[Bug tree-optimization/99856] [9/10/11 Regression] Alpha Compositing auto vectorization regression

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99856

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Version|unknown                     |9.3.0
           Priority|P3                          |P2
             Status|NEW                         |ASSIGNED
   Target Milestone|---                         |9.4
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
             Target|                            |x86_64-*-*

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
t.c:16:14: note:   op template: patt_266 = (<unnamed-unsigned:17>) _8;
t.c:16:14: note:        stmt 0 patt_266 = (<unnamed-unsigned:17>) _8;
t.c:16:14: note:        stmt 1 patt_243 = (<unnamed-unsigned:17>) _28;
t.c:16:14: note:        stmt 2 patt_220 = (<unnamed-unsigned:17>) _46;
t.c:16:14: note:        stmt 3 patt_198 = (<unnamed-unsigned:17>) _60;
t.c:16:14: note:        children 0x38267e0

WTF

t.c:16:14: note:   ==> examining pattern def stmt: patt_266 =
(<unnamed-unsigned:17>) _8;
t.c:16:14: note:   precomputed vectype: vector(8) unsigned int
t.c:16:14: note:   get vectype for smallest scalar type: unsigned char

the dumps do not show who creates this 17 bit precision temporary, but clearly
this causes

t.c:16:14: note:   ==> examining statement: patt_266 = (<unnamed-unsigned:17>)
_8;
t.c:16:14: note:   vect_is_simple_use: operand *_7, type of def: internal
t.c:16:14: missed:   type conversion to/from bit-precision unsupported.
t.c:16:14: missed:   bit-precision arithmetic not supported.
t.c:16:14: note:   vect_is_simple_use: operand *_7, type of def: internal
t.c:11:1: missed:   not vectorized: relevant stmt not supported: patt_266 =
(<unnamed-unsigned:17>) _8;
t.c:16:14: note:   removing SLP instance operations starting from: *_20 = _21;
t.c:16:14: missed:  unsupported SLP instances
t.c:16:14: note:  re-trying with SLP disabled
...
t.c:16:14: note:   ==> examining statement: patt_266 = (<unnamed-unsigned:17>)
_8;
t.c:16:14: note:   vect_is_simple_use: operand *_7, type of def: internal
t.c:16:14: note:   vect_is_simple_use: vectype const vector(32) unsigned char
t.c:16:14: missed:   type conversion to/from bit-precision unsupported.
t.c:16:14: missed:   bit-precision arithmetic not supported.
t.c:16:14: note:   vect_is_simple_use: operand *_7, type of def: internal
t.c:16:14: note:   vect_is_simple_use: vectype const vector(32) unsigned char
t.c:11:1: missed:   not vectorized: relevant stmt not supported: patt_266 =
(<unnamed-unsigned:17>) _8;
t.c:16:14: missed:  bad operation or unsupported loop bound.


Breakpoint 6, build_nonstandard_integer_type (precision=17, unsignedp=1)
    at /home/rguenther/src/gcc3/gcc/tree.c:8090
8090      if (unsignedp)
(gdb) bt
#0  build_nonstandard_integer_type (precision=17, unsignedp=1)
    at /home/rguenther/src/gcc3/gcc/tree.c:8090
#1  0x00000000024461ab in vect_recog_over_widening_pattern (vinfo=0x385bf60, 
    last_stmt_info=0x38d1130, type_out=0x7fffffffc7d0)
    at /home/rguenther/src/gcc3/gcc/tree-vect-patterns.c:1719
#2  0x0000000002453c10 in vect_pattern_recog_1 (vinfo=0x385bf60, 
    recog_func=0x359bf80 <vect_vect_recog_func_ptrs>, stmt_info=0x38d1130)
    at /home/rguenther/src/gcc3/gcc/tree-vect-patterns.c:5463
#3  0x0000000002453fee in vect_pattern_recog (vinfo=0x385bf60)
    at /home/rguenther/src/gcc3/gcc/tree-vect-patterns.c:5603

where

(gdb) p last_stmt_info->min_output_precision
$9 = 17
(gdb) p last_stmt_info->operation_precision
$10 = 8

and the min_output_precision is determined from the defs uses
min_input_precision.

I didn't trace further where we arrive at '17', but

diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c
index b575b456301..803de3fc287 100644
--- a/gcc/tree-vect-patterns.c
+++ b/gcc/tree-vect-patterns.c
@@ -1705,6 +1705,7 @@ vect_recog_over_widening_pattern (vec_info *vinfo,
   /* Apply the minimum efficient precision we just calculated.  */
   if (new_precision < min_precision)
     new_precision = min_precision;
+  new_precision = vect_element_precision (new_precision);
   if (new_precision >= TYPE_PRECISION (type))
     return NULL;

restores vectorization.