http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53538
Bug #: 53538
Summary: Compile error with ARM NEON intrinsics and certain
compile flags
Classification: Unclassified
Product: gcc
Version: 4.5.3
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: ah...@gmx.net
Created attachment 27529
--> http://gcc.gnu.org/bugzilla/attachment.cgi?id=27529
test program
The attached test program fails to compile with the following command line
(cross compiling for ARM neon on Ubuntu):
gcc -c test_an_bug.c -o test_an_bug.o -O2 -g -Wall -march=armv7-a
-mlittle-endian -mfloat-abi=hard -mfpu=neon -mtp=cp15 -mthumb
Error message:
In file included from test_an_bug.c:2:0:
[...]/4.5/include/arm_neon.h: In function 'main':
[...]/4.5/include/arm_neon.h:4968:49: error: argument must be a constant
[...]/4.5/include/arm_neon.h:4968: confused by earlier errors, bailing out
However it builds without warning when either:
1) use -O3 instead of -O2 or
2) use -funroll-loops
3) if you move the intrinsic (vget_lane_u16) outside the for loop
4) if you use literals instead of the variable j
I haven't found any detailed documentation on the use of this intrinsic, but it
looks a bit odd if it only accepts literals as parameters.
At what point did we give up pass-by-value for int?
And it certainly is odd that it compiles fine with different optimization
levels...
Note 1: This is not the original code, it is the minimum test program I derived
from it.
Note 2: It seems to fail on all compiler versions from 4.5 to 4.7 I've tried.
