switch lowering makes ugly code for a simple scanning loop

[lattner]

Bugzilla Link	964
Resolution	FIXED
Resolved on	Feb 22, 2010 12:41
Version	1.8
OS	All

Extended Description

Consider this function (very similar to strhr):

void foo(unsigned char C);
const char *FindChar(const char *CurPtr) {
unsigned char C;
do
C = *CurPtr++;
while (C != 'x' && C != '\0');

foo(C);
return CurPtr;
}

We currently compile the loop to:

LBB1_1: #bb
movb (%esi), %al
incl %esi
cmpb $119, %al
jg LBB1_4 #bb
LBB1_3: #bb
testb %al, %al
je LBB1_2 #bb7
jmp LBB1_1 #bb
LBB1_4: #bb
cmpb $120, %al
jne LBB1_1 #bb

It seems that switch lowering could produce something like:

    cmpb $120, %al
    je out
    testb %al, %al
    jnz LBB1_1       #bb7

which would be much faster.

-Chris

[lattner]

assigned to @lattner

[lattner]

GCC generates code very close to the desired code:

L3:
movzbl (%esi), %eax
addl $1, %esi
cmpb $120, %al
je L4
testb %al, %al
jne L3

[lattner]

Implemented. Testcase here: Regression/CodeGen/Generic/SwitchLowering.ll

Patch here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20061016/038932.html

We now compile the inner loop to:

LBB1_1: #bb
movb (%esi), %al
incl %esi
testb %al, %al
je LBB1_2 #bb7
LBB1_3: #bb
cmpb $120, %al
jne LBB1_1 #bb

-Chris