Fork#39565: Incorrect alignment of SIMD vectors returned by value from functions
Open Date: 2019-09-12 15:47
Last Update: 2021-04-14 03:53
URL for this Ticket:
https://osdn.net//projects/mingw/ticket/39565
RSS feed for this Ticket:
https://osdn.net/ticket/ticket_rss.php?group_id=3917&tid=39565
---------------------------------------------------------------------
Last Changes/Comment on this Ticket:
2021-04-14 03:53 Updated by: michal_fapso
Comment:
Thanks a lot for your answer, @keith, and for testing it on linux in the cross-compile mode. I wasn't aware of any trademark issues between MinGW and MSYS2. I'm sorry to hear that. I thought that msys2 includes the mingw gcc compiler package just like it includes other open source packages. I know that they apply some patches on top of MinGW, but I thought that this kind of bug is so low level that it has to be directly in MinGW sources, not in msys2 patches. Eventually, it seems this bug is msys2 specific. Anyway, I like your idea of cross-compiling on linux for windows. Thanks, you can close this issue.
---------------------------------------------------------------------
Ticket Status:
Reporter: michal_fapso
Owner: (None)
Type: Issues
Status: Closed
Priority: 5 - Medium
MileStone: (None)
Component: MSYS
Severity: 5 - Medium
Resolution: Invalid
---------------------------------------------------------------------
Ticket details:
Summary of the Issue
Working with AVX2 SIMD vectors. When a __m256i value is returned from a function, the SIMD register ymm0 is copied to memory using the aligned move instruction vmovdqa, so the destination memory address must be aligned to 32 bytes. This requirement is not met in my test case and the program crashes with SIGSEGV. More info follows in the "Minimal Test Case" section below.
I tested it also on Ubuntu's gcc, but it didn't crash there, neither did with clang++ on Windows and Ubuntu, so therefore I am submitting this issue to the MinGW tracker. But I might be wrong and the issue may exist in gcc as well.
Host Operating System Information and Version
Windows 10 Home, Version 10.0.17134 Build 17134
with latest updated MSYS2.
GCC Version
$ gcc -v
Using built-in specs.
COLLECT_GCC=D:\msys64_latest\mingw64\bin\gcc.exe
COLLECT_LTO_WRAPPER=D:/msys64_latest/mingw64/bin/../lib/gcc/x86_64-w64-mingw32/9.2.0/lto-wrapper.exe
Target: x86_64-w64-mingw32
Configured with: ../gcc-9.2.0/configure --prefix=/mingw64 --with-local-prefix=/mingw64/local --build=x86_64-w64-mingw32 --host=x86_64-w64-mingw32 --target=x86_64-w64-mingw32 --with-native-system-header-dir=/mingw64/x86_64-w64-mingw32/include --libexecdir=/mingw64/lib --enable-bootstrap --with-arch=x86-64 --with-tune=generic --enable-languages=c,lto,c++,fortran,ada,objc,obj-c++ --enable-shared --enable-static --enable-libatomic --enable-threads=posix --enable-graphite --enable-fully-dynamic-string --enable-libstdcxx-filesystem-ts=yes --enable-libstdcxx-time=yes --disable-libstdcxx-pch --disable-libstdcxx-debug --disable-isl-version-check --enable-lto --enable-libgomp --disable-multilib --enable-checking=release --disable-rpath --disable-win32-registry --disable-nls --disable-werror --disable-symvers --enable-plugin --with-libiconv --with-system-zlib --with-gmp=/mingw64 --with-mpfr=/mingw64 --with-mpc=/mingw64 --with-isl=/mingw64 --with-pkgversion='Rev2, Built by MSYS2 project' --with-bugurl=https://sourceforge.net/projects/msys2 --with-gnu-as --with-gnu-ld
Thread model: posix
gcc version 9.2.0 (Rev2, Built by MSYS2 project)
Binutils Version
$ ld -v
GNU ld (GNU Binutils) 2.32
MinGW Version
I don't know what should I look for in the mingw/include/_mingw.h file.
Build Environment
$ uname -a
MINGW64_NT-10.0-17134 MISO-PC 3.0.7-338.x86_64 2019-07-11 10:58 UTC x86_64 Msys
Minimal Self-Contained Test Case
main.cpp:
#include <iostream> #include <immintrin.h> #define DBG(var_name) std::cout<<#var_name": "<<(var_name)<<std::endl // Output operator for vector std::ostream& operator<<(std::ostream& oss, const __m256i& v) { constexpr size_t length_bytes = 32; unsigned char a[length_bytes]; _mm256_storeu_si256(reinterpret_cast<__m256i*>(a), v); oss << "["; std::string sep = ""; for (size_t i=0; i<length_bytes; i++) { oss << sep << int(a[i]); sep = " "; } return oss << "]"; } __m256i __attribute__ ((noinline)) crash_vector_ret(uint8_t* a) { __m256i v; v = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(a)); // Crash return v; } int main(int argc, char** argv) { // Setup memory from which the vector will be loaded const int a_size = ARRAY_SIZE; uint8_t a[a_size]; for (volatile int i=0; i<a_size; i++) { a[i] = i; } DBG(alignof(__m256i)); __m256i vr; vr = crash_vector_ret(a); DBG(vr); return 0; }Makefile:
CXXFLAGS += -mavx2 CXXFLAGS += -std=c++17 CXXFLAGS += -g CXX = g++ main.exe: main.cpp $(CXX) $(CXXFLAGS) -O1 -DARRAY_SIZE=32 -o $@ $< test: main.cpp for COMPILER in g++; do \ ASM_FLAGS="-S -fverbose-asm -masm=intel"; \ for ARRAY_SIZE in 32 36; do \ for OPT in 0 1 2 3; do \ NAME="main_O$${OPT}_$${COMPILER}_s$${ARRAY_SIZE}"; \ echo "--------------------------------------------------"; \ echo "NAME:$${NAME}"; \ rm -f $${NAME}.exe; \ $${COMPILER} $(CXXFLAGS) -DARRAY_SIZE=$${ARRAY_SIZE} -O$${OPT} -o $${NAME}.exe main.cpp; \ $${COMPILER} $(CXXFLAGS) -DARRAY_SIZE=$${ARRAY_SIZE} -O$${OPT} $${ASM_FLAGS} -o $${NAME}.s main.cpp; \ ./$${NAME}.exe; \ done; \ done; \ doneTrying various compilation options:
$ make test
for COMPILER in g++; do \
ASM_FLAGS="-S -fverbose-asm -masm=intel"; \
for ARRAY_SIZE in 32 36; do \
for OPT in 0 1 2 3; do \
NAME="main_O${OPT}_${COMPILER}_s${ARRAY_SIZE}"; \
echo "--------------------------------------------------"; \
echo "NAME:${NAME}"; \
rm -f ${NAME}.exe; \
${COMPILER} -mavx2 -std=c++17 -g -DARRAY_SIZE=${ARRAY_SIZE} -O${OPT} -o ${NAME}.exe main.cpp; \
${COMPILER} -mavx2 -std=c++17 -g -DARRAY_SIZE=${ARRAY_SIZE} -O${OPT} ${ASM_FLAGS} -o ${NAME}.s main.cpp; \
./${NAME}.exe; \
done; \
done; \
done
--------------------------------------------------
NAME:main_O0_g++_s32
alignof(__m256i): 32
/bin/sh: line 3: 15627 Segmentation fault ./${NAME}.exe
--------------------------------------------------
NAME:main_O1_g++_s32
alignof(__m256i): 32
/bin/sh: line 3: 15631 Segmentation fault ./${NAME}.exe
--------------------------------------------------
NAME:main_O2_g++_s32
alignof(__m256i): 32
/bin/sh: line 3: 15635 Segmentation fault ./${NAME}.exe
--------------------------------------------------
NAME:main_O3_g++_s32
alignof(__m256i): 32
/bin/sh: line 3: 15639 Segmentation fault ./${NAME}.exe
--------------------------------------------------
NAME:main_O0_g++_s36
alignof(__m256i): 32
/bin/sh: line 3: 15643 Segmentation fault ./${NAME}.exe
--------------------------------------------------
NAME:main_O1_g++_s36
alignof(__m256i): 32
vr: [0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31]
--------------------------------------------------
NAME:main_O2_g++_s36
alignof(__m256i): 32
vr: [0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31]
--------------------------------------------------
NAME:main_O3_g++_s36
alignof(__m256i): 32
vr: [0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31]
GDB session
Program compiled by
g++ -mavx2 -std=c++17 -g -DARRAY_SIZE=32 -O0 -o main_O0_g++_s32.exe main.cpp
$ gdb main_O0_g++_s32
GNU gdb (GDB) 8.3
Copyright (C) 2019 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.
Type "show copying" and "show warranty" for details.
This GDB was configured as "x86_64-w64-mingw32".
Type "show configuration" for configuration details.
For bug reporting instructions, please see:
<http://www.gnu.org/software/gdb/bugs/>.
Find the GDB manual and other documentation resources online at:
<http://www.gnu.org/software/gdb/documentation/>.
For help, type "help".
Type "apropos word" to search for commands related to "word"...
Reading symbols from main_O0_g++_s32...
(gdb) r
Starting program: D:\SourceCode\cpp_simdpp_crash_loadu\main_O0_g++_s32.exe
[New Thread 10112.0x4c8]
[New Thread 10112.0x409c]
[New Thread 10112.0x6bf0]
alignof(__m256i): 32
Thread 1 received signal SIGSEGV, Segmentation fault.
0x00000000004016ea in crash_vector_ret (a=0x66fe10 "") at main.cpp:24
24 v = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(a));
(gdb) bt
#0 0x00000000004016ea in crash_vector_ret (a=0x66fe10 "") at main.cpp:24
#1 0x000000000040179f in main (argc=1, argv=0xe64a60) at main.cpp:40
(gdb) set disassembly-flavor intel
(gdb) disas
Dump of assembler code for function crash_vector_ret(unsigned char*):
0x00000000004016b7 <+0>: push rbp
0x00000000004016b8 <+1>: mov rbp,rsp
0x00000000004016bb <+4>: sub rsp,0x10
0x00000000004016bf <+8>: mov QWORD PTR [rbp+0x10],rcx
0x00000000004016c3 <+12>: mov QWORD PTR [rbp+0x18],rdx
0x00000000004016c7 <+16>: mov rax,QWORD PTR [rbp+0x18]
0x00000000004016cb <+20>: mov QWORD PTR [rbp-0x8],rax
0x00000000004016cf <+24>: mov rax,QWORD PTR [rbp-0x8]
0x00000000004016d3 <+28>: vmovdqu xmm0,XMMWORD PTR [rax]
0x00000000004016d7 <+32>: vinserti128 ymm0,ymm0,XMMWORD PTR [rax+0x10],0x1
0x00000000004016de <+39>: vmovdqa ymm1,ymm0
0x00000000004016e2 <+43>: vmovdqa ymm0,ymm1
0x00000000004016e6 <+47>: mov rax,QWORD PTR [rbp+0x10]
=> 0x00000000004016ea <+51>: vmovdqa YMMWORD PTR [rax],ymm0
0x00000000004016ee <+55>: nop
0x00000000004016ef <+56>: mov rax,QWORD PTR [rbp+0x10]
0x00000000004016f3 <+60>: add rsp,0x10
0x00000000004016f7 <+64>: pop rbp
0x00000000004016f8 <+65>: ret
End of assembler dump.
(gdb) p $rax
$1 = 6749616
(gdb) p $rax/32*32
$2 = 6749600
As you can see, the $rax address is not aligned to 32 bytes and therefore the vmovdqa crashes.
I thought that because of the alignof(__m256i) == 32, the compiler should guarantee that also the return value is aligned to 32 bytes.
I might be wrong. Could you please shed more light on this issue?
Thanks for any hints.
--
Ticket information of MinGW - Minimalist GNU for Windows project
MinGW - Minimalist GNU for Windows Project is hosted on OSDN
Project URL: https://osdn.net/projects/mingw/
OSDN: https://osdn.net
URL for this Ticket:
https://osdn.net/projects/mingw/ticket/39565
RSS feed for this Ticket:
https://osdn.net/ticket/ticket_rss.php?group_id=3917&tid=39565