Re: [x86][patch] Add -march=cannonlake.

[Uros Bizjak <ubizjak@gmail.com>]

On Wed, Nov 8, 2017 at 9:02 AM, Koval, Julia <julia.koval@intel.com> wrote:
> Attachment got lost.
>
>> -----Original Message-----
>> From: Koval, Julia
>> Sent: Wednesday, November 08, 2017 9:01 AM
>> To: 'GCC Patches' <gcc-patches@gcc.gnu.org>
>> Cc: 'Uros Bizjak' <ubizjak@gmail.com>; 'Kirill Yukhin' <kirill.yukhin@gmail.com>
>> Subject: RE: [x86][patch] Add -march=cannonlake.
>>
>> Hi, this patch adds new option -march=cannonlake. Isasets defined in:
>> https://software.intel.com/sites/default/files/managed/c5/15/architecture-
>> instruction-set-extensions-programming-reference.pdf
>>
>> Ok for trunk?
>>
>> gcc/
>>       * config.gcc: Add -march=cannonlake.
>>       * config/i386/driver-i386.c (host_detect_local_cpu): Detect cannonlake.
>>       * config/i386/i386-c.c (ix86_target_macros_internal): Handle
>> cannonlake.
>>       * config/i386/i386.c (processor_costs): Add m_CANNONLAKE.
>>       (PTA_CANNONLAKE): New.
>>       (processor_target_table): Add cannonlake.
>>       (ix86_option_override_internal): Ditto.
>>       (fold_builtin_cpu): Ditto.
>>       (get_builtin_code_for_version): Handle cannonlake.
>>       (M_INTEL_CANNONLAKE): New.
>>       * config/i386/i386.h (TARGET_CANNONLAKE,
>> PROCESSOR_CANNONLAKE): New.
>>       * doc/invoke.texi: Add -march=cannonlake.
>> gcc/testsuite/
>>       * gcc.target/i386/funcspec-56.inc: Handle new march.

--- a/gcc/config/i386/driver-i386.c
+++ b/gcc/config/i386/driver-i386.c
@@ -803,9 +803,11 @@ const char *host_detect_local_cpu (int argc,
const char **argv)
     default:
       if (arch)
         {
+          if (has_avx512vbmi)
+        cpu = "cannonlake";
           /* This is unknown family 0x6 CPU.  */
           /* Assume Knights Landing.  */
-          if (has_avx512f)
+          else if (has_avx512f)
         cpu = "knl";
           /* Assume Knights Mill */
           else if (has_avx5124vnniw)

You should add correct model numbers under <case
PROCESSOR_PENTIUMPRO>. The above is for the unknown case (which should
not happen), and it should read (note that "knl" is already
misplaced):

    default:
      /* This is unknown family 0x6 CPU.  */
      if (arch)
        {
          /* Assume Cannonlake.  */
          if (has_avx512vbmi)
        cpu = "cannonlake";
          /* Assume Knights Mill */
          else if (has_avx5124vnniw)
        cpu = "knm";
          /* Assume Skylake.  */
          else if (has_clflushopt)
        cpu = "skylake";
          /* Assume Knights Landing.  */
          else if (has_avx512f)
        cpu = "knl";
          /* Assume Broadwell.  */
          else if (has_adx)
        ...


@@ -31832,7 +31839,8 @@ fold_builtin_cpu (tree fndecl, tree *args)
     M_INTEL_COREI7_HASWELL,
     M_INTEL_COREI7_BROADWELL,
     M_INTEL_COREI7_SKYLAKE,
-    M_INTEL_COREI7_SKYLAKE_AVX512
+    M_INTEL_COREI7_SKYLAKE_AVX512,
+    M_INTEL_CANNONLAKE
   };

Please also update libgcc/config/i386/cpuinfo.h, enum processor_features.

diff --git a/gcc/testsuite/gcc.target/i386/funcspec-56.inc
b/gcc/testsuite/gcc.target/i386/funcspec-56.inc
index 9ae74cb..ed0748b 100644
--- a/gcc/testsuite/gcc.target/i386/funcspec-56.inc
+++ b/gcc/testsuite/gcc.target/i386/funcspec-56.inc
@@ -144,6 +144,7 @@ extern void test_arch_core_avx2 (void)
__attribute__((__target__("arch=core-avx
 extern void test_arch_knl (void)
__attribute__((__target__("arch=knl")));
 extern void test_arch_knm (void)
__attribute__((__target__("arch=knm")));
 extern void test_arch_skylake_avx512 (void)
__attribute__((__target__("arch=skylake-avx512")));
+extern void test_arch_cannonlake (void)
__attribute__((__target__("arch=cannonlake")));
 extern void test_arch_k8 (void)
__attribute__((__target__("arch=k8")));
 extern void test_arch_k8_sse3 (void)
__attribute__((__target__("arch=k8-sse3")));
 extern void test_arch_opteron (void)
__attribute__((__target__("arch=opteron")));

Please also add new architecture to multiversioning testcases, see
gcc/testsuite/g++.dg/ext/mv*.C

Uros.