13d #hand coded
add ebx, ecx
.LBB1_3:
dec ebp
jne .LBB1_2
# BB#5: # in Loop: Header=BB1_1 Depth=1
inc r14d
cmp r14d, 500
jne .LBB1_1
# BB#6:
mov eax, ebx
pop rbx
pop r12
pop r14
pop r15
pop rbp
ret
.Ltmp11:
.size main, .Ltmp11-main
.cfi_endproc
.ident "Debian clang version 3.5.0-1~exp1 (trunk) (based on LLVM 3.5.0)"
.section ".note.GNU-stack","",@progbits