Re: Don't query the frontend for unsupported types

[Richard Sandiford <richard.sandiford@linaro.org>]

Richard Biener <richard.guenther@gmail.com> writes:
> On September 20, 2017 2:36:03 PM GMT+02:00, Richard Sandiford
> <richard.sandiford@linaro.org> wrote:
>>When forcing a constant of mode MODE into memory, force_const_mem
>>asks the frontend to provide the type associated with that mode.
>>In principle type_for_mode is allowed to return null, and although
>>one use site correctly handled that, the other didn't.
>>
>>I think there's agreement that it's bogus to use type_for_mode for
>>this kind of thing, since it forces frontends to handle types that
>>don't exist in that language.  See e.g. http://gcc.gnu.org/PR46805
>>where the Go frontend was forced to handle vector types even though
>>Go doesn't have vector types.
>>
>>Also, the frontends use code like:
>>
>>  else if (VECTOR_MODE_P (mode))
>>    {
>>      machine_mode inner_mode = GET_MODE_INNER (mode);
>>      tree inner_type = c_common_type_for_mode (inner_mode, unsignedp);
>>      if (inner_type != NULL_TREE)
>>        return build_vector_type_for_mode (inner_type, mode);
>>    }
>>
>>and there's no guarantee that every vector mode M used by backend
>>rtl has an associated vector type whose TYPE_MODE is M.  I think
>>really the type_for_mode hook should only return trees that _do_ have
>>the requested TYPE_MODE, but PR46805 linked above shows that this is
>>likely to have too many knock-on consequences.  It doesn't make sense
>>for force_const_mem to ask about vector modes that aren't valid for
>>vector types, so this patch handles the condition there instead.
>>
>>This is needed for SVE multi-register modes, which are modelled as
>>vector modes but are not usable as vector types.
>>
>>Tested on aarch64-linux-gnu, x86_64-linux-gnu and
>>powerpc64le-linus-gnu.
>>OK to install?
>
> I think we should get rid of the use entirely.

I first read this as not using type_for_mode at all in force_const_mem,
which sounded like a good thing :-)  I tried it overnight on the usual
at-least-one-target-per-CPU set and diffing the before and after
assembly for the testsuite.  And it looks like i686 relies on this
to get an alignment of 16 rather than 4 for XFmode constants:
GET_MODE_ALIGNMENT (XFmode) == 32 (as requested by i386-modes.def),
but i386's CONSTANT_ALIGNMENT increases it to 128 for static constants.

But now I wonder if you meant we should just get rid of:

  set_mem_attributes (def, lang_hooks.types.type_for_mode (mode, 0), 1);

and keep the other call to type_for_mode, as below.

Thanks,
Richard


2017-09-21  Richard Sandiford  <richard.sandiford@linaro.org>
	    Alan Hayward  <alan.hayward@arm.com>
	    David Sherwood  <david.sherwood@arm.com>

gcc/
	* varasm.c (force_const_mem): Don't ask the front end about
	vector modes that are not supported as vector types by the target.
	Remove call to set_mem_attributes.

Index: gcc/varasm.c
===================================================================
--- gcc/varasm.c	2017-09-21 11:17:14.726201207 +0100
+++ gcc/varasm.c	2017-09-21 13:54:22.209159021 +0100
@@ -3785,10 +3785,17 @@ force_const_mem (machine_mode mode, rtx
   desc = ggc_alloc<constant_descriptor_rtx> ();
   *slot = desc;
 
+  tree type = NULL_TREE;
+  if (mode != VOIDmode
+      /* Don't ask the frontend about vector modes if there cannot be a
+	 VECTOR_TYPE whose TYPE_MODE is MODE.  */
+      && (!VECTOR_MODE_P (mode)
+	  || targetm.vector_mode_supported_p (mode)))
+    type = lang_hooks.types.type_for_mode (mode, 0);
+
   /* Align the location counter as required by EXP's data type.  */
   align = GET_MODE_ALIGNMENT (mode == VOIDmode ? word_mode : mode);
 
-  tree type = lang_hooks.types.type_for_mode (mode, 0);
   if (type != NULL_TREE)
     align = CONSTANT_ALIGNMENT (make_tree (type, x), align);
 
@@ -3832,7 +3839,6 @@ force_const_mem (machine_mode mode, rtx
 
   /* Construct the MEM.  */
   desc->mem = def = gen_const_mem (mode, symbol);
-  set_mem_attributes (def, lang_hooks.types.type_for_mode (mode, 0), 1);
   set_mem_align (def, align);
 
   /* If we're dropping a label to the constant pool, make sure we