[LLVMdev] Compiling integer mod

[Brent Walker]

For the simple C program below I show the output of clang and the
output of the VS compiler (I am on windows).  Maybe this is obvious to
you, but is it really faster to do 2 multiplications, 3 movl
instructions, 2 shifts, 1 add, and 1 substract than to do 1 mov, 1
cdq, and 1 idiv?

I run into this while trying to understand why my code runs slower
with llvm than a comparable program on windows.

Thanks for any help,
Brent

int f(int n)
{
    return (n + 1) % 18;
}

"clang -O2 -S" produces this code:

_f:                                     # @f
# BB#0:
	movl	4(%esp), %ecx
	incl	%ecx
	movl	$954437177, %edx        # imm = 0x38E38E39
	movl	%ecx, %eax
	imull	%edx
	movl	%edx, %eax
	shrl	$31, %eax
	sarl	$2, %edx
	addl	%eax, %edx
	imull	$18, %edx, %eax
	subl	%eax, %ecx
	movl	%ecx, %eax
	ret


The visual studio compiler (/O2) instead issues the idiv instruction:

PUBLIC	_f
; Function compile flags: /Ogtpy
;	COMDAT _f
_TEXT	SEGMENT
_n$ = 8							; size = 4
_f	PROC						; COMDAT
; File c:\a.c
; Line 6
	mov	eax, DWORD PTR _n$[esp-4]
	inc	eax
	cdq
	mov	ecx, 18					; 00000012H
	idiv	ecx
	mov	eax, edx
; Line 7
	ret	0
_f	ENDP
_TEXT	ENDS
END