From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (qmail 12288 invoked by alias); 13 Aug 2002 06:06:01 -0000 Mailing-List: contact gcc-prs-help@gcc.gnu.org; run by ezmlm Precedence: bulk List-Archive: List-Post: List-Help: Sender: gcc-prs-owner@gcc.gnu.org Received: (qmail 12274 invoked by uid 71); 13 Aug 2002 06:06:00 -0000 Date: Tue, 13 Aug 2002 03:58:00 -0000 Message-ID: <20020813060600.12273.qmail@sources.redhat.com> To: nobody@gcc.gnu.org Cc: gcc-prs@gcc.gnu.org, From: Tim Prince Subject: Re: c++/7582: Intel intrinsics cause segfault with gcc 3.1.1 and 3.2 Reply-To: Tim Prince X-SW-Source: 2002-08/txt/msg00240.txt.bz2 List-Id: The following reply was made to PR c++/7582; it has been noted by GNATS. From: Tim Prince To: dholm@telia.com, gcc-gnats@gcc.gnu.org Cc: Subject: Re: c++/7582: Intel intrinsics cause segfault with gcc 3.1.1 and 3.2 Date: Mon, 12 Aug 2002 22:52:16 -0700 On Monday 12 August 2002 16:07, dholm@telia.com wrote: > >Number: 7582 > >Category: c++ > >Synopsis: Intel intrinsics cause segfault with gcc 3.1.1 and 3.2 > >Confidential: no > >Severity: critical > >Priority: medium > >Responsible: unassigned > >State: open > >Class: sw-bug > >Submitter-Id: net > >Arrival-Date: Mon Aug 12 16:16:01 PDT 2002 > >Closed-Date: > >Last-Modified: > >Originator: David Holm > >Release: gcc version 3.2 2002-07-26 (prerelease) > >Organization: > >Environment: > > Gentoo Linux 1.4, Pentium 3 (Coppermine) > > >Description: > > The following code executes perfectly when compiled with the Intel C++ > Compiler v6.0 but segfaults when compiled with gcc 3.1.1 or 3.2 > (2002-07-26). It's compiled with "g++ (-g3) -Wall -msse intrin.cpp -o > intrin" and runs without any output. g++ gives no warnings during > compilation. > intrin segfaults on this line "_mm_stream_ps((float*) dst, xmm0);" > > "g++ -v" returns: > Reading specs from /usr/lib/gcc-lib/i686-pc-linux-gnu/3.2/specs > Configured with: /var/tmp/portage/gcc-3.2_pre/work/gcc-3.2/configure > --prefix=/usr --mandir=/usr/share/man --infodir=/usr/share/info > --enable-shared --host=i686-pc-linux-gnu --build=i686-pc-linux-gnu > --target=i686-pc-linux-gnu --with-system-zlib > --enable-languages=c,c++,ada,f77,objc,java --enable-threads=posix > --enable-long-long --disable-checking --enable-cstdio=stdio > --enable-clocale=generic --enable-version-specific-runtime-libs > --with-gxx-include-dir=/usr/include/g++-v32 --with-local-prefix=/usr/local > --enable-shared --enable-nls --without-included-gettext Thread model: posix > gcc version 3.2 2002-07-26 (prerelease) > > I haven't got 3.1.1 anymore, so I can't give you the -v output from it. > > >How-To-Repeat: > > #include > #include > #include > #include > #include > > #define small_memcpy(dst,src,n) \ > register unsigned long int dummy; \ > asm volatile ( \ > "rep; movsb\n\t" \ > > :"=&D"(dst), "=&S"(src), "=&c"(dummy) \ > : > :"0" (dst), "1" (src),"2" (n) \ > : > : "memory"); > > /** > * SIMD Optimized memcpy's are graciously borrowed from DirectFB. > */ > > # define SSE_MMREG_SIZE 16 > # define MIN_LEN 0x40 /* 64-byte blocks */ > > void *memcpy_sse( void *dst, const void *src, size_t len ) > { > void *retval = dst; > size_t i; > > _mm_prefetch((char*) src, _MM_HINT_NTA); > _mm_prefetch((char*) src + 64, _MM_HINT_NTA); > _mm_prefetch((char*) src + 128, _MM_HINT_NTA); > _mm_prefetch((char*) src + 192, _MM_HINT_NTA); > _mm_prefetch((char*) src + 256, _MM_HINT_NTA); > > if (len >= MIN_LEN) > { > register unsigned long int delta; > delta = ((unsigned long int) dst) & (SSE_MMREG_SIZE - 1); > if (delta) > { > delta = SSE_MMREG_SIZE - delta; > len -= delta; > small_memcpy(dst, src, delta); > } > i = len >> 6; > len &= 63; > > if (((unsigned long) src) & 15) > for (; i > 0; i--) > { > __m128 xmm0, xmm1, xmm2, xmm3; > _mm_prefetch((char*) src + 320, > _MM_HINT_NTA); xmm0 = _mm_loadu_ps((float*) src); xmm1 = > _mm_loadu_ps((float*) src + 4); xmm2 = _mm_loadu_ps((float*) src + 8); xmm3 > = _mm_loadu_ps((float*) src + 12); _mm_stream_ps((float*) dst, xmm0); > _mm_stream_ps((float*) dst + 4, xmm1); _mm_stream_ps((float*) dst + 8, > xmm2); _mm_stream_ps((float*) dst + 12, xmm3); #ifdef __GNUC__ > (char*) src += 64; > (char*) dst += 64; > #else > src += 64; > dst += 64; > #endif > } > else > for (; i > 0; i--) > { > __m128 xmm0, xmm1, xmm2, xmm3; > _mm_prefetch((char*) src + 320, > _MM_HINT_NTA); xmm0 = _mm_load_ps((float*) src); > xmm1 = _mm_load_ps((float*) src + 4); > xmm2 = _mm_load_ps((float*) src + 8); > xmm3 = _mm_load_ps((float*) src + 12); > _mm_stream_ps((float*) dst, xmm0); > _mm_stream_ps((float*) dst + 4, xmm1); > _mm_stream_ps((float*) dst + 8, xmm2); > _mm_stream_ps((float*) dst + 12, xmm3); > #ifdef __GNUC__ > (char*) src += 64; > (char*) dst += 64; > #else > src += 64; > dst += 64; > #endif > } > } > > if (len) > memcpy(dst, src, len); > > return retval; > } > > int main(void) > { > char *tmp1, *tmp2; > > (void*) tmp1 = malloc(1024 * 1024 * 10); > (void*) tmp2 = malloc(1024 * 1024 * 10); > > memcpy_sse(tmp1, tmp2, 1024 * 1024 * 10); > > free(tmp1); > free(tmp2); > > return 0; > } > > >Fix: > > > >Release-Note: > >Audit-Trail: > >Unformatted: You start off with no attempt to get aligned storage for tmp1 and tmp2. In an ideal world, malloc would take care of this, but gcc doesn't take responsibility for which malloc you use. I take it you are using whatever glibc gives you. Mine gives me 8-byte alignment, but not the required 16-byte alignment. Since you didn't mention whether you stepped into your code with your favorite debugger to check for such problems, some of us may assume you haven't begun to do your homework. If you did use a strategy to assure alignment, you haven't informed us what it might be. You've gone out of your way to obscure your code, yet you ignore what seems most evident. As I understand it, the preference for the Intel compiler would be to use the special aligned entry point _mm_malloc(), in order to make your code portable to Windows, so you are lucky it works with icc. -- Tim Prince