[Bug target/87354] New: x86-64: 16- and 32-byte register variables cannot be put in XMM16/YMM16 and up without -mavx512vl

[jbeulich at novell dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87354

            Bug ID: 87354
           Summary: x86-64: 16- and 32-byte register variables cannot be
                    put in XMM16/YMM16 and up without -mavx512vl
           Product: gcc
           Version: 8.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jbeulich at novell dot com
  Target Milestone: ---

The same issue had been present in gas, but was corrected by commit 6e041cf4b0:
YMM (and of course also XMM) registers can certainly be used with AVX512F
alone, just that the set of insns is pretty limited. I realize that making this
work may not be a trivial change, as assumptions to this effect appear to be
made all over the place, but this code should imo compile (and assemble) fine
with just -mavx512f, while currently only the first function compiles without
error (QI mode vectors have been used just for simplicity and to make things
look reasonably uniform):

asm(".arch generic64");
asm(".arch .avx512f");

typedef char __attribute__((vector_size(64))) v64qi_t;
typedef char __attribute__((vector_size(16))) v16qi_t;
typedef char __attribute__((vector_size(32))) v32qi_t;

v64qi_t test512(v64qi_t x) {
        register v64qi_t y asm("zmm16");
        asm("vmovdqa32 %1,%0" : "=v" (y) : "v" (x));
        return y;
}

v16qi_t test128(v64qi_t x) {
        register v16qi_t y asm("xmm16");
        asm("vpmovqw %1,%0" : "=v" (y) : "v" (x));
        return y;
}

v32qi_t test256(v64qi_t x) {
        register v32qi_t y asm("ymm16");
        asm("vpmovqd %1,%0" : "=v" (y) : "v" (x));
        return y;
}