Bug 77845

Summary: LTO accumulates CPU requirements from all input objects (reopen)
Product: gcc Reporter: andysem
Component: targetAssignee: Not yet assigned to anyone <unassigned>
Status: NEW ---    
Severity: normal CC: hubicka, rguenth, tulipawn, uros
Priority: P3 Keywords: lto
Version: 5.4.0   
Target Milestone: ---   
Host: Target: x86_64-*-*, i?86-*-*
Build: Known to work:
Known to fail: 5.4.0, 6.2.0 Last reconfirmed: 2016-10-05 00:00:00
Attachments: The testcase that miscompiles with gcc 5.4

Description andysem 2016-10-04 17:43:25 UTC 
Created attachment 39752 [details]
The testcase that miscompiles with gcc 5.4

This is a continuation of bug 61043.

In the attached testcase, the compiler produces my_simd.so library with two functions (my_simd_func_sse2 and my_simd_func_avx), which are implemented in separate TUs, which are compiled with -msse2 and -mavx, respectively. However, the disassembler shows that both my_simd_func_sse2 and my_simd_func_avx contain VEX-encoded instructions (i.e. the compiler used AVX in both cases). This only happens when LTO is used. See build.sh for compiler command lines.

$ g++ -v
Using built-in specs.
COLLECT_GCC=g++
COLLECT_LTO_WRAPPER=/usr/lib/gcc/x86_64-linux-gnu/5/lto-wrapper
Target: x86_64-linux-gnu
Configured with: ../src/configure -v --with-pkgversion='Ubuntu 5.4.0-6ubuntu1~16.04.2' --with-bugurl=file:///usr/share/doc/gcc-5/README.Bugs --enable-languages=c,ada,c++,java,go,d,fortran,objc,obj-c++ --prefix=/usr --program-suffix=-5 --enable-shared --enable-linker-build-id --libexecdir=/usr/lib --without-included-gettext --enable-threads=posix --libdir=/usr/lib --enable-nls --with-sysroot=/ --enable-clocale=gnu --enable-libstdcxx-debug --enable-libstdcxx-time=yes --with-default-libstdcxx-abi=new --enable-gnu-unique-object --disable-vtable-verify --enable-libmpx --enable-plugin --with-system-zlib --disable-browser-plugin --enable-java-awt=gtk --enable-gtk-cairo --with-java-home=/usr/lib/jvm/java-1.5.0-gcj-5-amd64/jre --enable-java-home --with-jvm-root-dir=/usr/lib/jvm/java-1.5.0-gcj-5-amd64 --with-jvm-jar-dir=/usr/lib/jvm-exports/java-1.5.0-gcj-5-amd64 --with-arch-directory=amd64 --with-ecj-jar=/usr/share/java/eclipse-ecj.jar --enable-objc-gc --enable-multiarch --disable-werror --with-arch-32=i686 --with-abi=m64 --with-multilib-list=m32,m64,mx32 --enable-multilib --with-tune=generic --enable-checking=release --build=x86_64-linux-gnu --host=x86_64-linux-gnu --target=x86_64-linux-gnu
Thread model: posix
gcc version 5.4.0 20160609 (Ubuntu 5.4.0-6ubuntu1~16.04.2)
Comment 1 Richard Biener 2016-10-05 09:12:39 UTC 
the lto_opts section still contains -mavx and -msse2 for the respective objects
which are passed down at link time.  It looks like the DECL_FUNCTION_SPECIFIC_TARGET node free-lang-data (is supposed to) installs
doesn't "undo" -mavx fully.  I suppose the ix86_sse2avx flag is to handle this.

Confirmed for GCC 5 and GCC 6.

Breakpoint 1, ix86_function_specific_restore (opts=0x2440b80 <global_options>, 
    ptr=0x7ffff6a6cbd0)
    at /space/rguenther/src/svn/gcc-6-branch/gcc/config/i386/i386.c:5748
warning: Source file is more recent than executable.
5748      int i;
Missing separate debuginfos, use: zypper install libgmp10-debuginfo-6.0.0-71.1.x86_64 libisl15-debuginfo-0.16.1-0.x86_64 libmpc3-debuginfo-1.0.2-38.2.x86_64 libmpfr4-debuginfo-3.1.2-3.1.2.x86_64
(gdb) p opts->x_ix86_sse2avx
$1 = 0
(gdb) p ptr->x_ix86_sse2avx
$2 = 0
(gdb) p cfun
$3 = (function *) 0x7ffff6a702a0
(gdb) p debug_generic_expr (cfun->decl)
my_simd_func_avx

so I may be mistaken.  Not sure how saving/restoring of flags like -mavx or -msse2 is supposed to work.