On Thu, Mar 19, 2015 at 7:21 PM, Julius Werner jwerner@chromium.org wrote:
That said, I went back and looked at the assembly dump. It was using 0x14 as the size of the structure when sequencing through the entries.
001465dc R _bs_init_begin 001465e0 r cbmem_bscb 00146600 r gcov_bscb 0014663c R _bs_init_end
Each *entry* was aligned to 0x20. Just aligning the symbol wouldn't have fixed it. That means we'd need to mark the structure declaration as aligned to a specific value and annotate it accordingly in the linker script. That's one route or just go w/ pointers to the structures.
Yeah, that's what I was talking about in the first mail... I think this means that the compiler would consider the sizeof() for each structure to be 0x20. I'm pretty sure there's some kind of rule that the sizeof() for any object must be evenly divisible by its minimum alignment... that's how normal arrays work as well (e.g. if you had struct mystruct { uint32_t a; uint8_t b; } and don't mark it __packed__, then sizeof(struct mystruct) is 0x8 and an array of them would have 3 bytes padding after each entry). So I'm pretty sure the pointer++ mechanism to walk through it would've still worked if the initial address had been right. (can we repro this? then we could simply try it out instead of guessworking.)
I think you misunderstood what I wrote; I wasn't guessworking.
The bscb structure arrays were aligned to 0x20. But the structure size to the compiler is 0x14. It's 0x14 in gdb and it's 0x14 in the instructions calculating the next entry. That's how I diagnosed this to begin with. From 4.92/ramstage.debug:
145991: bb dc 65 14 00 mov $0x1465dc,%ebx ... 145a36: 81 fb 3c 66 14 00 cmp $0x14663c,%ebx 145a3c: 74 5f je 145a9d <main+0x11c> ... 145a8a: 83 c3 14 add $0x14,%ebx ... 145a9b: eb 99 jmp 145a36 <main+0xb5>
So we know the first entry in the array is wrong based on _bs_init_begin != cbmem_bscb. However, ignoring that the entry after cbmem_bscb is wrong too because the 0x14 increment will not result in resolving to gcov_bscb.
-Aaron