| Summary: | [3.3/3.4 regression] selection of %dil or %sil on ia32 by valid C source | ||
|---|---|---|---|
| Product: | gcc | Reporter: | etienne.lorrain |
| Component: | inline-asm | Assignee: | Not yet assigned to anyone <unassigned> |
| Status: | RESOLVED INVALID | ||
| Severity: | normal | CC: | adam, aoliva, dacut, etienne.lorrain, gcc-bugs, nkoch, normvcr, richardpku, xnavara |
| Priority: | P2 | Keywords: | wrong-code |
| Version: | 3.2 | ||
| Target Milestone: | 3.3.1 | ||
| Host: | Target: | ||
| Build: | Known to work: | ||
| Known to fail: | Last reconfirmed: | 2003-05-28 04:48:23 | |
From: Dara Hazeghi <dhazeghi@yahoo.com> To: gcc-gnats@gcc.gnu.org, etienne.lorrain@masroudeau.com Cc: Subject: Re: inline-asm/10153: selection of %dil or %sil on ia32 arch by valid C source Date: Sat, 10 May 2003 13:51:39 -0700 http://gcc.gnu.org/cgi-bin/gnatsweb.pl?cmd=view%20audit- trail&database=gcc&pr=10153 Hello, on gcc 3.2.3 and 3.3 branch, I get the same error as mentioned in this report, but on mainline (20030509), this code compiles and assembles fine. Dara From: etienne.lorrain@masroudeau.com To: "Dara Hazeghi" <dhazeghi@yahoo.com> Cc: gcc-gnats@gcc.gnu.org,etienne.lorrain@masroudeau.com Subject: Re: inline-asm/10153: selection of %dil or %sil on ia32 arch by valid C source Date: Mon, 12 May 2003 06:59:00 +0100 (BST) > http://gcc.gnu.org/cgi-bin/gnatsweb.pl?cmd=view%20audit-trail&database=gcc&pr=10153 > > Hello, > > on gcc 3.2.3 and 3.3 branch, I get the same error as mentioned in this > report, but on mainline (20030509), this code compiles and assembles > fine. > > Dara > Hello, I did this morning, in an empty directory: lynx http://prdownloads.sourceforge.net/gujin/gujin-0.7.tgz?download ncftpget ftp://sources.redhat.com/pub/binutils/releases/binutils-2.13.2.1.tar.gz ncftpget ftp://sources.redhat.com/pub/gcc/snapshots/gcc-3.3-20030508.tar.gz tar -xvzf gujin-0.7.tgz ------------------------------------------- Here I had to modify gujin/Makefile : @@ -1564,7 +1564,7 @@ cd $(notdir $(basename $(basename $(subst core-,,$(LAST_GCC))))) ; \ mkdir build ; \ cd build/ ; \ - ../configure --prefix=$(P_PWD)/toolchain --enable-languages=C ; \ + ../configure --prefix=$(P_PWD)/toolchain --enable-languages=c ; \ make bootstrap-lean ; \ make install ) endif ------------------- Is that a philosophycal change, no more "C" language, only "c" ???? ------------------------------------------- cd gujin make toolchain make dep boot.exe (all of this is shown without their output - quite long) At some point I got: .... GCC: gcc (GCC) 3.3 20030509 (prerelease), Binutils: GNU ld version 2.13.2.1 .... /home/etienne/prjgcc/toolchain/bin/gcc -fstrict-aliasing -falign-loops=1 -falign-jumps=1 -falign-functions=2 -mno-align-double -mpreferred-stack-boundary=2 -fomit-frame-pointer -march=i386 -mrtd -fno-builtin -fno-schedule-insns2 -finline-limit=10000 -Os -Wall -Wno-format -Wno-main -Wstrict-prototypes -Wundef -Wpointer-arith -Wcast-align -Wwrite-strings -Wsign-compare -Wdisabled-optimization -ffunction-sections -include code16.h -DDEBUG="(DEBUG_STACK)" -DUSER_SUPPORT="(VGA_SUPPORT|VGA_EXTENDED|VESA_WINDOW|VESA_WINFUNCTION|VESA_LINEAR|VESA_EDID|VESA_RECALC|VESA_2WINDOWS|VESA_HARDWINDOW|VESA_PMINFO|VESA_16BPP|VESA_32BPP|VESA_24BPP|VESA_4BPP_EGA|VESA_4BPP_TEXT|VESA_8BPP|SERIAL_VT100|SERIAL_VT420|BIOS_MOUSE_SUPPORT|SERIAL_MOUSE_SUPPORT|JOYSTICK_SUPPORT)" -DDISK_SUPPORT="(BIOS_SUPPORT|EBIOS_SUPPORT|IDE_SUPPORT|DOS_SUPPORT|E2FS_PROBE|FAT12_PROBE|FAT16_PROBE|FAT32_PROBE|BOOTSECT_PROBE)" -DSETUP="(APPLICATION|CODE_SEGMENT|EXTRA_SEGMENT|GC_SECTION|ASSEMBLY_CISC|UPDATE_BOOTPARAM|USE_INT1587|MINIDEBUG)" -DLANGUAGE=ENGLISH -S -o user.S user.c In file included from user.c:43: ega.h: In function `_EGA_read_write_mode': ega.h:172: warning: dereferencing type-punned pointer will break strict-aliasing rules ega.h: In function `_EGA_setup': ega.h:178: warning: dereferencing type-punned pointer will break strict-aliasing rules sed '/.align 32/d' < user.S > user.s /home/etienne/prjgcc/toolchain/bin/as -acdhls=user.lis -o user.o user.s user.s: Assembler messages: user.s:2962: Error: bad register name `%dil' make: *** [user.o] Error 1 rm user.S boot.s user.s boot.S So the bug is still present in this version. I do not know about my "reduced to the simplest source" bug report. I have to check about new warning "dereferencing type-punned". Regards, Etienne. I can see the problem with the mainline (20030527) with `gcc -Os pr10153.i -march=i386 -c -save-temps', `gcc -Os pr10153.i -march=i486 -c -save-temps', `gcc -Os pr10153.i - march=i586 -c -save-temps' I think Dara was getting lucky with i686. I think sil and dil are both x86-64 registers which are not defined on ia32. I think you were luck that gcc-3.0.4 works, but still it is a regression. Since the "g" constrant is right here, but someone with more knowledge of the ia32 backend should look at it. I'm of the opinion that the source is invalid. Using a "=q" constraint instead of an "=r" constraint properly shows that the compiler must choose one of the four "low byte" registers. You might object that the compiler should "know" that it cannot place a character object into %esi/%edi/%ebp, but that's not true at all -- we cannot store a character to memory from those registers, but we can hold a character value in those registers. *** Bug 12856 has been marked as a duplicate of this bug. *** Richard -- agreed with your comment, but I think the larger issue is that GCC is choosing phantom registers (%sil/%dil). Yonetani Tomokazu on the DragonFly BSD lists has seen this issue compiling straight C code with "-march=pentium4 -O". Any ideas where to look? I'll try to get a sample of failing source (along with GCC versions and the other goodies) and see where it's dying. After some further discussion on the DragonFly lists and further reading on my part, I found out that %sil and %dil are the 8-bit equivalents of %al, %bl, etc. (Watch the light bulb undim in my head here :-) It would be better if gcc were able to diagnose the problem code better (ensure that the architecture is valid before emitting a %sil/%dil/%bpl/%spl register) -- I'll see if I can't ponder on this and propose a small patch. *** Bug 19979 has been marked as a duplicate of this bug. *** *** Bug 29579 has been marked as a duplicate of this bug. *** *** Bug 33674 has been marked as a duplicate of this bug. *** *** Bug 48347 has been marked as a duplicate of this bug. *** The master branch has been updated by Andrew Pinski <pinskia@gcc.gnu.org>: https://gcc.gnu.org/g:8819419ba1d397c0444d89079ec16657a09914fb commit r12-2473-g8819419ba1d397c0444d89079ec16657a09914fb Author: Andrew Pinski <apinski@marvell.com> Date: Tue Jul 20 11:25:43 2021 -0700 Fix PR 10153: tail recusion for vector types. The problem here is we try to an initialized value from a scalar constant. For vectors we need to do a vect_dup instead. This fixes that issue by using build_{one,zero}_cst instead of integer_{one,zero}_node when calling create_tailcall_accumulator. Changes from v1: * v2: Use build_{one,zero}_cst and get the correct type before. OK? Bootstrapped and tested on aarch64-linux-gnu with no regressions. gcc/ChangeLog: PR tree-optimization/10153 * tree-tailcall.c (create_tailcall_accumulator): Don't call fold_convert as the type should be correct already. (tree_optimize_tail_calls_1): Use build_{one,zero}_cst instead of integer_{one,zero}_node for the call of create_tailcall_accumulator. gcc/testsuite/ChangeLog: PR tree-optimization/10153 * gcc.c-torture/compile/pr10153-1.c: New test. * gcc.c-torture/compile/pr10153-2.c: New test. FTR, the fix and the testcases were for bug 101534. |
Just compile this, and look at the comment "automatic size decision there" (lots of comment removed and macro expanded...) Works up to gcc-3.0.4, not after 3.1 etienne@fulbert:~/gcc/gujin$ gcc -v Reading specs from /usr/lib/gcc-lib/i386-linux/3.2.3/specs Configured with: ../src/configure -v --enable-languages=c,c++,java,f77,proto,pascal,objc,ada --prefix=/usr --mandir=/usr/share/man --infodir=/usr/share/info --with-gxx-include-dir=/usr/include/c++/3.2 --enable-shared --with-system-zlib --enable-nls --without-included-gettext --enable-__cxa_atexit --enable-clocale=gnu --enable-java-gc=boehm --enable-objc-gc i386-linux Thread model: posix gcc version 3.2.3 20030309 (Debian prerelease) etienne@fulbert:~/gcc/gujin$ etienne@fulbert:~/gcc/gujin$ cat user.i extern unsigned char VGA_valid_mode[]; extern inline unsigned divrem8 (unsigned short oper, unsigned char *rem) { unsigned divider; asm ( " aam $8 # operand in ax \n" " movb %%al,%%cl # remainder \n" " movzbl %%ah,%%eax # divider \n" : "=a" (divider), "=c" (*rem) : "a" (oper) ); return divider; } extern inline void VIDEO_mode_invalidate (unsigned char mode) { unsigned char shift; unsigned index = divrem8 (mode, &shift); unsigned char tmp = ({ typeof (VGA_valid_mode[index]) returned; asm volatile ( " # automatic size decision there: \n" " mov %%cs:%a1,%0 \n" : "=r" (returned) : "g" (&VGA_valid_mode[index])); returned; }) & ~(1 << shift); asm volatile ( " movb %b0,%%cs:%a1 \n" : : "qi" (tmp), "r" (&VGA_valid_mode[index]) ); } void VGA_init (void) { extern char tmpstatictable[15]; unsigned i; for (i = 0; i <= 0x7F; i++) { unsigned char shift; unsigned short offset = divrem8 (i, &shift); if ((tmpstatictable[offset] & (1 << shift)) == 0) VIDEO_mode_invalidate (i); } } etienne@fulbert:~/gcc/gujin$ gcc -Os user.i /tmp/ccWPbdd5.s: Assembler messages: /tmp/ccWPbdd5.s:35: Error: bad register name `%sil' Release: gcc version 3.2.3 20030309 (Debian prerelease) Environment: Linux Debian ia32