the ffmpeg test idct8x8 (NEON) fails when built with binutils from the trunk. The ffmpeg (2.8.1) tests succeed when built with binutils from the 2.25 branch. TEST idct8x8 ./tests/fate-run.sh fate-idct8x8 "" "" "/home/ubuntu/tmp/ffmpeg-2.8.1" 'run libavcodec/dct-test -i' 'null' '/dev/null' '' '1' '' '' '' '' '' '' '' '' /home/ubuntu/tmp/ffmpeg-2.8.1/libavcodec/dct-test -i ffmpeg DCT/IDCT test 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 IDCT REF-DBL: max_err=0 omse=0.00000000 ome=0.00000000 syserr=0.00000000 maxout=288 blockSumErr=0 54 -33 10 21 24 33 13 -2 23 36 -6 1 12 13 34 33 5 37 18 38 -13 41 38 32 -4 22 28 11 13 18 10 18 -2 9 38 -12 10 18 -5 50 32 43 5 23 17 36 25 -11 18 9 33 -1 17 3 34 -10 5 23 41 0 15 10 23 29 IDCT INT: max_err=1 omse=0.01303203 ome=0.00086172 syserr=0.00270000 maxout=288 blockSumErr=64 1 -15 -2 2 -2 -2 -4 14 -12 -5 2 -5 -4 -5 -16 -10 9 -6 5 -19 -3 7 -4 5 -7 5 11 -19 -18 -11 -4 -2 -19 -6 4 19 -7 -6 -2 -3 -8 9 20 -13 -3 13 9 -6 16 8 -17 -11 4 -17 5 -7 -7 15 17 -16 1 16 -6 -8 IDCT SIMPLE-C: max_err=1 omse=0.00675937 ome=-0.00009375 syserr=0.00100000 maxout=288 blockSumErr=64 19 -18 5 38 75 29 2 47 40 77 25 18 30 19 -13 29 28 43 -13 31 25 18 -6 32 22 48 66 12 8 14 11 25 1 3 28 0 22 31 48 55 20 17 40 8 45 52 21 4 16 28 36 2 45 34 69 33 42 21 9 36 -3 17 31 24 IDCT PR-C: max_err=1 omse=0.02551797 ome=0.00126641 syserr=0.00385000 maxout=288 blockSumErr=64 2 -3 0 1 2 0 0 1 -5 -4 -2 0 -4 -4 -4 1 0 1 2 -1 0 -1 -1 3 1 1 -2 2 -1 1 1 2 2 -1 0 -1 2 1 1 1 2 -2 -1 3 2 -2 -2 3 2 -2 -1 3 2 -2 -2 3 1 1 2 2 1 1 1 2 IDCT FAANI: max_err=1 omse=0.00046484 ome=0.00000859 syserr=0.00025000 maxout=288 blockSumErr=64 364 346 377 372 368 377 344 368 151 144 125 164 155 166 138 155 152 152 143 143 146 149 145 160 -129 -140 -116 -131 -152 -133 -124 -122 -137 -142 -108 -144 -129 -151 -138 -138 203 225 230 223 208 208 204 208 225 219 258 200 213 214 244 237 52 53 39 29 39 66 32 34 IDCT XVID: max_err=1 omse=0.01386797 ome=0.00549453 syserr=0.01885000 maxout=288 blockSumErr=48 -2384 -15 230638 2 21676 -1 -278477 15 15651 -3 -87455 -5 -1232 -4 77956 -10 28105 -6 -26767 -19 -28995 7 59199 7 -13554 5 -18582 -18 48593 -11 -4184 -2 12675 -5 5166 19 -30168 -6 -36187 -2 20222 9 82445 -12 -20193 13 -59735 -5 30518 8 43440 -11 -41067 -17 -18556 -7 -27035 15 1879 -16 -16166 16 -40682 -6 IDCT SIMPLE-ARM: max_err=2019 omse=32415.48331328 ome=-0.05728203 syserr=13.92385000 maxout=2040 blockSumErr=21082 52 -11 2 21 23 23 32 3 21 32 -5 8 11 17 30 35 17 37 36 37 -4 26 40 27 -3 31 30 14 6 10 7 20 -5 7 48 -13 12 26 -3 54 30 49 -2 32 7 32 27 -18 12 9 36 -5 17 -1 26 -4 9 12 27 -1 19 17 20 29 IDCT INT-ARM: max_err=1 omse=0.01301562 ome=0.00088281 syserr=0.00270000 maxout=288 blockSumErr=64 1 -15 -2 2 -2 -2 -4 14 -12 -5 2 -5 -4 -5 -16 -10 9 -6 5 -19 -3 7 -4 5 -7 5 11 -19 -18 -11 -4 -2 -19 -6 4 19 -7 -6 -2 -3 -8 9 20 -13 -3 13 9 -6 16 8 -17 -11 4 -17 5 -7 -7 15 17 -16 1 16 -6 -8 IDCT SIMPLE-ARMV5TE: max_err=1 omse=0.00675937 ome=-0.00009375 syserr=0.00100000 maxout=288 blockSumErr=64 4 -15 -1 2 -2 -1 -4 15 -12 -3 3 -5 -4 -4 -16 -10 9 -6 5 -19 -3 7 -4 7 -5 5 12 -18 -18 -11 -3 -2 -19 -5 4 19 -6 -6 -2 -2 -8 9 20 -12 -2 13 9 -5 18 8 -17 -11 4 -17 7 -7 -7 15 17 -16 1 16 -6 -6 IDCT SIMPLE-ARMV6: max_err=1 omse=0.00675703 ome=-0.00007109 syserr=0.00100000 maxout=288 blockSumErr=64 1 -15 -2 2 -2 -2 -4 14 -12 -5 2 -5 -4 -5 -16 -10 9 -6 5 -19 -3 7 -4 5 -7 5 11 -19 -18 -11 -4 -2 -19 -6 4 19 -7 -6 -2 -3 -8 9 20 -13 -3 13 9 -6 16 8 -17 -11 4 -17 5 -7 -7 15 17 -16 1 16 -6 -8 IDCT SIMPLE-NEON: max_err=1 omse=0.00675937 ome=-0.00009375 syserr=0.00100000 maxout=288 blockSumErr=64 Error: 1. Test idct8x8 failed. Look at tests/data/fate/idct8x8.err for details. tests/Makefile:203: recipe for target 'fate-idct8x8' failed make[2]: *** [fate-idct8x8] Error 1 The mentioned .err file is empty.
I'll take a look.
more info at https://launchpad.net/bugs/1513985
(In reply to Matthias Klose from comment #2) > more info at https://launchpad.net/bugs/1513985 Could you add the options used to build the test case, either here or on the other bug report. The command line passed to the ffmpeg configure script should be enough. Thanks, Matthew
Created attachment 8774 [details] test case from https://bugs.launchpad.net/ubuntu/+source/binutils/+bug/1513985/comments/3 """ Attached testcase test.S doesn't require compiling ffmpeg, but simply: $ gcc -g -c -o test.o test.S Analyzing the difference of 'objdump -d test.o' between variants built with working and broken binutils show that the working binutils always always uses 'ldr.w' for ldr, while the broken version sometimes optimizes this to 'movw', which is a correct alternative, but one time to 'movt', which only sets half of the register." """
This appears to be due to an optimization introduced by PR gas/18499. The patch submitted to fix it is at https://sourceware.org/ml/binutils/2015-11/msg00072.html.
The master branch has been updated by Ramana Radhakrishnan <ramana@sourceware.org>: https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=582cfe03cb7496371bc6d83af49b6324d0c36887 commit 582cfe03cb7496371bc6d83af49b6324d0c36887 Author: Ramana Radhakrishnan <ramana.radhakrishnan@arm.com> Date: Thu Nov 12 10:50:22 2015 +0000 Fix PR gas/19217 2015-11-11 Matthew Wahab <matthew.wahab@arm.com> PR gas/19217 * config/tc-arm.c (move_or_literal_pool): Remove redundant feature check. Fix some code formatting. Drop use of MOVT. Add some comments. 2015-11-11 Matthew Wahab <matthew.wahab@arm.com> PR gas/19217 * gas/arm/thumb2_ldr_immediate_armv6t2.d: Update expected output.