On 02.02.2021 02:00, Ivan Sorokin wrote:
> [PATCH] [PR91400] build only one __cpu_model variable
This is my first patch to GCC. So I might have done something totally
stupid or totally wrong. Caution is required for reviewing. :-)
> gcc/ChangeLog:
>
> PR target/91400
> * config/i386/i386-builtins.c (fold_builtin_cpu): Extract
> building of __cpu_model and __processor_model into new
> function.
> * config/i386/i386-builtins.c (init_cpu_model_var): New.
> Cache creation of __cpu_model and __processor_model.
>
> gcc/testsuite/Changelog:
>
> PR target/91400
> * gcc.target/i386/pr91400.c: New.
I wrote the change log text manually. I hope I didn't mess up the
formatting.
> +static GTY(()) tree __cpu_model_var;
> +static GTY(()) tree __processor_model_type;
I felt a bit uneasy writing global variables, but this file contains
other global variables already and they are used for similar purpose.
> +static void
> +init_cpu_model_var()
> +{
> + if (__cpu_model_var != NULL_TREE)
> + {
> + gcc_assert(__processor_model_type != NULL_TREE);
> + return;
> + }
> +
> + __processor_model_type = build_processor_model_struct ();
> + __cpu_model_var = make_var_decl (__processor_model_type,
> + "__cpu_model");
> +
> + varpool_node::add (__cpu_model_var);
I have no idea what this line does. But I decided that perhaps we want
to do it only once instead of once for each usage of __builtin_cpu_supports.
> diff --git a/gcc/testsuite/gcc.target/i386/pr91400.c
> b/gcc/testsuite/gcc.target/i386/pr91400.c
> new file mode 100644
> index 00000000000..e8b7d9285f9
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr91400.c
> @@ -0,0 +1,11 @@
> +/* PR target/91400 */
> +/* { dg-do compile { target { ! ia32 } } } */
> +/* { dg-options "-O2" } */
> +/* { dg-final { scan-assembler-times "andl" 1 } } */
> +/* { dg-final { scan-assembler-times "68" 2 } } */
> +/* { dg-final { scan-assembler-not "je" } } */
> +
> +_Bool f()
> +{
> + return __builtin_cpu_supports("popcnt") &&
> __builtin_cpu_supports("ssse3");
> +}
The test was the most complicated thing for me. Previous versions of GCC
did two instructions andl, so I written check for them. Current master
does a conditional jump, so I add test for it too. Any suggestions for
better checks are welcomed.
