http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55342
Bug #: 55342
Summary: [LRA,x86] Non-optimal code for simple loop with LRA
Classification: Unclassified
Product: gcc
Version: 4.8.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: ysrum...@gmail.com
Target: x86
For a simple test-case we got -15% regression with LRA on x86 in 32-bit mode.
The test-case is
#define byte unsigned char
#define MIN(a, b) ((a) > (b)?(b):(a))
void convert_image(byte *in, byte *out, int size) {
int i;
byte * read = in,
* write = out;
for(i = 0; i < size; i++) {
byte r = *read++;
byte g = *read++;
byte b = *read++;
byte c, m, y, k, tmp;
c = 255 - r;
m = 255 - g;
y = 255 - b;
if (c < m)
k = MIN (c, y);
else
k = MIN (m, y);
*write++ = c - k;
*write++ = m - k;
*write++ = y - k;
*write++ = k;
}
}
The essential part of assembly is (it is correspondent to write-part of loop):
without LRA
.L4:
movl %esi, %ecx
addl $4, %eax
subl %ecx, %ebx
movzbl 3(%esp), %ecx
movb %bl, -4(%eax)
movl %esi, %ebx
subl %ebx, %edx
movb %dl, -2(%eax)
subl %ebx, %ecx
movb %cl, -3(%eax)
cmpl %ebp, 4(%esp)
movb %bl, -1(%eax)
je .L1
with LRA
.L4:
movl %esi, %eax
subl %eax, %ebx
movl 28(%esp), %eax
movb %bl, (%eax)
movl %esi, %eax
subl %eax, %ecx
movl 28(%esp), %eax
movb %cl, 1(%eax)
movl %esi, %eax
subl %eax, %edx
movl 28(%esp), %eax
movb %dl, 2(%eax)
addl $4, %eax
movl %eax, 28(%esp)
movl 28(%esp), %ecx
movl %esi, %eax
cmpl %ebp, (%esp)
movb %al, -1(%ecx)
je .L1
I also wonder why additional moves are required to perform subtraction:
movl %esi, %eax
subl %eax, %ebx
whereas only one instruction is required:
subl %esi, %ebx.
I assume that this part is not related to LRA.
