Be stricter about CONST_VECTOR operands

Be stricter about CONST_VECTOR operands

[Richard Sandiford]

The recent gen_vec_duplicate patches used CONST_VECTOR for all
constants, but the documentation says:

  @findex const_vector
  @item (const_vector:@var{m} [@var{x0} @var{x1} @dots{}])
  Represents a vector constant.  The square brackets stand for the vector
  containing the constant elements.  @var{x0}, @var{x1} and so on are
  the @code{const_int}, @code{const_double} or @code{const_fixed} elements.

Both the AArch32 and AArch64 ports relied on the elements having
this form and would ICE if the element was something like a CONST
instead.  This showed up as a failure in vect-102.c for both arm-eabi
and aarch64-elf (but not aarch64-linux-gnu, which is what the series
was tested on).

The two obvious options were to redefine CONST_VECTOR to accept all
constants or make gen_vec_duplicate honour the existing documentation.
It looks like other code also assumes that integer CONST_VECTORs contain
CONST_INTs, so the patch does the latter.

I deliberately didn't add an assert to gen_const_vec_duplicate
because it looks like the SPU port *does* expect to be able to create
CONST_VECTORs of symbolic constants.

Also, I think the list above should include const_wide_int for vectors
of TImode and wider.

The new routine takes a mode for consistency with the generators,
and because I think it does make sense to accept all constants for
variable-length:

    (const (vec_duplicate ...))

rather than have some rtxes for which we instead use:

    (vec_duplicate (const ...))

Tested on aarch64-elf, aarch64-linux-gnu, x86_64-linux-gnu and
powerpc64-linux-gnu.  OK to install?

Richard


2017-11-06  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
	* doc/rtl.texi (const_vector): Say that elements can be
	const_wide_ints too.
	* emit-rtl.h (valid_for_const_vec_duplicate_p): Declare.
	* emit-rtl.c (valid_for_const_vec_duplicate_p): New function.
	(gen_vec_duplicate): Use it instead of CONSTANT_P.
	* optabs.c (expand_vector_broadcast): Likewise.

Index: gcc/doc/rtl.texi
===================================================================
--- gcc/doc/rtl.texi	2017-11-06 08:56:27.608223621 +0000
+++ gcc/doc/rtl.texi	2017-11-06 09:01:44.150672938 +0000
@@ -1625,7 +1625,8 @@ accessed with @code{CONST_FIXED_VALUE_LO
 @item (const_vector:@var{m} [@var{x0} @var{x1} @dots{}])
 Represents a vector constant.  The square brackets stand for the vector
 containing the constant elements.  @var{x0}, @var{x1} and so on are
-the @code{const_int}, @code{const_double} or @code{const_fixed} elements.
+the @code{const_int}, @code{const_wide_int}, @code{const_double} or
+@code{const_fixed} elements.
 
 The number of units in a @code{const_vector} is obtained with the macro
 @code{CONST_VECTOR_NUNITS} as in @code{CONST_VECTOR_NUNITS (@var{v})}.
Index: gcc/emit-rtl.h
===================================================================
--- gcc/emit-rtl.h	2017-11-06 08:56:27.157323659 +0000
+++ gcc/emit-rtl.h	2017-11-06 09:01:44.151672938 +0000
@@ -438,6 +438,7 @@ get_max_uid (void)
   return crtl->emit.x_cur_insn_uid;
 }
 
+extern bool valid_for_const_vec_duplicate_p (machine_mode, rtx);
 extern rtx gen_const_vec_duplicate (machine_mode, rtx);
 extern rtx gen_vec_duplicate (machine_mode, rtx);
 
Index: gcc/emit-rtl.c
===================================================================
--- gcc/emit-rtl.c	2017-11-06 08:56:27.608223621 +0000
+++ gcc/emit-rtl.c	2017-11-06 09:01:44.151672938 +0000
@@ -5757,6 +5757,17 @@ init_emit (void)
 #endif
 }
 
+/* Return true if X is a valid element for a duplicated vector constant
+   of the given mode.  */
+
+bool
+valid_for_const_vec_duplicate_p (machine_mode, rtx x)
+{
+  return (CONST_SCALAR_INT_P (x)
+	  || CONST_DOUBLE_AS_FLOAT_P (x)
+	  || CONST_FIXED_P (x));
+}
+
 /* Like gen_const_vec_duplicate, but ignore const_tiny_rtx.  */
 
 static rtx
@@ -5792,7 +5803,7 @@ gen_const_vec_duplicate (machine_mode mo
 rtx
 gen_vec_duplicate (machine_mode mode, rtx x)
 {
-  if (CONSTANT_P (x))
+  if (valid_for_const_vec_duplicate_p (mode, x))
     return gen_const_vec_duplicate (mode, x);
   return gen_rtx_VEC_DUPLICATE (mode, x);
 }
Index: gcc/optabs.c
===================================================================
--- gcc/optabs.c	2017-11-06 08:56:27.466923633 +0000
+++ gcc/optabs.c	2017-11-06 09:01:44.152672938 +0000
@@ -377,7 +377,7 @@ expand_vector_broadcast (machine_mode vm
 
   gcc_checking_assert (VECTOR_MODE_P (vmode));
 
-  if (CONSTANT_P (op))
+  if (valid_for_const_vec_duplicate_p (vmode, op))
     return gen_const_vec_duplicate (vmode, op);
 
   /* ??? If the target doesn't have a vec_init, then we have no easy way

Re: Be stricter about CONST_VECTOR operands

[James Greenhalgh]

On Mon, Nov 06, 2017 at 09:10:23AM +0000, Richard Sandiford wrote:
> The recent gen_vec_duplicate patches used CONST_VECTOR for all
> constants, but the documentation says:
> 
>   @findex const_vector
>   @item (const_vector:@var{m} [@var{x0} @var{x1} @dots{}])
>   Represents a vector constant.  The square brackets stand for the vector
>   containing the constant elements.  @var{x0}, @var{x1} and so on are
>   the @code{const_int}, @code{const_double} or @code{const_fixed} elements.
> 
> Both the AArch32 and AArch64 ports relied on the elements having
> this form and would ICE if the element was something like a CONST
> instead.  This showed up as a failure in vect-102.c for both arm-eabi
> and aarch64-elf (but not aarch64-linux-gnu, which is what the series
> was tested on).
> 
> The two obvious options were to redefine CONST_VECTOR to accept all
> constants or make gen_vec_duplicate honour the existing documentation.
> It looks like other code also assumes that integer CONST_VECTORs contain
> CONST_INTs, so the patch does the latter.
> 
> I deliberately didn't add an assert to gen_const_vec_duplicate
> because it looks like the SPU port *does* expect to be able to create
> CONST_VECTORs of symbolic constants.
> 
> Also, I think the list above should include const_wide_int for vectors
> of TImode and wider.
> 
> The new routine takes a mode for consistency with the generators,
> and because I think it does make sense to accept all constants for
> variable-length:
> 
>     (const (vec_duplicate ...))
> 
> rather than have some rtxes for which we instead use:
> 
>     (vec_duplicate (const ...))
> 
> Tested on aarch64-elf, aarch64-linux-gnu, x86_64-linux-gnu and
> powerpc64-linux-gnu.  OK to install?

For what it is worth, I can see why this would fix the bug in the AArch64
port, and the patch looks good to me (though I can't approve it).

> (but not aarch64-linux-gnu, which is what the series was tested on).

Presumably it would fail there for -mcmodel=tiny too, we just don't run
that variant by default.

Thanks,
James

> 2017-11-06  Richard Sandiford  <richard.sandiford@linaro.org>
> 
> gcc/
> 	* doc/rtl.texi (const_vector): Say that elements can be
> 	const_wide_ints too.
> 	* emit-rtl.h (valid_for_const_vec_duplicate_p): Declare.
> 	* emit-rtl.c (valid_for_const_vec_duplicate_p): New function.
> 	(gen_vec_duplicate): Use it instead of CONSTANT_P.
> 	* optabs.c (expand_vector_broadcast): Likewise.

Re: Be stricter about CONST_VECTOR operands

[Jeff Law]

On 11/06/2017 02:10 AM, Richard Sandiford wrote:
> The recent gen_vec_duplicate patches used CONST_VECTOR for all
> constants, but the documentation says:
> 
>   @findex const_vector
>   @item (const_vector:@var{m} [@var{x0} @var{x1} @dots{}])
>   Represents a vector constant.  The square brackets stand for the vector
>   containing the constant elements.  @var{x0}, @var{x1} and so on are
>   the @code{const_int}, @code{const_double} or @code{const_fixed} elements.
> 
> Both the AArch32 and AArch64 ports relied on the elements having
> this form and would ICE if the element was something like a CONST
> instead.  This showed up as a failure in vect-102.c for both arm-eabi
> and aarch64-elf (but not aarch64-linux-gnu, which is what the series
> was tested on).
> 
> The two obvious options were to redefine CONST_VECTOR to accept all
> constants or make gen_vec_duplicate honour the existing documentation.
> It looks like other code also assumes that integer CONST_VECTORs contain
> CONST_INTs, so the patch does the latter.
> 
> I deliberately didn't add an assert to gen_const_vec_duplicate
> because it looks like the SPU port *does* expect to be able to create
> CONST_VECTORs of symbolic constants.
> 
> Also, I think the list above should include const_wide_int for vectors
> of TImode and wider.
> 
> The new routine takes a mode for consistency with the generators,
> and because I think it does make sense to accept all constants for
> variable-length:
> 
>     (const (vec_duplicate ...))
> 
> rather than have some rtxes for which we instead use:
> 
>     (vec_duplicate (const ...))
> 
> Tested on aarch64-elf, aarch64-linux-gnu, x86_64-linux-gnu and
> powerpc64-linux-gnu.  OK to install?
> 
> Richard
> 
> 
> 2017-11-06  Richard Sandiford  <richard.sandiford@linaro.org>
> 
> gcc/
> 	* doc/rtl.texi (const_vector): Say that elements can be
> 	const_wide_ints too.
> 	* emit-rtl.h (valid_for_const_vec_duplicate_p): Declare.
> 	* emit-rtl.c (valid_for_const_vec_duplicate_p): New function.
> 	(gen_vec_duplicate): Use it instead of CONSTANT_P.
> 	* optabs.c (expand_vector_broadcast): Likewise.
OK.
jeff