Dear Pavel Machek,
Hi!
In message <20121105200340.GA15821@xo-6d-61-c0.localdomain> you wrote:/* Append length in bits and transform */
ctx->in32[14] = ctx->bits[0];ctx->in32[15] = ctx->bits[1];
memcpy(ctx->in + 14 * sizeof(__u32), ctx->bits, 2 *sizeof(__u32));
This makes the code actually unreadable. Please add at least a comment what this is doing.
Actually I think this shoul dbe split into two memcpy commands,
using
the addresses of the respective array elements directly, without
such
manual pointer arithmetics.
I guess bigger question is: why does gcc miscompile that, and is it guaranteed that it will not miscompile the memcpy?
I did not see Simon mentioning anythin about incorrect compilation. My understanding was that it's just the usual "dereferencing type-punned pointer" warnings issue.Is there some alternate solution? The memcpy is really ugly...
Plus... does it solve the issue? The code does not look like being compatible with strict pointer aliasing... and I don't think memcpy() helps.
arch/nds32/config.mk:PLATFORM_RELFLAGS += -fno-strict-aliasing -fno-common -mrelax arch/x86/config.mk:PLATFORM_CPPFLAGS += -fno-strict-aliasing
We should really do that globally. Pavel
Were you even able to replicate this problem in the first place? Isn't this whole "problem" a problem of a broken (ubuntu/linaro) toolchain again?
Best regards, Marek Vasut