Compiling the code below with -O2 generates, at the start of the function: sub sp, sp, #8 mov r2, #2032 ldrh r3, [sp, #6] @ movhi add r2, r2, #15 and r3, r2, r3, lsr #4 cmp r3, r2 stmia sp, {r0-r1} As far as I can see, this allocates eight bytes of stack, loads from it (ldrh), messes around a bit and then stores the value it was expecting to load (stmia). Obviously the stmia shouldn't have been moved to after the ldrh. This problem also exists in GCC 3.0.1 and the NetBSD version of GCC 2.95.3 (where I first found it). Release: 3.1 20011023 (experimental) Environment: System: Linux wraith 2.4.12 #13 Mon Oct 29 11:21:47 GMT 2001 i686 unknown Architecture: i686 host: i686-pc-linux-gnu build: i686-pc-linux-gnu target: arm-unknown-elf configured with: ./configure --target=arm-unknown-elf How-To-Repeat: use "./cc1 -O2" typedef unsigned int u_int; struct ieee_double { u_int dbl_fracl; u_int dbl_frach:20; u_int dbl_exp:11; u_int dbl_sign:1; }; int isinf (d) double d; { register struct ieee_double *p = (struct ieee_double *)(void *)&d; return (p->dbl_exp == 2047 && (p->dbl_frach == 0 && p->dbl_fracl == 0)); }
Fix: -fno-schedule-insns -fno-schedule-insns2 causes the problem to go away.
From: Richard Earnshaw <rearnsha@arm.com> To: bjh21@netbsd.org Cc: gcc-gnats@gcc.gnu.org, Richard.Earnshaw@arm.com Subject: Re: optimization/4733: Incorrect code when aliasing double and struct on ARM -O2 Date: Mon, 29 Oct 2001 17:43:44 +0000 > >How-To-Repeat: > use "./cc1 -O2" > > typedef unsigned int u_int; > struct ieee_double { > u_int dbl_fracl; > u_int dbl_frach:20; > u_int dbl_exp:11; > u_int dbl_sign:1; > }; > int > isinf (d) > double d; > { > register struct ieee_double *p = (struct ieee_double *)(void *)&d; > > return (p->dbl_exp == 2047 && > (p->dbl_frach == 0 && p->dbl_fracl == 0)); > } Nope, this code violates the ANSI memory aliasing rules. According to ANSI, the only way to do this portably is to memcpy between the types; gcc will do the right thing if you use a union, but not otherwise. R.
From: Ben Harris <bjh21@netbsd.org> To: <Richard.Earnshaw@arm.com> Cc: <gcc-gnats@gcc.gnu.org> Subject: Re: optimization/4733: Incorrect code when aliasing double and struct on ARM -O2 Date: Mon, 29 Oct 2001 18:08:14 +0000 (GMT) On Mon, 29 Oct 2001, Richard Earnshaw wrote: > Nope, this code violates the ANSI memory aliasing rules. Ah yes, I'd tried -fno-strict-aliasing on 2.95.3, but not on 20011023. Thanks. > According to > ANSI, the only way to do this portably is to memcpy between the types; Yeah, that works. The code it produces isn't all that wonderful, though: ... mov r3, sp stmia r3, {r0-r1} ldmia r3, {r0, r1} add r2, sp, #8 stmia r2, {r0, r1} mov r2, #2032 ldrh r3, [sp, #14] @ movhi ... -- Ben Harris <bjh21@netbsd.org> Portmaster, NetBSD/arm26 <URL:http://www.netbsd.org/Ports/arm26/>
State-Changed-From-To: open->closed State-Changed-Why: User error. The compiler is allowed to assume that the struct and the double aren't aliased. Use -fno-strict-aliasing if you want to compile this code with gcc.
toggling to mark as duplicate of 4410
oops
...
.
Reopening to ...
Mark as a dup of bug 21920. *** This bug has been marked as a duplicate of 21920 ***