Re: [PATCH 2/6] x86: bug: prevent gcc distortions

From: Nadav Amit
Date: Fri May 18 2018 - 09:34:22 EST


Ingo Molnar <mingo@xxxxxxxxxx> wrote:

>
> * Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
>> On Thu, May 17, 2018 at 09:13:58AM -0700, Nadav Amit wrote:
>>> +asm(".macro __BUG_FLAGS ins:req file:req line:req flags:req size:req\n"
>>> + "1:\t \\ins\n\t"
>>> + ".pushsection __bug_table,\"aw\"\n"
>>> + "2:\t "__BUG_REL(1b) "\t# bug_entry::bug_addr\n\t"
>>> + __BUG_REL(\\file) "\t# bug_entry::file\n\t"
>>> + ".word \\line" "\t# bug_entry::line\n\t"
>>> + ".word \\flags" "\t# bug_entry::flags\n\t"
>>> + ".org 2b+\\size\n\t"
>>> + ".popsection\n\t"
>>> + ".endm");
>>> +
>>> +#define _BUG_FLAGS(ins, flags) \
>>> do { \
>>> + asm volatile("__BUG_FLAGS \"" ins "\" %c0 %c1 %c2 %c3" \
>>> + : : "i" (__FILE__), "i" (__LINE__), \
>>> + "i" (flags), \
>>> "i" (sizeof(struct bug_entry))); \
>>> } while (0)
>>
>> This is an awesome hack, but is there really nothing we can do to make
>> it more readable? Esp, that global asm doing the macro definition is a
>> pain to read.
>>
>> Also, can we pretty please used named operands in 'new' code?
>
> Yes, that's my main worry too about all these inlining changes:
> the very, very marked reduction in the readability of assembly code.
>
> It's bad to an extent that I'm questioning the wisdom of pandering to a compiler
> limitation to begin with?
>
> How about asking GCC for an attribute where we can specify the inlined size of an
> asm() function? Even if we'll just approximate it due to some vagaries of actual
> code generation related to how arguments interact with GCC, an explicit byte value
> will do a heck of a better job of it than some sort of implied, vague 'number of
> newlines' heuristics ...

If it were to become a GCC feature, I think it is best to be a builtin that
says: consider the enclosed expression as âfreeâ. The problem of poor
inlining decisions is not specific to inline asm. As I mentioned in the RFC,
when there are two code paths for constants and variables based on
__builtin_constant_p(), you can get the âcostâ of the constant path for
variables.

It is not hard to add such a feature to GCC, but I donât know how easy it is
to get new features into the compiler.