On 23/02/2016 12:25, Peter Maydell wrote:
> On 23 February 2016 at 10:43, Peter Maydell <peter.mayd...@linaro.org> wrote:
>> That code fragment you suggest compiles fine normally, but not if I
>> add -save-temps:
>>
>> $ cat /tmp/zz9.c
>> #pragma GCC target("avx2")
>> #include <immintrin.h>
>> __m256i foo;
>> $ gcc -g -Wall -o /tmp/zz9.o -c /tmp/zz9.c
>> $ echo $?
>> 0
>> $ gcc -g -Wall -o /tmp/zz9.o -c /tmp/zz9.c -save-temps
>> /tmp/zz9.c:4:1: error: unknown type name ‘__m256i’
>> __m256i foo;
>> ^
>> /tmp/zz9.c: In function ‘bar’:
>> /tmp/zz9.c:7:19: error: ‘__m256i’ undeclared (first use in this function)
>> return sizeof(__m256i);
>> ^
>> /tmp/zz9.c:7:19: note: each undeclared identifier is reported only
>> once for each function it appears in
>> /tmp/zz9.c:8:1: warning: control reaches end of non-void function
>> [-Wreturn-type]
>> }
>> ^
>>
>> This seems to be because -save-temps causes the #pragma not to
>> actually #define __AVX__.
>
> This is because -save-temps causes gcc to invoke the
> preprocessor and the compiler as separate passes, and the
> standalone preprocessor doesn't know that the target pragma
> should result in a new #define, so the result is that the
> immintrin.h doesn't pull in what it should.
>
> This is also the reason why my build failed -- I use ccache,
> which is another tool that results in the preprocessor being
> done as a standalone pass rather than in the same pass as
> compilation proper.
>
> Arguably it's a gcc bug that the target pragma doesn't cause
> the standalone preprocessor to define the same #defines that
> you get if it's all in one pass, but regardless I don't think
> we can break ccache builds, so you'll need to find a different
> way to do this, I'm afraid.
It's a bug in the header file, it was fixed in 4.9.
https://gcc.gnu.org/ml/gcc-patches/2013-06/txtvBBiTsFs8g.txt
Amit or Liang, can you restrict the new optimization to GCC 4.9+?
Paolo
