On Thu, 2017-09-14 at 09:38 -0500, Bill Schmidt wrote:
> On Sep 14, 2017, at 5:15 AM, Richard Biener <richard.guent...@gmail.com>
> wrote:
> >
> > On Wed, Sep 13, 2017 at 10:14 PM, Bill Schmidt
> > <wschm...@linux.vnet.ibm.com> wrote:
> >> On Sep 13, 2017, at 10:40 AM, Bill Schmidt <wschm...@linux.vnet.ibm.com>
> >> wrote:
> >>>
> >>> On Sep 13, 2017, at 7:23 AM, Richard Biener <richard.guent...@gmail.com>
> >>> wrote:
> >>>>
> >>>> On Tue, Sep 12, 2017 at 11:08 PM, Will Schmidt
> >>>> <will_schm...@vnet.ibm.com> wrote:
> >>>>> Hi,
> >>>>>
> >>>>> [PATCH, rs6000] [v2] Folding of vector loads in GIMPLE
> >>>>>
> >>>>> Folding of vector loads in GIMPLE.
> >>>>>
> >>>>> Add code to handle gimple folding for the vec_ld builtins.
> >>>>> Remove the now obsoleted folding code for vec_ld from rs6000-c.c.
> >>>>> Surrounding
> >>>>> comments have been adjusted slightly so they continue to read OK for the
> >>>>> existing vec_st code.
> >>>>>
> >>>>> The resulting code is specifically verified by the
> >>>>> powerpc/fold-vec-ld-*.c
> >>>>> tests which have been posted separately.
> >>>>>
> >>>>> For V2 of this patch, I've removed the chunk of code that prohibited the
> >>>>> gimple fold from occurring in BE environments. This had fixed an issue
> >>>>> for me earlier during my development of the code, and turns out this was
> >>>>> not necessary. I've sniff-tested after removing that check and it looks
> >>>>> OK.
> >>>>>
> >>>>>> + /* Limit folding of loads to LE targets. */
> >>>>>> + if (BYTES_BIG_ENDIAN || VECTOR_ELT_ORDER_BIG)
> >>>>>> + return false;
> >>>>>
> >>>>> I've restarted a regression test on this updated version.
> >>>>>
> >>>>> OK for trunk (assuming successful regression test completion) ?
> >>>>>
> >>>>> Thanks,
> >>>>> -Will
> >>>>>
> >>>>> [gcc]
> >>>>>
> >>>>> 2017-09-12 Will Schmidt <will_schm...@vnet.ibm.com>
> >>>>>
> >>>>> * config/rs6000/rs6000.c (rs6000_gimple_fold_builtin): Add handling
> >>>>> for early folding of vector loads (ALTIVEC_BUILTIN_LVX_*).
> >>>>> * config/rs6000/rs6000-c.c (altivec_resolve_overloaded_builtin):
> >>>>> Remove obsoleted code for handling ALTIVEC_BUILTIN_VEC_LD.
> >>>>>
> >>>>> diff --git a/gcc/config/rs6000/rs6000-c.c b/gcc/config/rs6000/rs6000-c.c
> >>>>> index fbab0a2..bb8a77d 100644
> >>>>> --- a/gcc/config/rs6000/rs6000-c.c
> >>>>> +++ b/gcc/config/rs6000/rs6000-c.c
> >>>>> @@ -6470,92 +6470,19 @@ altivec_resolve_overloaded_builtin (location_t
> >>>>> loc, tree fndecl,
> >>>>> convert (TREE_TYPE (stmt), arg0));
> >>>>> stmt = build2 (COMPOUND_EXPR, arg1_type, stmt, decl);
> >>>>> return stmt;
> >>>>> }
> >>>>>
> >>>>> - /* Expand vec_ld into an expression that masks the address and
> >>>>> - performs the load. We need to expand this early to allow
> >>>>> + /* Expand vec_st into an expression that masks the address and
> >>>>> + performs the store. We need to expand this early to allow
> >>>>> the best aliasing, as by the time we get into RTL we no longer
> >>>>> are able to honor __restrict__, for example. We may want to
> >>>>> consider this for all memory access built-ins.
> >>>>>
> >>>>> When -maltivec=be is specified, or the wrong number of arguments
> >>>>> is provided, simply punt to existing built-in processing. */
> >>>>> - if (fcode == ALTIVEC_BUILTIN_VEC_LD
> >>>>> - && (BYTES_BIG_ENDIAN || !VECTOR_ELT_ORDER_BIG)
> >>>>> - && nargs == 2)
> >>>>> - {
> >>>>> - tree arg0 = (*arglist)[0];
> >>>>> - tree arg1 = (*arglist)[1];
> >>>>> -
> >>>>> - /* Strip qualifiers like "const" from the pointer arg. */
> >>>>> - tree arg1_type = TREE_TYPE (arg1);
> >>>>> - if (!POINTER_TYPE_P (arg1_type) && TREE_CODE (arg1_type) !=
> >>>>> ARRAY_TYPE)
> >>>>> - goto bad;
> >>>>> -
> >>>>> - tree inner_type = TREE_TYPE (arg1_type);
> >>>>> - if (TYPE_QUALS (TREE_TYPE (arg1_type)) != 0)
> >>>>> - {
> >>>>> - arg1_type = build_pointer_type (build_qualified_type
> >>>>> (inner_type,
> >>>>> - 0));
> >>>>> - arg1 = fold_convert (arg1_type, arg1);
> >>>>> - }
> >>>>> -
> >>>>> - /* Construct the masked address. Let existing error handling
> >>>>> take
> >>>>> - over if we don't have a constant offset. */
> >>>>> - arg0 = fold (arg0);
> >>>>> -
> >>>>> - if (TREE_CODE (arg0) == INTEGER_CST)
> >>>>> - {
> >>>>> - if (!ptrofftype_p (TREE_TYPE (arg0)))
> >>>>> - arg0 = build1 (NOP_EXPR, sizetype, arg0);
> >>>>> -
> >>>>> - tree arg1_type = TREE_TYPE (arg1);
> >>>>> - if (TREE_CODE (arg1_type) == ARRAY_TYPE)
> >>>>> - {
> >>>>> - arg1_type = TYPE_POINTER_TO (TREE_TYPE (arg1_type));
> >>>>> - tree const0 = build_int_cstu (sizetype, 0);
> >>>>> - tree arg1_elt0 = build_array_ref (loc, arg1, const0);
> >>>>> - arg1 = build1 (ADDR_EXPR, arg1_type, arg1_elt0);
> >>>>> - }
> >>>>> -
> >>>>> - tree addr = fold_build2_loc (loc, POINTER_PLUS_EXPR,
> >>>>> arg1_type,
> >>>>> - arg1, arg0);
> >>>>> - tree aligned = fold_build2_loc (loc, BIT_AND_EXPR, arg1_type,
> >>>>> addr,
> >>>>> - build_int_cst (arg1_type,
> >>>>> -16));
> >>>>> -
> >>>>> - /* Find the built-in to get the return type so we can convert
> >>>>> - the result properly (or fall back to default handling if
> >>>>> the
> >>>>> - arguments aren't compatible). */
> >>>>> - for (desc = altivec_overloaded_builtins;
> >>>>> - desc->code && desc->code != fcode; desc++)
> >>>>> - continue;
> >>>>> -
> >>>>> - for (; desc->code == fcode; desc++)
> >>>>> - if (rs6000_builtin_type_compatible (TREE_TYPE (arg0),
> >>>>> desc->op1)
> >>>>> - && (rs6000_builtin_type_compatible (TREE_TYPE (arg1),
> >>>>> - desc->op2)))
> >>>>> - {
> >>>>> - tree ret_type = rs6000_builtin_type (desc->ret_type);
> >>>>> - if (TYPE_MODE (ret_type) == V2DImode)
> >>>>> - /* Type-based aliasing analysis thinks vector long
> >>>>> - and vector long long are different and will put
> >>>>> them
> >>>>> - in distinct alias classes. Force our return type
> >>>>> - to be a may-alias type to avoid this. */
> >>>>> - ret_type
> >>>>> - = build_pointer_type_for_mode (ret_type, Pmode,
> >>>>> -
> >>>>> true/*can_alias_all*/);
> >>>>> - else
> >>>>> - ret_type = build_pointer_type (ret_type);
> >>>>> - aligned = build1 (NOP_EXPR, ret_type, aligned);
> >>>>> - tree ret_val = build_indirect_ref (loc, aligned,
> >>>>> RO_NULL);
> >>>>> - return ret_val;
> >>>>> - }
> >>>>> - }
> >>>>> - }
> >>>>>
> >>>>> - /* Similarly for stvx. */
> >>>>> if (fcode == ALTIVEC_BUILTIN_VEC_ST
> >>>>> && (BYTES_BIG_ENDIAN || !VECTOR_ELT_ORDER_BIG)
> >>>>> && nargs == 3)
> >>>>> {
> >>>>> tree arg0 = (*arglist)[0];
> >>>>> diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
> >>>>> index 1338371..1fb5f44 100644
> >>>>> --- a/gcc/config/rs6000/rs6000.c
> >>>>> +++ b/gcc/config/rs6000/rs6000.c
> >>>>> @@ -16547,10 +16547,61 @@ rs6000_gimple_fold_builtin
> >>>>> (gimple_stmt_iterator *gsi)
> >>>>> res = gimple_build (&stmts, VIEW_CONVERT_EXPR, TREE_TYPE (lhs),
> >>>>> res);
> >>>>> gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
> >>>>> update_call_from_tree (gsi, res);
> >>>>> return true;
> >>>>> }
> >>>>> + /* Vector loads. */
> >>>>> + case ALTIVEC_BUILTIN_LVX_V16QI:
> >>>>> + case ALTIVEC_BUILTIN_LVX_V8HI:
> >>>>> + case ALTIVEC_BUILTIN_LVX_V4SI:
> >>>>> + case ALTIVEC_BUILTIN_LVX_V4SF:
> >>>>> + case ALTIVEC_BUILTIN_LVX_V2DI:
> >>>>> + case ALTIVEC_BUILTIN_LVX_V2DF:
> >>>>> + {
> >>>>> + gimple *g;
> >>>>> + arg0 = gimple_call_arg (stmt, 0); // offset
> >>>>> + arg1 = gimple_call_arg (stmt, 1); // address
> >>>>> +
> >>>>> + lhs = gimple_call_lhs (stmt);
> >>>>> + location_t loc = gimple_location (stmt);
> >>>>> +
> >>>>> + tree arg1_type = TREE_TYPE (arg1);
> >>>>> + tree lhs_type = TREE_TYPE (lhs);
> >>>>> +
> >>>>> + /* POINTER_PLUS_EXPR wants the offset to be of type
> >>>>> 'sizetype'. Create
> >>>>> + the tree using the value from arg0. The resulting type
> >>>>> will match
> >>>>> + the type of arg1. */
> >>>>> + tree temp_offset = create_tmp_reg_or_ssa_name (sizetype);
> >>>>> + g = gimple_build_assign (temp_offset, NOP_EXPR, arg0);
> >>>>> + gimple_set_location (g, loc);
> >>>>> + gsi_insert_before (gsi, g, GSI_SAME_STMT);
> >>>>> + tree temp_addr = create_tmp_reg_or_ssa_name (arg1_type);
> >>>>> + g = gimple_build_assign (temp_addr, POINTER_PLUS_EXPR, arg1,
> >>>>> + temp_offset);
> >>>>> + gimple_set_location (g, loc);
> >>>>> + gsi_insert_before (gsi, g, GSI_SAME_STMT);
> >>>>> +
> >>>>> + /* Mask off any lower bits from the address. */
> >>>>> + tree alignment_mask = build_int_cst (arg1_type, -16);
> >>>>> + tree aligned_addr = create_tmp_reg_or_ssa_name (arg1_type);
> >>>>> + g = gimple_build_assign (aligned_addr, BIT_AND_EXPR,
> >>>>> + temp_addr, alignment_mask);
> >>>>> + gimple_set_location (g, loc);
> >>>>> + gsi_insert_before (gsi, g, GSI_SAME_STMT);
> >>>>
> >>>> You could use
> >>>>
> >>>> gimple_seq stmts = NULL;
> >>>> tree temp_offset = gimple_convert (&stmts, loc, sizetype, arg0);
> >>>> tree temp_addr = gimple_build (&stmts, loc, POINTER_PLUS_EXPR,
> >>>> arg1_type, arg1, temp_offset);
> >>>> tree aligned_addr = gimple_build (&stmts, loc, BIT_AND_EXPR,
> >>>> arg1_type, temp_addr, build_int_cst (arg1_type, -16));
> >>>> gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
> >>>>
> >>>>> + /* Use the build2 helper to set up the mem_ref. The MEM_REF
> >>>>> could also
> >>>>> + take an offset, but since we've already incorporated the
> >>>>> offset
> >>>>> + above, here we just pass in a zero. */
> >>>>> + g = gimple_build_assign (lhs, build2 (MEM_REF, lhs_type,
> >>>>> aligned_addr,
> >>>>> + build_int_cst
> >>>>> (arg1_type, 0)));
> >>>>
> >>>> are you sure about arg1_type here? I'm sure not. For
> >>>>
> >>>> ... foo (struct S *p)
> >>>> {
> >>>> return __builtin_lvx_v2df (4, (double *)p);
> >>>> }
> >>>>
> >>>> you'd end up with p as arg1 and thus struct S * as arg1_type and thus
> >>>> TBAA using 'struct S' to access the memory.
> >>>
> >>> Hm, is that so? Wouldn't arg1_type be double* since arg1 is (double *)p?
> >>> Will, you should probably test this example and see, but I'm pretty
> >>> confident
> >>> about this (see below).
> >>
> >> But, as I should have suspected, you're right. For some reason
> >> gimple_call_arg is returning p, stripped of the cast information where the
> >> user asserted that p points to a double*.
> >>
> >> Can you explain to me why this should be so? I assume that somebody
> >> has decided to strip_nops the argument and lose the cast.
> >
> > pointer types have no meaning in GIMPLE so we aggressively prune them.
> >
> >> Using ptr_type_node loses all type information, so that would be a
> >> regression from what we do today. In some cases we could reconstruct
> >> that this was necessarily, say, a double*, but I don't know how we would
> >> recover the signedness for an integer type.
> >
> > How did we handle the expansion previously - ah - it was done earlier
> > in the C FE. So why are you moving it to GIMPLE? The function is called
> > resolve_overloaded_builtin - what kind of overloading do you resolve here?
> > As said argument types might not be preserved.
>
> The AltiVec builtins allow overloaded names based on the argument types,
> using a special callout during parsing to convert the overloaded names to
> type-specific names. Historically these have then remained builtin calls
> until RTL expansion, which loses a lot of useful optimization. Will has been
> gradually implementing gimple folding for these builtins so that we can
> optimize simple vector arithmetic and so on. The overloading is still dealt
> with during parsing.
>
> As an example:
>
> double a[64];
> vector double x = vec_ld (0, a);
>
> will get translated into
>
> vector double x = __builtin_altivec_lvx_v2df (0, a);
>
> and
>
> unsigned char b[64];
> vector unsigned char y = vec_ld (0, b);
>
> will get translated into
>
> vector unsigned char y = __builtin_altivec_lvx_v16qi (0, b);
>
> So in resolving the overloading we still maintain the type info for arg1.
>
> Earlier I had dealt with the performance issue in a different way for the
> vec_ld and vec_st overloaded builtins, which created the rather grotty
> code in rs6000-c.c to modify the parse trees instead. My hope was that
> we could simplify the code by having Will deal with them as gimple folds
> instead. But if in so doing we lose type information, that may not be the
> right call.
>
> However, since you say that gimple aggressively removes the casts
> from pointer types, perhaps the code that we see in early gimple from
> the existing method might also be missing the type information? Will,
> it would be worth looking at that code to see. If it's no different then
> perhaps we still go ahead with the folding.
The rs6000-c.c version of the code did not fold unless arg0 was
constant; and if it was a constant, it appears the operation got turned
directly into a * reference. So there isn't a good before/after compare
there.
What I see:
return vec_ld (ll1, (vector double *)p);
at gimple-time after the rs6000-c.c folding was a mostly un-folded
D.3207 = __builtin_altivec_lvx_v2dfD.1443 (ll1D.3192, pD.3193);
while this, with a constant value for arg0:
return vec_ld (16, (vector double *)p);
at gimple time after rs6000-c.c folding became a reference:
_1 = p + 16;
_2 = _1 & -16B;
D.3196 = *_2;
with the rs6000.c gimple folding code (the changes I've got locally),
the before/after with arg0 is constant reads the same. When arg0 is a
variable:
return vec_ld (ll1, (vector double *)p);
at dump-gimple time it then becomes:
D.3208 = (sizetype) ll1D.3192;
D.3209 = pD.3193 + D.3208;
D.3210 = D.3209 & -16B;
D.3207 = MEM[(struct S *)D.3210];
And if I change the code such that arg1_type is instead ptr_type_node:
D.3207 = MEM[(voidD.44 *)D.3210];
So...
> Another note for Will: The existing code gives up when -maltivec=be has
> been specified, and you probably want to do that as well. That may be
> why you initially turned off big endian -- it is easy to misread that code.
> -maltivec=be is VECTOR_ELT_ORDER_BIG && !BYTES_BIG_ENDIAN.
Yeah, I apparently inverted and confused the logic when I made that
change. My current snippet reads as:
/* Do not fold for -maltivec=be on LE targets. */
if (VECTOR_ELT_ORDER_BIG && !BYTES_BIG_ENDIAN)
return false;
> Thanks,
> Bill
> >
> > Richard.
> >
> >> Bill
> >>>
> >>>>
> >>>> I think if the builtins have any TBAA constraints you need to build those
> >>>> explicitely, if not, you should use ptr_type_node aka no TBAA.
> >>>
> >>> The type signatures are constrained during parsing, so we should only
> >>> see allowed pointer types on arg1 by the time we get to gimple folding. I
> >>> think that using arg1_type should work, but I am probably missing
> >>> something subtle, so please feel free to whack me on the temple until
> >>> I get it. :-)
> >>>
> >>> Bill
> >>>>
> >>>> Richard.
> >>>>
> >>>>> + gimple_set_location (g, loc);
> >>>>> + gsi_replace (gsi, g, true);
> >>>>> +
> >>>>> + return true;
> >>>>> +
> >>>>> + }
> >>>>> +
> >>>>> default:
> >>>>> if (TARGET_DEBUG_BUILTIN)
> >>>>> fprintf (stderr, "gimple builtin intrinsic not matched:%d %s
> >>>>> %s\n",
> >>>>> fn_code, fn_name1, fn_name2);
> >>>>> break;
>
