Re: [PATCH] constructor: Elide expand_constructor when can move by pieces is true

[Richard Biener <richard.guenther@gmail.com>]

On Thu, May 20, 2021 at 4:04 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>
> On Thu, May 20, 2021 at 12:51 AM Richard Biener
> <richard.guenther@gmail.com> wrote:
> >
> > On Wed, May 19, 2021 at 3:22 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> > >
> > > On Wed, May 19, 2021 at 2:33 AM Richard Biener
> > > <richard.guenther@gmail.com> wrote:
> > > >
> > > > On Tue, May 18, 2021 at 9:16 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> > > > >
> > > > > When expanding a constant constructor, don't call expand_constructor if
> > > > > it is more efficient to load the data from the memory via move by pieces.
> > > > >
> > > > > gcc/
> > > > >
> > > > >         PR middle-end/90773
> > > > >         * expr.c (expand_expr_real_1): Don't call expand_constructor if
> > > > >         it is more efficient to load the data from the memory.
> > > > >
> > > > > gcc/testsuite/
> > > > >
> > > > >         PR middle-end/90773
> > > > >         * gcc.target/i386/pr90773-24.c: New test.
> > > > >         * gcc.target/i386/pr90773-25.c: Likewise.
> > > > > ---
> > > > >  gcc/expr.c                                 | 10 ++++++++++
> > > > >  gcc/testsuite/gcc.target/i386/pr90773-24.c | 22 ++++++++++++++++++++++
> > > > >  gcc/testsuite/gcc.target/i386/pr90773-25.c | 20 ++++++++++++++++++++
> > > > >  3 files changed, 52 insertions(+)
> > > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr90773-24.c
> > > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr90773-25.c
> > > > >
> > > > > diff --git a/gcc/expr.c b/gcc/expr.c
> > > > > index d09ee42e262..80e01ea1cbe 100644
> > > > > --- a/gcc/expr.c
> > > > > +++ b/gcc/expr.c
> > > > > @@ -10886,6 +10886,16 @@ expand_expr_real_1 (tree exp, rtx target, machine_mode tmode,
> > > > >                 unsigned HOST_WIDE_INT ix;
> > > > >                 tree field, value;
> > > > >
> > > > > +               /* Check if it is more efficient to load the data from
> > > > > +                  the memory directly.  FIXME: How many stores do we
> > > > > +                  need here if not moved by pieces?  */
> > > > > +               unsigned HOST_WIDE_INT bytes
> > > > > +                 = tree_to_uhwi (TYPE_SIZE_UNIT (type));
> > > >
> > > > that's prone to fail - it could be a VLA.
> > >
> > > What do you mean by fail?  Is it ICE or missed optimization?
> > > Do you have a testcase?
> > >
> > > >
> > > > > +               if ((bytes / UNITS_PER_WORD) > 2
> > > > > +                   && MOVE_MAX_PIECES > UNITS_PER_WORD
> > > > > +                   && can_move_by_pieces (bytes, TYPE_ALIGN (type)))
> > > > > +                 goto normal_inner_ref;
> > > > > +
> > > >
> > > > It looks like you're concerned about aggregate copies but this also handles
> > > > non-aggregates (which on GIMPLE might already be optimized of course).
> > >
> > > Here I check if we copy more than 2 words and we can move more than
> > > a word in a single instruction.
> > >
> > > > Also you say "if it's cheaper" but I see no cost considerations.  How do
> > > > we generally handle immed const vs. load from constant pool costs?
> > >
> > > This trades 2 (update to 8) stores with one load plus one store.  Is there
> > > a way to check which one is faster?
> >
> > I'm not sure - it depends on whether the target can do stores from immediates
> > at all or what restrictions apply, what the immediate value actually is
> > (zero or all-ones should be way cheaper than sth arbitrary) and how the
> > pressure on the load unit is.  can_move_by_pieces (bytes, TYPE_ALIGN (type))
> > also does not guarantee it will actually move pieces larger than UNITS_PER_WORD,
> > that might depend on alignment.  There's by_pieces_ninsns that might provide
> > some hint here.
> >
> > I'm sure it works well for x86.
> >
> > I wonder if the existing code is in the appropriate place and we
> > shouldn't instead
> > handle this somewhere upthread where we ask to copy 'exp' into some other
> > memory location.  For your testcase that's expand_assignment but I can
> > imagine passing array[0] by value to a function resulting in similar copying.
> > Testing that shows we get
> >
> >         pushq   array+56(%rip)
> >         .cfi_def_cfa_offset 24
> >         pushq   array+48(%rip)
> >         .cfi_def_cfa_offset 32
> >         pushq   array+40(%rip)
> >         .cfi_def_cfa_offset 40
> >         pushq   array+32(%rip)
> >         .cfi_def_cfa_offset 48
> >         pushq   array+24(%rip)
> >         .cfi_def_cfa_offset 56
> >         pushq   array+16(%rip)
> >         .cfi_def_cfa_offset 64
> >         pushq   array+8(%rip)
> >         .cfi_def_cfa_offset 72
> >         pushq   array(%rip)
> >         .cfi_def_cfa_offset 80
> >         call    bar
> >
> > for that.  We do have the by-pieces infrastructure to generally do this kind of
> > copying but in both of these cases we do not seem to use it.  I also wonder
> > if the by-pieces infrastructure can pick up constant initializers automagically
> > (we could native_encode the initializer part and feed the by-pieces
> > infrastructure with an array of bytes).  There for example might be easy to
> > immediate-store byte parts and difficult ones where we could decide on a
> > case-by-case basis whether to load+store or immediate-store them.
>
> I opened:
>
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100704
>
> > For example if I change your testcase to have the array[] initializer
> > all-zero we currently emit
> >
> >         pxor    %xmm0, %xmm0
> >         movups  %xmm0, (%rdi)
> >         movups  %xmm0, 16(%rdi)
> >         movups  %xmm0, 32(%rdi)
> >         movups  %xmm0, 48(%rdi)
> >         ret
> >
> > will your patch cause us to emit 4 loads?  OTHO if I do
> >
> > const struct S array[] = {
> >   { 0, 0, 0, 7241, 124764, 48, 16, 33, 10, 96, 2, 0, 0, 4 }
> > };
> >
> > we get
> >
> >         movq    $0, (%rdi)
> >         movl    $0, 8(%rdi)
> >         movl    $0, 12(%rdi)
> >         movl    $7241, 16(%rdi)
> > ...
> >
> > ideally we'd have sth like
> >
> >     pxor %xmm0, %xmm0
> >     movups  %xmm0, (%rdi)
> >     movaps array+16(%rip), %xmm0
> >     movups %xmm0, 16(%rdi)
> > ...
> >
> > thus have the zeros written as immediates and the remaining pieces
> > with load+stores.
> >
> > The by-pieces infrastructure eventually get's to see
> >
> > (mem/u/c:BLK (symbol_ref:DI ("array") [flags 0x2] <var_decl
> > 0x7ffff7ff5b40 array>) [1 array+0 S64 A256])
> >
> > where the MEM_EXPR should provide a way to access the constant initializer.
> >
> > That said I do agree the current code is a bit premature optimization
> > - but maybe
> > it should be fend off in expand_constructor which has the cheap clear_storage
> > first and which already does check can_move_by_pieces with some heuristics,
> > but that seems to be guarded by
> >
> >            || (tree_fits_uhwi_p (TYPE_SIZE_UNIT (type))
> >                && (! can_move_by_pieces
> >                    (tree_to_uhwi (TYPE_SIZE_UNIT (type)),
> >                     TYPE_ALIGN (type)))
> >                && ! mostly_zeros_p (exp))))
> >
> > which is odd (we _can_ move by pieces, but how does this apply to
> > TREE_CONSTANT CTORs and avoid_temp_mem?).
> >
> > That said, I wonder if we want to elide expand_constructor when the
> > CTOR is TREE_STATIC && TREE_CONSTANT and !mostly_zeros_p
> > and we can_move_by_pieces.
> >
> > So sth like
> >
> > diff --git a/gcc/expr.c b/gcc/expr.c
> > index 7139545d543..76b3bdf0c01 100644
> > --- a/gcc/expr.c
> > +++ b/gcc/expr.c
> > @@ -8504,6 +8504,12 @@ expand_constructor (tree exp, rtx target, enum
> > expand_modifier modifier,
> >                && (! can_move_by_pieces
> >                    (tree_to_uhwi (TYPE_SIZE_UNIT (type)),
> >                     TYPE_ALIGN (type)))
> > +              && ! mostly_zeros_p (exp))
> > +          || (TREE_CONSTANT (exp)
> > +              && tree_fits_uhwi_p (TYPE_SIZE_UNIT (type))
> > +              && (can_move_by_pieces
> > +                  (tree_to_uhwi (TYPE_SIZE_UNIT (type)),
> > +                   TYPE_ALIGN (type)))
> >                && ! mostly_zeros_p (exp))))
> >        || ((modifier == EXPAND_INITIALIZER || modifier == EXPAND_CONST_ADDRESS)
> >           && TREE_CONSTANT (exp)))
> >
> > which handles your initializer and the all-zero one optimal?
> >
>
> It works.  Here is the updated patch.

So just looking at the code again I think we probably want to add
&& avoid_temp_mem here, at least that's the case we're looking
at.  Not sure if we ever arrive with TREE_CONSTANT CTORs
and !avoid_temp_mem but if so we'd create a temporary here
which of course would be pointless.

So maybe it's then clearer to split the condition out as

diff --git a/gcc/expr.c b/gcc/expr.c
index 7139545d543..ee8f25f9abd 100644
--- a/gcc/expr.c
+++ b/gcc/expr.c
@@ -8523,6 +8523,19 @@ expand_constructor (tree exp, rtx target, enum
expand_modifier modifier,
       return constructor;
     }

+  /* If the CTOR is available in static storage and not mostly
+     zeros and we can move it by pieces prefer to do so since
+     that's usually more efficient than performing a series of
+     stores from immediates.  */
+  if (avoid_temp_mem
+      && TREE_STATIC (exp)
+      && TREE_CONSTANT (exp)
+      && tree_fits_uhwi_p (TYPE_SIZE_UNIT (type))
+      && can_move_by_pieces (tree_to_uhwi (TYPE_SIZE_UNIT (type)),
+                            TYPE_ALIGN (type))
+      && ! mostly_zeros_p (exp))
+    return NULL_RTX;
+
   /* Handle calls that pass values in multiple non-contiguous
      locations.  The Irix 6 ABI has examples of this.  */
   if (target == 0 || ! safe_from_p (target, exp, 1)


OK with that change.

Thanks,
Richard.

> Thanks.
>
> --
> H.J.