| Message ID | fd2ce78bb11d457481cdcca122c1ad64@amazon.com |
|---|---|
| State | New |
| Headers | show |
| Series | [FFmpeg-devel,v2] add a configure flag to enabled tree-vecorization with gcc | expand |
| Context | Check | Description |
|---|---|---|
| yinshiyou/commit_msg_loongarch64 | warning | The first line of the commit message must start with a context terminated by a colon and a space, for example "lavu/opt: " or "doc: ". |
| andriy/commit_msg_x86 | warning | The first line of the commit message must start with a context terminated by a colon and a space, for example "lavu/opt: " or "doc: ". |
| yinshiyou/make_loongarch64 | success | Make finished |
| yinshiyou/make_fate_loongarch64 | success | Make fate finished |
| andriy/make_fate_x86 | success | Make fate finished |
| andriy/make_x86 | warning | New warnings during build |
Aug 8, 2022, 17:25 by jswinney@amazon.com: > Recent version of gcc improve the automatic vectorization. This flag > allows adventurous users to enable vectorization. Known problems with > this are primarily related to inline assembly for x86 and so to address > those, add a pragma to explicitly disable automatic vectorization for > those files. > > Signed-off-by: Jonathan Swinney <jswinney@amazon.com> > > -- > > Thank you considering this patch. I believe this addresses the primary > concerns that were raised by my previous submission. There may be more > files which require the pragma add `-fno-tree-vectorize`, and I welcome > suggestions. This should strike a compromise, allowing some users to > enable vectorization while not breaking mainstream builds. This should > give time to work out additional problems if they arise before enabling > vectorization more broadly. > I dislike this, pretty soon we'll end up with compiler version checks whenever vectorization breaks. Either gcc should fix miscompilation, or patches should be sent to write assembly (which they should be anyway).
diff --git a/configure b/configure index cbbb4dd9c8..8e842da1b8 100755 --- a/configure +++ b/configure @@ -110,6 +110,7 @@ Configuration options: --disable-swscale-alpha disable alpha channel support in swscale --disable-all disable building components, libraries and programs --disable-autodetect disable automatically detected external libraries [no] + --enable-auto-vectorization enable compiler auto vectorization Program options: --disable-programs do not build command line programs @@ -1945,6 +1946,7 @@ FEATURE_LIST=" small static swscale_alpha + auto_vectorization " # this list should be kept in linking order @@ -7176,7 +7178,9 @@ if enabled icc; then disable aligned_stack fi elif enabled gcc; then - check_optflags -fno-tree-vectorize + if disabled auto_vectorization; then + check_optflags -fno-tree-vectorize + fi check_cflags -Werror=format-security check_cflags -Werror=implicit-function-declaration check_cflags -Werror=missing-prototypes @@ -7569,6 +7573,7 @@ echo "pod2man enabled ${pod2man-no}" echo "makeinfo enabled ${makeinfo-no}" echo "makeinfo supports HTML ${makeinfo_html-no}" echo "xmllint enabled ${xmllint-no}" +echo "auto-vectorization ${auto_vectorization-no}" test -n "$random_seed" && echo "random seed ${random_seed}" echo diff --git a/libavcodec/x86/cabac.h b/libavcodec/x86/cabac.h index b046a56a6b..782e4cbda4 100644 --- a/libavcodec/x86/cabac.h +++ b/libavcodec/x86/cabac.h @@ -39,6 +39,10 @@ #if HAVE_INLINE_ASM +#ifdef __GNUC__ + __attribute__((optimize("-fno-tree-vectorize"))) +#endif + #ifndef UNCHECKED_BITSTREAM_READER #define UNCHECKED_BITSTREAM_READER !CONFIG_SAFE_BITSTREAM_READER #endif
Recent version of gcc improve the automatic vectorization. This flag allows adventurous users to enable vectorization. Known problems with this are primarily related to inline assembly for x86 and so to address those, add a pragma to explicitly disable automatic vectorization for those files. Signed-off-by: Jonathan Swinney <jswinney@amazon.com> -- Thank you considering this patch. I believe this addresses the primary concerns that were raised by my previous submission. There may be more files which require the pragma add `-fno-tree-vectorize`, and I welcome suggestions. This should strike a compromise, allowing some users to enable vectorization while not breaking mainstream builds. This should give time to work out additional problems if they arise before enabling vectorization more broadly. --- configure | 7 ++++++- libavcodec/x86/cabac.h | 4 ++++ 2 files changed, 10 insertions(+), 1 deletion(-)