Platform info: Red Hat Enterprise Linux WS release 3 (Taroon) Linux legless 2.4.21-4.ELsmp #1 SMP Fri Oct 3 17:52:56 EDT 2003 i686 i686 i386 GNU/Linux g++ (GCC) 3.2.3 20030502 (Red Hat Linux 3.2.3-20) gcc SVN snapshot: 2006-10-07 10:14 PDT Commands used: ../gcc_trunk/configure --prefix=/some/path --enable-languages=c,c++ make bootstrap End of make bootstrap output: Comparing stages 2 and 3 warning: ./cc1-checksum.o differs warning: ./cc1plus-checksum.o differs Bootstrap comparison failure! ./cfg.o differs ./cfgloopanal.o differs ./loop-iv.o differs ./predict.o differs ./profile.o differs ./value-prof.o differs ./ipa-inline.o differs make[2]: *** [compare] Error 1 make[2]: Leaving directory `/net/legless/scratch1/rwgk/gcc_build' make[1]: *** [stage3-bubble] Error 2 make[1]: Leaving directory `/net/legless/scratch1/rwgk/gcc_build' make: *** [bootstrap] Error 2 Comments: I had the same problem with the SVN snapshot from 2006-09-21 22:03 PDT. Both SVN snapshots work without a problem under Fedora Core 5 x86_64.
This works for me on i686-linux-gnu. Can you first try compiling 3.4.x and then trying compiling the mainline with that?
No feedback in 3 months so closing.
[tru@magneto ~]$ gcc -v Reading specs from /usr/lib/gcc-lib/i386-redhat-linux/3.2.3/specs Configured with: ../configure --prefix=/usr --mandir=/usr/share/man --infodir=/usr/share/info --enable-shared --enable-threads=posix --disable-checking --with-system-zlib --enable-__cxa_atexit --host=i386-redhat-linux Thread model: posix gcc version 3.2.3 20030502 (Red Hat Linux 3.2.3-58) [tru@magneto ~]$ rpm -q gcc gcc-3.2.3-58.i386 is failing too, but a plain gcc-3.4.6 can bootstrap gcc-4.2.0
When using gcc-3.2.3 as bootstrap compiler, i386.c is miscompiled with -O0 -fkeep-inline-functions (the latter option is what is new in gcc 4.2 and why 4.1.x bootstrap didn't suffer from this, see 2006-07-04 Eric Botcazou <ebotcazou@libertysurf.fr> PR bootstrap/18058 * configure.in: Add -fkeep-inline-functions to CFLAGS for stage 1 if the bootstrap compiler is a GCC version that supports it. * configure: Regenerate. ). typedef struct rtx_def *rtx; enum machine_mode { VOIDmode = 0, DImode = 13 }; extern rtx gen_rtx_CONST_INT (enum machine_mode, long); extern rtx gen_x86_shld_1 (rtx, rtx, rtx); static __inline__ rtx gen_x86_64_shld(rtx a, rtx b, rtx c) { return 0; } extern rtx emit_insn (rtx); void foo (rtx *high, rtx *low, enum machine_mode mode, int count) { emit_insn ((mode == DImode ? gen_x86_shld_1 : gen_x86_64_shld) (high[0], low[0], gen_rtx_CONST_INT (VOIDmode, count))); } (extracted from i386.i) is miscompiled (tested latest RHEL3 gcc-3.2.3 as well as current branches/gcc-3_2-branch), it calls gen_x86_64_shld unconditionally. Removing -fkeep-inline-functions (or adding asm volatile ("") into the inline function or making gen_x86_64_shld extern instead of static inline cures this). Now, gcc-3_2-branch is long time closed, so IMNSHO gcc-4.2+ should work around this bug.
> Now, gcc-3_2-branch is long time closed, so IMNSHO gcc-4.2+ should work around > this bug. Agreed, we can simply say that GCC 3.2.x is not "is a GCC version that supports it". Would you mind writing the patch? I don't have GCC 3.2.x handy. TIA.
Seems there were 2 separate bugs that are causing this miscompilation. 1) common_type (in contemporary gcc's common_pointer_type) will for the type of the whole conditional expression use pointer to attribute const function rather than non-const: int fn1 (void); int fn2 (void) __attribute__((const)); ... (cond ? fn1 : fn2) I'd say that's an unfortunate effect from representing __attribute__((const)) as TREE_READONLY on the FUNCTION_TYPE rather than as an attribute. Say for cond expr where one ptr target is volatile and the other is not the result needs to be ptr to volatile and similarly for say int *ptr1; const int *ptr2; cond ? ptr1 : ptr2 where the result should be pointer to const, a conservative choice. But for __attribute__((const)) functions the conservative choise is the exact opposite, if I have (cond ? fn1 : fn2) one of the functions isn't const and therefore nothing should assume the whole thing is pointer to __attribute__((const)) function. gcc trunk handles this the same way. 2) because of 1), but also e.g. on typedef struct rtx_def *rtx; enum machine_mode { VOIDmode = 0, DImode = 13 }; extern rtx gen_rtx_CONST_INT (enum machine_mode, long); extern rtx gen_x86_shld_1 (rtx, rtx, rtx) __attribute__((const)); extern rtx gen_x86_64_shld (rtx, rtx, rtx) __attribute__((const)); extern rtx emit_insn (rtx); void foo (rtx *high, rtx *low, enum machine_mode mode, int count) { emit_insn ((mode == DImode ? gen_x86_shld_1 : gen_x86_64_shld) (high[0], low[0], gen_rtx_CONST_INT (VOIDmode, count))); } where 1) doesn't apply a properly emitted expanded COND_EXPR into rtl is passed through emit_libcall_block which changes: (insn 31 0 32 (set (reg:CCZ 17 flags) (compare:CCZ (mem/f:SI (plus:DI (reg/f:DI 54 virtual-stack-vars) (const_int -20 [0xffffffffffffffec])) [0 mode+0 S4 A32]) (const_int 13 [0xd]))) -1 (nil) (nil)) (jump_insn 32 31 34 (set (pc) (if_then_else (ne (reg:CCZ 17 flags) (const_int 0 [0x0])) (label_ref 37) (pc))) -1 (nil) (nil)) (insn 34 32 35 (set (reg:DI 63) (symbol_ref:DI ("gen_x86_shld_1"))) -1 (nil) (nil)) (jump_insn 35 34 36 (set (pc) (label_ref 40)) -1 (nil) (nil)) (barrier 36 35 37) (code_label 37 36 39 2 "" "" [0 uses]) (insn 39 37 40 (set (reg:DI 63) (symbol_ref:DI ("gen_x86_64_shld"))) -1 (nil) (nil)) (code_label 40 39 42 3 "" "" [0 uses]) (insn 42 40 44 (set (reg:DI 1 rdx) (reg:DI 60)) -1 (nil) (nil)) (insn 44 42 46 (set (reg:DI 4 rsi) (mem:DI (reg/f:DI 61) [0 S8 A64])) -1 (nil) (nil)) (insn 46 44 47 (set (reg:DI 5 rdi) (mem:DI (reg/f:DI 62) [0 S8 A64])) -1 (nil) (nil)) (call_insn/u 47 46 0 (set (reg:DI 0 rax) (call (mem:QI (reg:DI 63) [0 S1 A8]) (const_int 0 [0x0]))) -1 (nil) (nil) (expr_list (use (reg:DI 5 rdi)) (expr_list (use (reg:DI 4 rsi)) (expr_list (use (reg:DI 1 rdx)) (nil))))) into (insn 34 30 39 (set (reg:DI 63) (symbol_ref:DI ("gen_x86_shld_1"))) -1 (nil) (nil)) (insn 39 34 31 (set (reg:DI 63) (symbol_ref:DI ("gen_x86_64_shld"))) -1 (nil) (nil)) (insn 31 39 32 (set (reg:CCZ 17 flags) (compare:CCZ (mem/f:SI (plus:DI (reg/f:DI 54 virtual-stack-vars) (const_int -20 [0xffffffffffffffec])) [0 mode+0 S4 A32]) (const_int 13 [0xd]))) -1 (nil) (nil)) (jump_insn 32 31 35 (set (pc) (if_then_else (ne (reg:CCZ 17 flags) (const_int 0 [0x0])) (label_ref 37) (pc))) -1 (nil) (nil)) (jump_insn 35 32 36 (set (pc) (label_ref 40)) -1 (nil) (nil)) (insn 34 30 39 (set (reg:DI 63) (symbol_ref:DI ("gen_x86_shld_1"))) -1 (nil) (nil)) (insn 39 34 31 (set (reg:DI 63) (symbol_ref:DI ("gen_x86_64_shld"))) -1 (nil) (nil)) (insn 31 39 32 (set (reg:CCZ 17 flags) (compare:CCZ (mem/f:SI (plus:DI (reg/f:DI 54 virtual-stack-vars) (const_int -20 [0xffffffffffffffec])) [0 mode+0 S4 A32]) (const_int 13 [0xd]))) -1 (nil) (nil)) (jump_insn 32 31 35 (set (pc) (if_then_else (ne (reg:CCZ 17 flags) (const_int 0 [0x0])) (label_ref 37) (pc))) -1 (nil) (nil)) (jump_insn 35 32 36 (set (pc) (label_ref 40)) -1 (nil) (nil)) (barrier 36 35 37) (code_label 37 36 40 2 "" "" [0 uses]) (code_label 40 37 42 3 "" "" [0 uses]) (insn 42 40 44 (set (reg:DI 1 rdx) (reg:DI 60)) -1 (nil) (nil)) (insn 44 42 46 (set (reg:DI 4 rsi) (mem:DI (reg/f:DI 61) [0 S8 A64])) -1 (nil) (nil)) (insn 46 44 47 (set (reg:DI 5 rdi) (mem:DI (reg/f:DI 62) [0 S8 A64])) -1 (nil) (nil)) (call_insn/u 47 46 49 (set (reg:DI 0 rax) (call (mem:QI (reg:DI 63) [0 S1 A8]) (const_int 0 [0x0]))) -1 (nil) (expr_list:REG_EH_REGION (const_int -1 [0xffffffffffffffff]) (nil)) (expr_list (use (reg:DI 5 rdi)) (expr_list (use (reg:DI 4 rsi)) (expr_list (use (reg:DI 1 rdx)) (nil))))) Haven't checked yet what exactly fixed this on 3.4 (or already 3.3) branch.
2) is apparently PR11557, fixed in GCC 3.3.1+. So, I'd say as workaround we should not use -fkeep-inline-functions for GCC < 3.3.1. Testing a patch for that.
Subject: Bug 29382 Author: jakub Date: Wed May 30 13:32:34 2007 New Revision: 125182 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=125182 Log: PR bootstrap/29382 * configure.in: Don't use -fkeep-inline-functions for GCC < 3.3.1. * configure: Rebuilt. Modified: trunk/ChangeLog trunk/configure trunk/configure.ac
Subject: Bug 29382 Author: jakub Date: Wed May 30 13:48:07 2007 New Revision: 125184 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=125184 Log: PR bootstrap/29382 * configure.in: Don't use -fkeep-inline-functions for GCC < 3.3.1. * configure: Rebuilt. Modified: branches/gcc-4_2-branch/ChangeLog branches/gcc-4_2-branch/configure branches/gcc-4_2-branch/configure.in
Worked around in 4.2.1+ and on the trunk.