[Bug tree-optimization/92819] [10 Regression] Worse code generated on avx2 due to simplify_vector_constructor

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92819

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
Update:

--- t2.s.9      2019-12-06 09:03:03.104331362 +0100
+++ t2.s.10     2019-12-06 09:02:52.700181055 +0100
@@ -6,8 +6,9 @@
 foo:
 .LFB0:
        .cfi_startproc
-       vunpckhpd       %xmm0, %xmm0, %xmm0
-       vmovhpd (%rdi), %xmm0, %xmm0
+       vbroadcastsd    (%rdi), %ymm1
+       vperm2f128      $48, %ymm1, %ymm0, %ymm0
+       vshufpd $11, %ymm1, %ymm0, %ymm0
        ret
        .cfi_endproc
 .LFE0:
@@ -29,8 +30,8 @@
 baz:
 .LFB2:
        .cfi_startproc
-       vunpckhpd       %xmm0, %xmm0, %xmm0
-       vmovhpd (%rdi), %xmm0, %xmm0
+       vmovddup        (%rdi), %xmm1
+       vunpckhpd       %xmm1, %xmm0, %xmm0
        ret
        .cfi_endproc
 .LFE2:
@@ -52,12 +53,12 @@
 corge:
 .LFB4:
        .cfi_startproc
-       vextractf128    $0x1, %ymm0, %xmm0
-       vunpckhpd       %xmm0, %xmm0, %xmm0
-       vmovhpd (%rdi), %xmm0, %xmm0
+       vbroadcastsd    (%rdi), %ymm1
+       vperm2f128      $49, %ymm1, %ymm0, %ymm0
+       vshufpd $11, %ymm1, %ymm0, %ymm0
        ret
        .cfi_endproc
 .LFE4:
        .size   corge, .-corge
-       .ident  "GCC: (GNU) 9.2.0"
+       .ident  "GCC: (GNU) 10.0.0 20191206 (experimental) [trunk revision
270575]"
        .section        .note.GNU-stack,"",@progbits

"foo" is clearly worse, "corge" very likely and "baz" _might_ be slightly
better (the load happens one instruction earlier, eventually hiding some
latency and a dup from a load should be able to micro-fuse to the load op
in my naiive wishful thinking).

foo (v4df x, double * p)
{
  double _1;
  vector(4) double _5;
  vector(4) double _6;
  vector(2) double _7;

  <bb 2> [local count: 1073741824]:
  _1 = *p_4(D);
  _5 = {_1, _1, _1, _1};
  _6 = VEC_PERM_EXPR <x_2(D), _5, { 1, 5, 6, 7 }>;
  _7 = BIT_FIELD_REF <_6, 128, 0>;
  return _7;

baz (v2df x, double * p)
{
  double _1;
  vector(2) double _5;
  vector(2) double _6;

  <bb 2> [local count: 1073741824]:
  _1 = *p_4(D);
  _5 = {_1, _1};
  _6 = VEC_PERM_EXPR <x_2(D), _5, { 1, 3 }>;
  return _6;

corge (v4df x, double * p, v2df y)
{
  double _1;
  vector(4) double _5;
  vector(4) double _6;
  vector(2) double _7;

  <bb 2> [local count: 1073741824]:
  _1 = *p_4(D); 
  _5 = {_1, _1, _1, _1};
  _6 = VEC_PERM_EXPR <x_2(D), _5, { 3, 5, 6, 7 }>;
  _7 = BIT_FIELD_REF <_6, 128, 0>;
  return _7;


Now disabling VEC_PERM generation won't be a 100% solution in general since
the user can present us with the above as well.

Looking at "corge" gcc9 expanded this as

  extract upper half of x
  extract lane 1 from this
  load the scalar
  concat both

so it performed the permute on smaller vectors rather than permuting
the orignals and then selecting the lowpart.

I'll see if I can cook up sth for simplify_bitfield_ref.