UNINIT reports on libjpeg_turbo [SSE2 instructions]

[derekbruening]

From [email protected] on August 04, 2011 10:49:28

http://build.chromium.org/p/chromium.fyi/builders/Windows%20Tests%20%28DrMemory%29/builds/4691/steps/memory%20test%3A%20unit_1/logs/stdio UNINITIALIZED READ: reading 0x407e0738-0x407e0740 8 byte(s) within 0x407e0730-0x407e0740
#1 jsimd_ycc_extbgrx_convert_sse2
#2 jsimd_ycc_rgb_convert third_party\libjpeg_turbo\simd\jsimd_i386.c:199
#3 jsimd_ycc_rgb_convert third_party\libjpeg_turbo\simd\jsimd_i386.c:199
#4 sep_upsample third_party\libjpeg_turbo\jdsample.c:133
#5 process_data_context_main third_party\libjpeg_turbo\jdmainct.c:432
#6 chromium_jpeg_read_scanlines third_party\libjpeg_turbo\jdapistd.c:173
#7 gfx::JPEGCodec::Decode ui\gfx\codec\jpeg_codec.cc:539
#8 gfx::JPEGCodec::Decode ui\gfx\codec\jpeg_codec.cc:594
#9 history::ExpireHistoryTest::AddExampleData chrome\browser\history\expire_history_backend_unittest.cc:232
#10 history::ExpireHistoryTest_ArchiveHistoryBeforeStarred_Test::TestBody chrome\browser\history\expire_history_backend_unittest.cc:779
// the instruction in the global.PID.log during a local run is "movdqa (%esi) -> %xmm5"
// TODO: find out where does it propagate from. [ issue #170 would help here]

If "set JSIMD_FORCEMMX=1" to switch SSE2 implementation with MMX, the same test has 0 uninit reports.

SSE2 deals with floating-point operands and we require all bits to be set for FP registers [xref issue #471 ], that looks like the reason for the reports.

FTR, jsimd_h2v2_fancy_upsample_sse2 assembly is codesearch'able.

Original issue: http://code.google.com/p/drmemory/issues/detail?id=540

[mistmist]

Guess I've run into the same problem with LibreOffice, although somehow I don't get any uninit warnings from libjpeg-turbo itself; the earliest warnings I get are when the internal bitmap is written out as a PNG, so it took me some time to track down...

The uninit bytes can first be detected in the scan line buffer that comes out of jpeg-turbo's jpeg_read_scanlines(); naturally there's no conditionals in that place so it's not possible to write a suppression entry for it.

JSIMD_FORCEMMX=1 does work around the problem.

By the way, DrMemory really needs a equivalent of valgrind's memcheck's --track-origins=yes.

[zhaoqin]

On Nov 26, 2015 4:34 PM, "mistmist" [email protected] wrote:

Guess I've run into the same problem with LibreOffice, although somehow I
don't get any uninit warnings from libjpeg-turbo itself; the earliest
warnings I get are when the internal bitmap is written out as a PNG, so it
took me some time to track down...

The uninit bytes can first be detected in the scan line buffer that comes
out of jpeg-turbo's jpeg_read_scanlines(); naturally there's no
conditionals in that place so it's not possible to write a suppression
entry for it.

JSIMD_FORCEMMX=1 does work around the problem.

By the way, DrMemory really needs a equivalent of valgrind's memcheck's
--track-origins=yes.

It takes significant efforts to add that, we welcome any contributions!
—
Reply to this email directly or view it on GitHub.