[Bug middle-end/77399] New: Fails to use native instructions for vector casts

[amonakov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77399

            Bug ID: 77399
           Summary: Fails to use native instructions for vector casts
           Product: gcc
           Version: 7.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: amonakov at gcc dot gnu.org
                CC: nsz at gcc dot gnu.org
  Target Milestone: ---

For the following testcase:

typedef int   v4si __attribute__((vector_size(16)));
typedef float v4sf __attribute__((vector_size(16)));
v4sf vec_cast(v4si f)
{
  return (v4sf){f[0], f[1], f[2], f[3]};
}

(where unfortunately the c-style cast of the vector type has to be spelled out
per-component, as '(v4sf)f' would mean a bitwise copy of the vector
representation)

on x86-64 the generated should be just:

        cvtdq2ps        %xmm0, %xmm0
        retq

(and that is what Clang/LLVM generates), but GCC generates code that unpacks
the input vector and converts each component separately.

Note that the issue arises only with vector types; when auto-vectorizing scalar
code, GCC can use 'cvtdq2ps'. I believe this is because the original code is
lowered to gimple using 4 BIT_FIELD_REFS, and the vectorizer doesn't handle
that.