https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119374
Bug ID: 119374
Summary: armv7ve+neon-vfpv4: Miscompilation with -O2
Product: gcc
Version: 14.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: myokaski at gmail dot com
Target Milestone: ---
Created attachment 60823
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=60823&action=edit
testcase
gcc -march=armv7ve+neon-vfpv4 -O2 ftest.c
produces wrong result with GCC 14.2.0 and 12.2.0.
11.4.0 seems to be ok.
Correct output:
shift = 4
xx = 4096
xy = 0
yx = 0
yy = 4096
temp1 = 536870912
temp2 = 33554432
check = 1
Wrong output:
shift = 4
xx = 65536
xy = 0
yx = 0
yy = 65536
temp1 = 0
temp2 = 0
check = 0
Works normally with -O0/-fno-tree-vectorize.
The assembly output of this part:
if ( shift > 0 ) // shift is 4 here: ( 31 - __builtin_clz( val ) ) - 12
{
xx >>= shift;
xy >>= shift;
yx >>= shift;
yy >>= shift;
}
is
cmp r4, #0
ble .L2
vshl.i32 q4, q4, #0 // here nothing is done
...
.L2
If shift is set to immediate by 'shift = 4;', asm is generated correctly:
'vshr.s32 q4, q4, #4'
Correct result is also generated if I remove temp1/temp2-related code.
Original code was FT_Matrix_Check function from freetype library sources.
