On Fri, 18 Aug 2023 14:17:07 -0700 Sam Edwards cfsworks@gmail.com wrote:
On 8/18/23 10:40, Sam Edwards wrote:
On 8/18/23 07:11, Andre Przywara wrote:
Hi Andre,
The resulting object file is different (8 byte larger, even), so it's hard to prove
I'm no stranger to reading object code. Since the output should be identical in principle, I'll spend a little bit of time today seeing if I can identify what's changing. If it's easy enough, I'd like to adjust my patch so that the optimizer does produce the same output. (Keep in mind I'm on Clang, though. If Clang already gives the same output for both, I'll just report back to use that when comparing.)
I built only psci.o from every ARM32 sunxi for which we have a defconfig (and for which PSCI is supported), for 81 targets total (though there are only 4 variations: R40, sun7i, H3/sun6i, and "everything else"). I am working with Clang version 16.0.6.
I compared only the secure text section. The command to extract this looks like: llvm-objcopy -O binary --only-section=._secure.text psci.o text.bin This is important because there are debug sections that will change when the source file line numbers change, so we must ignore those when comparing.
In the majority of cases, there are no changes to the text section introduced by this patch. In the R40 case, there's a small change where the compiler adds a NULL check onto the result of the `(void *)cpucfg + SUN8I_R40_PWR_CLAMP(cpu)` computation, which we can ignore as it won't affect anything in practice. In the sun7i case, the only changes are because I am NOT hardcoding the CPU to 0, which does look like I broke it (since that means it will use cpu=1). So I'm going to need to fix that in v3.
^^^^^^^ Do you have an update on this? I will try to test it on an R40 ASAP.
Cheers, Andre
For good measure, I also applied the same methodology to patch 2 in this series, and that introduces no text section changes whatsoever in any of the tested cases. So patch 2 (theoretically, anyway) needs no bugfixes or hardware testing.
Patch 3 does cause a text section change for all targets. I will have to investigate why, in case I messed up any of the offsets when migrating off of structs.
Regards, Sam