The ivopts pass fails to set the REG_POINTER attribute on a pseudo that is
equal to two pseudos, one of which has teh REG_POINTER attribute and the other
is just a n offset from it. This leads us to not sort the operands on an
indexed load/store instruction causing performance problems on POWER6.
The test case used to recreate the problem is:
linux> cat > bug.c <<EOF
unsigned int *quadrant;
unsigned int
fullGtU (int i1, int n)
{
unsigned int s1 = 0;
for (; n >= 0; n--)
{
s1 = quadrant[i1];
i1 += n;
}
return s1;
}
EOF
linux> /path/to/gcc -O1 -S -da bug.c
Resulting assembly:
fullGtU:
li 0,0
cmpwi 0,4,0
blt 0,.L3
lis 9,[EMAIL PROTECTED]
lwz 11,[EMAIL PROTECTED](9)
addi 0,4,1
mtctr 0
addi 0,4,-1
cmpwi 7,0,-1
bge+ 7,.L4
li 0,1
mtctr 0
.L4:
slwi 9,3,2
lwzx 0,9,11 <- bad operand ordering
add 3,3,4
addi 4,4,-1
bdnz .L4
.L3:
mr 3,0
blr
Bad RTL in bug.c.127r.expand:
(insn 21 20 22 bug.c:8 (set (reg/f:SI 129)
(plus:SI (reg:SI 128)
(reg:SI 121 [ quadrant.0 ]))) -1 (nil))
(insn 22 21 0 bug.c:8 (set (reg/v:SI 120 [ s1 ])
(mem:SI (reg/f:SI 129) [0 S4 A32])) -1 (nil))
Eventually, pseudo 129 )which has the REG_POINTER attribute) gets exchanged
with the (plus:SI (reg:SI 128) (reg:SI 121 [ quadrant.0 ])) and pseudo 121
doesn't have the attribute, so we end up failing to swap them.
--
Summary: ivopts fails to set REG_POINTER attribute on pseudo
Product: gcc
Version: 4.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: bergner at gcc dot gnu dot org
GCC build triplet: powerpc64-linux
GCC host triplet: powerpc64-linux
GCC target triplet: powerpc64-linux
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32940
