Segher Boessenkool <seg...@kernel.crashing.org> writes:
> Hi!
>
> On Tue, Oct 15, 2019 at 05:15:07PM +0800, Jiufu Guo wrote:
>> And another issue: Behavior is still inconsistent between "-mno-vsx
>> -flto" and "-mno-vsx" for user code. Previous patch makes it consistent
>> between "-mvsx -flto" and "-mvsx".
>
>> $GCC_BUILD/gcc/xgcc -B$GCC_BUILD/gcc novsx.c -O2 -mno-vsx
>> /home/guojiufu/temp/novsx.c: In function 'main':
>> /home/guojiufu/temp/novsx.c:6:1: error: inlining failed in call to
>> 'always_inline' 'foo': target specific option mismatch
>
> So what should we do about this? There are arguments for *both*
> behaviours, and apparently with LTO we do not know which flags are
> explicit?
With LTO, rs6000_isa_flags_explicit and rs6000_isa_flags are also
checkable. rs6000_isa_flags indicates effective features;
bits of rs6000_isa_flags_explicit indicate if it is specified
explicitly.
The difference between 'with LTO' and 'without LTO' is:
Wit LTO, if a function does not have target attribute in source code,
then using option attribute from command line(e.g. -mcpu=xxx) as
isa_flags; and if features (e.g. -mno-vsx, -mvsx) are specified on
command line, isa_flags_explicit is also set.
If a function has target attribute in source code explicitly,
then isa_flags_explicit is set to reflect feature is specified
explicitly besides explicit features from command line; and isa_flags is
set as the effective features.
Without LTO, if a function does not have target attribute in source code,
then it's isa_flags_explicit and isa_flags are as NULL (target_options
== NULL).
If a function has target attribute in source code, it is same as 'with
LTO'.
>
>> + /* Propogate global flags to caller. */
>> + HOST_WIDE_INT caller_isa = rs6000_isa_flags;
>
> I don't think that is right, or, I don't see why, anyway?
Propogating global flags is exiting behavior for function during LTO
compiling; and global flags are also used to caculate
isa_flags/isa_flags_explicit for function which has target attribute in
source code.
So, it would make sense to using global flags for caller here.
One more exmaple to explain this:
For example a function "int foo () {xxxxx;}", to compile it using
command "gcc -mcpu=power8 -mno-vsx"; we should using power8+no-vsx
instructions for this function.
>
>> +
>> + if (((caller_isa & callee_isa) == callee_isa)
>> + && (caller_isa & explicit_isa) == (callee_isa & explicit_isa))
>> + ret = true;
>> + }
>
>
> Segher
