diff mbox series

[rs6000] Finish implementation of __builtin_atlivec_lvx_v1ti

Message ID 8365537f-20b3-6616-0180-3713bc0db215@linux.vnet.ibm.com
State New
Headers show
Series [rs6000] Finish implementation of __builtin_atlivec_lvx_v1ti | expand

Commit Message

Kelvin Nilsen March 14, 2018, 9:27 p.m. UTC 
During code review, it was discovered that the implementation of
__builtin_altivec_lvx_v1ti is not complete.  The constant
ALTIVEC_BUILTINLVX_V1TI is introduced and is bound to the function
__builtin_altivec_lvx_v1ti.  However, this function's implementation is
incomplete because there is no call to the def_builtin function for this
binding.

This patch provides the missing pieces to add support for this function.
Additionally, this patch introduces four new __int128-based prototypes
of the overloaded __builtin_vec_ld function.  This is the function that
implements the vec_ld () macro expansion.  A new test case has been
provided to exercise each of these prototypes.

This patch has been bootstrapped and tested without regressions on both
powerpc64le-unknown-linux (P8) and on powerpc-linux (P7 big-endian, with
both -m32 and -m64 target options).

Is this patch ok for trunk?

gcc/ChangeLog:

2018-03-14  Kelvin Nilsen  <kelvin@gcc.gnu.org>

	* config/rs6000/rs6000-c.c (altivec_overloaded_builtins): Add
	entries for V1TI variants of __builtin_altivec_ld builtin.
	* config/rs6000/rs6000.c (altivec_expand_lv_builtin): Add test and
	handling of V1TI variant of LVX icode pattern.
	(altivec_expand_builtin): Add case for ALTIVEC_BUILTIN_LVX_V1TI.
	(rs6000_gimple_fold_builtin): Likewise.
	(altivec_init_builtins): Add code to define
	__builtin_altivec_lvx_v1ti function.
	* doc/extend.texi: Add four new prototypes for vec_ld.

gcc/testsuite/ChangeLog:

2018-03-14  Kelvin Nilsen  <kelvin@gcc.gnu.org>

	* gcc.target/powerpc/altivec-ld-1.c: New test.

Comments

Segher Boessenkool March 17, 2018, 12:07 a.m. UTC | #1 
On Wed, Mar 14, 2018 at 04:27:31PM -0500, Kelvin Nilsen wrote:
> During code review, it was discovered that the implementation of
> __builtin_altivec_lvx_v1ti is not complete.  The constant
> ALTIVEC_BUILTINLVX_V1TI is introduced and is bound to the function
> __builtin_altivec_lvx_v1ti.  However, this function's implementation is
> incomplete because there is no call to the def_builtin function for this
> binding.
> 
> This patch provides the missing pieces to add support for this function.
> Additionally, this patch introduces four new __int128-based prototypes
> of the overloaded __builtin_vec_ld function.  This is the function that
> implements the vec_ld () macro expansion.  A new test case has been
> provided to exercise each of these prototypes.

Do we want/need stvx_v1ti as well?

> --- gcc/config/rs6000/rs6000.c	(revision 258341)
> +++ gcc/config/rs6000/rs6000.c	(working copy)
> @@ -14452,6 +14452,7 @@ altivec_expand_lv_builtin (enum insn_code icode, t
>       LVXL and LVE*X expand to use UNSPECs to hide their special behavior,
>       so the raw address is fine.  */
>    if (icode == CODE_FOR_altivec_lvx_v2df_2op
> +      || icode == CODE_FOR_altivec_lvx_v1ti_2op
>        || icode == CODE_FOR_altivec_lvx_v2di_2op
>        || icode == CODE_FOR_altivec_lvx_v4sf_2op
>        || icode == CODE_FOR_altivec_lvx_v4si_2op

That's a weird ordering; put v1ti first?

> @@ -15811,6 +15812,9 @@ altivec_expand_builtin (tree exp, rtx target, bool
>      case ALTIVEC_BUILTIN_LVX_V2DI:
>        return altivec_expand_lv_builtin (CODE_FOR_altivec_lvx_v2di_2op,
>  					exp, target, false);
> +    case ALTIVEC_BUILTIN_LVX_V1TI:
> +      return altivec_expand_lv_builtin (CODE_FOR_altivec_lvx_v1ti_2op,
> +					exp, target, false);
>      case ALTIVEC_BUILTIN_LVX_V4SF:
>        return altivec_expand_lv_builtin (CODE_FOR_altivec_lvx_v4sf_2op,
>  					exp, target, false);

Same here.

> --- gcc/testsuite/gcc.target/powerpc/altivec-ld-1.c	(nonexistent)
> +++ gcc/testsuite/gcc.target/powerpc/altivec-ld-1.c	(working copy)
> @@ -0,0 +1,67 @@
> +/* { dg-do run { target powerpc*-*-* } } */
> +/* { dg-require-effective-target powerpc_altivec_ok } */
> +/* { dg-require-effective-target lp64 } */
> +/* { dg-options "-maltivec -O0 -Wall -Wno-deprecated" } */

Since this is a run testcase you need vmx_hw instead of powerpc_altivec_ok.
You also need a test for int128.  I don't think you need lp64, please check.

Okay for trunk with that fixed.  Thanks!


Segher
Peter Bergner March 20, 2018, 5:36 p.m. UTC | #2 
On 3/14/18 4:27 PM, Kelvin Nilsen wrote:
> @@ -14452,6 +14452,7 @@ altivec_expand_lv_builtin (enum insn_code icode, t
>       LVXL and LVE*X expand to use UNSPECs to hide their special behavior,
>       so the raw address is fine.  */
>    if (icode == CODE_FOR_altivec_lvx_v2df_2op
> +      || icode == CODE_FOR_altivec_lvx_v1ti_2op
>        || icode == CODE_FOR_altivec_lvx_v2di_2op
>        || icode == CODE_FOR_altivec_lvx_v4sf_2op
>        || icode == CODE_FOR_altivec_lvx_v4si_2op
> @@ -15811,6 +15812,9 @@ altivec_expand_builtin (tree exp, rtx target, bool
>      case ALTIVEC_BUILTIN_LVX_V2DI:
>        return altivec_expand_lv_builtin (CODE_FOR_altivec_lvx_v2di_2op,
>  					exp, target, false);
> +    case ALTIVEC_BUILTIN_LVX_V1TI:
> +      return altivec_expand_lv_builtin (CODE_FOR_altivec_lvx_v1ti_2op,
> +					exp, target, false);
>      case ALTIVEC_BUILTIN_LVX_V4SF:
>        return altivec_expand_lv_builtin (CODE_FOR_altivec_lvx_v4sf_2op,
>  					exp, target, false);

FYI, these hunks will need updating due to the fix for PR83789 which
I just committed to trunk.

Peter
Peter Bergner March 21, 2018, 2:37 p.m. UTC | #3 
On 3/14/18 4:27 PM, Kelvin Nilsen wrote:
> 2018-03-14  Kelvin Nilsen  <kelvin@gcc.gnu.org>
>
> 	* config/rs6000/rs6000-c.c (altivec_overloaded_builtins): Add
> 	entries for V1TI variants of __builtin_altivec_ld builtin.
> 	* config/rs6000/rs6000.c (altivec_expand_lv_builtin): Add test and
> 	handling of V1TI variant of LVX icode pattern.
> 	(altivec_expand_builtin): Add case for ALTIVEC_BUILTIN_LVX_V1TI.
> 	(rs6000_gimple_fold_builtin): Likewise.
> 	(altivec_init_builtins): Add code to define
> 	__builtin_altivec_lvx_v1ti function.
> 	* doc/extend.texi: Add four new prototypes for vec_ld.
>
> gcc/testsuite/ChangeLog:
>
> 2018-03-14  Kelvin Nilsen  <kelvin@gcc.gnu.org>
>
> 	* gcc.target/powerpc/altivec-ld-1.c: New test.

You'll also need to add:

	PR target/84760

...to both ChangeLog entries.

Peter
diff mbox series

Patch

Index: gcc/config/rs6000/rs6000-c.c
===================================================================
--- gcc/config/rs6000/rs6000-c.c	(revision 258341)
+++ gcc/config/rs6000/rs6000-c.c	(working copy)
@@ -1562,6 +1562,15 @@  const struct altivec_builtin_types altivec_overloa
   { VSX_BUILTIN_VEC_FLOATO, VSX_BUILTIN_UNS_FLOATO_V2DI,
     RS6000_BTI_V4SF, RS6000_BTI_unsigned_V2DI, 0, 0 },
 
+  { ALTIVEC_BUILTIN_VEC_LD, ALTIVEC_BUILTIN_LVX_V1TI,
+    RS6000_BTI_V1TI, RS6000_BTI_INTSI, ~RS6000_BTI_V1TI, 0 },
+  { ALTIVEC_BUILTIN_VEC_LD, ALTIVEC_BUILTIN_LVX_V1TI,
+    RS6000_BTI_unsigned_V1TI, RS6000_BTI_INTSI, ~RS6000_BTI_unsigned_V1TI, 0 },
+  { ALTIVEC_BUILTIN_VEC_LD, ALTIVEC_BUILTIN_LVX_V1TI,
+    RS6000_BTI_V1TI, RS6000_BTI_INTSI, ~RS6000_BTI_INTTI, 0 },
+  { ALTIVEC_BUILTIN_VEC_LD, ALTIVEC_BUILTIN_LVX_V1TI,
+    RS6000_BTI_unsigned_V1TI, RS6000_BTI_INTSI, ~RS6000_BTI_UINTTI, 0 },
+
   { ALTIVEC_BUILTIN_VEC_LD, ALTIVEC_BUILTIN_LVX_V2DF,
     RS6000_BTI_V2DF, RS6000_BTI_INTSI, ~RS6000_BTI_V2DF, 0 },
   { ALTIVEC_BUILTIN_VEC_LD, ALTIVEC_BUILTIN_LVX_V2DI,
Index: gcc/config/rs6000/rs6000.c
===================================================================
--- gcc/config/rs6000/rs6000.c	(revision 258341)
+++ gcc/config/rs6000/rs6000.c	(working copy)
@@ -14452,6 +14452,7 @@  altivec_expand_lv_builtin (enum insn_code icode, t
      LVXL and LVE*X expand to use UNSPECs to hide their special behavior,
      so the raw address is fine.  */
   if (icode == CODE_FOR_altivec_lvx_v2df_2op
+      || icode == CODE_FOR_altivec_lvx_v1ti_2op
       || icode == CODE_FOR_altivec_lvx_v2di_2op
       || icode == CODE_FOR_altivec_lvx_v4sf_2op
       || icode == CODE_FOR_altivec_lvx_v4si_2op
@@ -15811,6 +15812,9 @@  altivec_expand_builtin (tree exp, rtx target, bool
     case ALTIVEC_BUILTIN_LVX_V2DI:
       return altivec_expand_lv_builtin (CODE_FOR_altivec_lvx_v2di_2op,
 					exp, target, false);
+    case ALTIVEC_BUILTIN_LVX_V1TI:
+      return altivec_expand_lv_builtin (CODE_FOR_altivec_lvx_v1ti_2op,
+					exp, target, false);
     case ALTIVEC_BUILTIN_LVX_V4SF:
       return altivec_expand_lv_builtin (CODE_FOR_altivec_lvx_v4sf_2op,
 					exp, target, false);
@@ -16542,6 +16546,7 @@  rs6000_gimple_fold_builtin (gimple_stmt_iterator *
     case ALTIVEC_BUILTIN_LVX_V4SF:
     case ALTIVEC_BUILTIN_LVX_V2DI:
     case ALTIVEC_BUILTIN_LVX_V2DF:
+    case ALTIVEC_BUILTIN_LVX_V1TI:
       {
 	arg0 = gimple_call_arg (stmt, 0);  // offset
 	arg1 = gimple_call_arg (stmt, 1);  // address
@@ -17443,6 +17448,10 @@  altivec_init_builtins (void)
     = build_function_type_list (V2DI_type_node,
 				long_integer_type_node, pcvoid_type_node,
 				NULL_TREE);
+  tree v1ti_ftype_long_pcvoid
+    = build_function_type_list (V1TI_type_node,
+				long_integer_type_node, pcvoid_type_node,
+				NULL_TREE);
 
   tree void_ftype_opaque_long_pvoid
     = build_function_type_list (void_type_node,
@@ -17540,6 +17549,8 @@  altivec_init_builtins (void)
   def_builtin ("__builtin_altivec_lvx", v4si_ftype_long_pcvoid, ALTIVEC_BUILTIN_LVX);
   def_builtin ("__builtin_altivec_lvx_v2df", v2df_ftype_long_pcvoid,
 	       ALTIVEC_BUILTIN_LVX_V2DF);
+  def_builtin ("__builtin_altivec_lvx_v1ti", v1ti_ftype_long_pcvoid,
+	       ALTIVEC_BUILTIN_LVX_V1TI);
   def_builtin ("__builtin_altivec_lvx_v2di", v2di_ftype_long_pcvoid,
 	       ALTIVEC_BUILTIN_LVX_V2DI);
   def_builtin ("__builtin_altivec_lvx_v4sf", v4sf_ftype_long_pcvoid,
Index: gcc/doc/extend.texi
===================================================================
--- gcc/doc/extend.texi	(revision 258341)
+++ gcc/doc/extend.texi	(working copy)
@@ -18396,6 +18396,10 @@  vector double vec_div (vector double, vector doubl
 vector long vec_div (vector long, vector long);
 vector unsigned long vec_div (vector unsigned long, vector unsigned long);
 vector double vec_floor (vector double);
+vector __int128 vec_ld (int, const vector __int128 *);
+vector unsigned __int128 vec_ld (int, const vector unsigned __int128 *);
+vector __int128 vec_ld (int, const __int128 *);
+vector unsigned __int128 vec_ld (int, const unsigned __int128 *);
 vector double vec_ld (int, const vector double *);
 vector double vec_ld (int, const double *);
 vector double vec_ldl (int, const vector double *);
Index: gcc/testsuite/gcc.target/powerpc/altivec-ld-1.c
===================================================================
--- gcc/testsuite/gcc.target/powerpc/altivec-ld-1.c	(nonexistent)
+++ gcc/testsuite/gcc.target/powerpc/altivec-ld-1.c	(working copy)
@@ -0,0 +1,67 @@ 
+/* { dg-do run { target powerpc*-*-* } } */
+/* { dg-require-effective-target powerpc_altivec_ok } */
+/* { dg-require-effective-target lp64 } */
+/* { dg-options "-maltivec -O0 -Wall -Wno-deprecated" } */
+
+#include <altivec.h>
+#include <stdlib.h>
+
+static __vector __int128 v;
+static __vector __int128 *pv;
+
+static __vector unsigned __int128 uv;
+static __vector unsigned __int128 *puv;
+
+static __int128 i128;
+static __int128 *pi128;
+
+static unsigned __int128 u128;
+static unsigned __int128 *pu128;
+
+
+void
+doInitialization ()
+{
+  v[0] = -1;
+  pv = &v;
+
+  uv[0] = 0xcafebabe;
+  puv = &uv;
+
+  i128 = 0xfabeabe;
+  pi128 = &i128;
+
+  u128 = 0xabefabe;
+  pu128 = &u128;
+}
+
+int
+main (int argc, char *argv[])
+{
+  __vector __int128 loaded_v;
+  __vector unsigned __int128 loaded_uv;
+
+  /* Usage:
+   *  <Type> result = vec_ld (int index, __vector <Type> v)
+   * is equivalent to:
+   *  result = v [index];
+   */
+  doInitialization ();
+  loaded_v = vec_ld (0, pv);
+  if (loaded_v[0] != -1)
+    abort ();
+
+  loaded_uv = vec_ld (0, puv);
+  if (loaded_uv[0] != 0xcafebabe)
+    abort ();
+
+  loaded_v = vec_ld (0, pi128);
+  if (loaded_v[0] != 0xfabeabe)
+    abort ();
+
+  loaded_uv = vec_ld (0, pu128);
+  if (loaded_uv[0] != 0xabefabe)
+    abort ();
+
+  return 0;
+}