Re: optimization for simd intrinsics vld2_dup_* on aarch64-none-elf

[Kyrill Tkachov]

Hi Shanyao,

On 03/09/14 16:02, shanyao chen wrote:
Hi,
I found there is a performance problem with some simd intrinsics
(vld2_dup_*) on aarch64-none-elf. Now the vld2_dup_* are defined as
follows:
#define __LD2R_FUNC(rettype, structtype, ptrtype, \
     regsuffix, funcsuffix, Q) \
   __extension__ static __inline rettype \
   __attribute__ ((__always_inline__))  \
   vld2 ## Q ## _dup_ ## funcsuffix (const ptrtype *ptr) \
   { \
     rettype result; \
     __asm__ ("ld2r {v16." #regsuffix ", v17." #regsuffix "}, %1\n\t" \
      "st1 {v16." #regsuffix ", v17." #regsuffix "}, %0\n\t" \
      : "=Q"(result) \
      : "Q"(*(const structtype *)ptr) \
      : "memory", "v16", "v17"); \
     return result; \
   }

It loads from memory to registers, and then store the value of
registers to memory as a result. Such code is terribly low in
performance because of redundant memory visit and limited registers
allocation.

Some intinsics like vld2_* were similar to vld2_dup_*, but now they
are realized by builtin functions.

__extension__ static __inline int16x4x2_t __attribute__ ((__always_inline__))
vld2_s16 (const int16_t * __a)
{
   int16x4x2_t ret;
   __builtin_aarch64_simd_oi __o;
   __o = __builtin_aarch64_ld2v4hi ((const __builtin_aarch64_simd_hi *) __a);
   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
   return ret;
}

Could vld2_dup_* also be written as builtin ?  If not, i think the
inline assembler can be optimized as follows :

The arm port implements these with builtins, it should possible to 
implement them that way on aarch64 as well.

Could you log an issue in bugzilla please, including some source code 
demonstrating the poor
codegen if possible?

Thanks,
Kyrill

#define __LD2R_FUNC(rettype, structtype, ptrtype, \
     regsuffix, funcsuffix, Q) \
   __extension__ static __inline rettype \
   __attribute__ ((__always_inline__))  \
   vld2 ## Q ## _dup_ ## funcsuffix (const ptrtype *ptr) \
   { \
     rettype result; \
     __asm__ (
              "ld2r {%0.4h, %1.4h}, %2"                           \
              : "=V16"(result.val[0]), "=V17"(result.val[1]) \
      : "Q"(*(const structtype *)ptr) \
      : "memory", "v16", "v17"); \
     return result; \
   }

It need to add a reg_class_name v16&v17 and add constraints V16 &  V17
for them. For this, aarch64.h、aarch64.c、constraints.md should be
modified.