Hi. I've a reduced test case showing that a function like: struct x { uint8_t a; uint16_t b; }; f(struct x *y) ... will expect that 'y' is correctly aligned: it uses 'ldrh' to access y->b and ldrh documentation says "if the memory is not halfworld-aligned and no data abort occurs, the value written to the destination register is unpredictable." Test case: typedef unsigned char uint8_t; typedef unsigned short int uint16_t; struct x { uint16_t d; uint8_t c; }; int f(struct x *y) { diag_printf("y=%p y->c=%d(0x%x) y->d=%d(0x%x)\n\r",y,y->c,y->c,y->d,y->d); } int main(int argc,char **argv ) { void *s=malloc(sizeof(struct x)+1); struct x *y; s++; // <- problem origin y=s; y->c=0xAA; y->d=0xBBCC; f(y); while(1); } Without "s++;": y=0x000088e0 y->c=170(0xaa) y->d=48076(0xbbcc) With "s++;": y=0x000088e1 y->c=170(0xaa) y->d=-872415045(0xcc0000bb) Generated code with -Os is: int f(struct x *y) { 40000698: b500 push {lr} diag_printf("y=%p y->c=%d(0x%x) y->d=%d(0x%x)\n\r",y,y->c,y->c,y->d,y->d); 4000069a: 7883 ldrb r3, [r0, #2] 4000069c: 8802 ldrh r2, [r0, #0] <--- *HERE* 4000069e: b082 sub sp, #8 400006a0: 9200 str r2, [sp, #0] 400006a2: 9201 str r2, [sp, #4] 400006a4: 1c01 mov r1, r0 (add r1, r0, #0) 400006a6: 1c1a mov r2, r3 (add r2, r3, #0) 400006a8: 4802 ldr r0, [pc, #8] (400006b4 <.text+0x674>) 400006aa: ff4ff002 bl 4000354c <diag_printf> } 400006ae: b002 add sp, #8 400006b0: bc02 pop {r1} 400006b2: 4708 bx r1 400006b4: 98e4 ldr r0, [sp, #912] 400006b6: 4000 and r0, r0 Tested on arm-elf-gcc 3.4.4, 3.4.6, 4.1.1, the latest two with and without the patch related to bug 29373. Thanks for your help.
The compiler expects 'y' to be correctly aligned because the ABI says that it can expect this to be so. If you write a program that then provides an unaligned value, then the behaviour is undefined and you can get any result the compiler feels like giving you. Hint: don't increment void* pointers (it's a gcc extension that permits this anyway) and then try to assign them to a structure. Malloc provides memory that is correctly aligned for direct assignment to any structure type.
Subject: Re: [3.4.4/3.4.6/4.1.1] ARM structure pointer alignment problem with optimization rearnsha at gcc dot gnu dot org wrote: > ------- Comment #1 from rearnsha at gcc dot gnu dot org 2006-11-22 10:50 ------- > The compiler expects 'y' to be correctly aligned because the ABI says that it > can expect this to be so. If you write a program that then provides an > unaligned value, then the behaviour is undefined and you can get any result the > compiler feels like giving you. > > Hint: don't increment void* pointers (it's a gcc extension that permits this > anyway) and then try to assign them to a structure. Malloc provides memory > that is correctly aligned for direct assignment to any structure type. > > > Sorry for having taken your time for nothing. I'll read the ABI spec better next time. Too bad my target does not have alignment traps, I've fallen in the problem while building a packet with an odd size header. Regards, Bernard