[rs6000] (v2) Gimple folding of splat_uX

Message ID 1512752906.11602.139.camel@brimstone.rchland.ibm.com
State New
Headers show
Series
  • [rs6000] (v2) Gimple folding of splat_uX
Related show

Commit Message

will schmidt Dec. 8, 2017, 5:08 p.m.
Hi,
Add support for gimple folding of splat_u{8,16,32}.
Testcase coverage is primarily handled by existing tests
testsuite/gcc.target/powerpc/fold-vec-splat_*.c
    
One new test added to verify we continue to receive
an 'invalid argument, must be a 5-bit immediate' error
when we try to splat a non-constant value.
    
V2 updates include..
  Use the gimple_convert() helper.
  Use the build_vector_from_val() helper.
  whitespace fix-ups.
Those changes actually simplify the code here significantly, which is good. :-)
    
Per comments from last iteration, regarding the early exit for
arg0 != INTEGER_CST, if I remove the early exit when arg0 is not constant,
the code appears to handle things OK, and the generated code looks
reasonable, but this is a behavior change with respect to when folding is
disabled, which is to error out with an ("argument 1 must be a 5-bit signed
literal") error.  For that reason, I'm planning to leave that check in place.
I'll defer to Bill/Segher/David, etc on whether it would be appropriate to
change that behavior.
    
This intrinsic only takes INTegral types, so no need to check for FLOAT_CST.
(The other vec_splat* intrinsics will deal more with floats, those are later
in my queue).
    
Regtests across assorted power systems look clean.
OK for trunk?
    
Thanks,
-Will
    
[gcc]

2017-12-08  Will Schmidt  <will_schmidt@vnet.ibm.com>
    
	* config/rs6000/rs6000.c (rs6000_gimple_fold_builtin): Add support for
	early folding of splat_u{8,16,32}.
    
[testsuite]

2017-12-08  Will Schmidt  <will_schmidt@vnet.ibm.com>

	* gcc.target/powerpc/fold-vec-splat-misc-invalid.c: New.

Comments

Bill Schmidt Dec. 8, 2017, 9:23 p.m. | #1
On Dec 8, 2017, at 11:08 AM, Will Schmidt <will_schmidt@vnet.ibm.com> wrote:
> 

> 

> Hi,

> Add support for gimple folding of splat_u{8,16,32}.

> Testcase coverage is primarily handled by existing tests

> testsuite/gcc.target/powerpc/fold-vec-splat_*.c

> 

> One new test added to verify we continue to receive

> an 'invalid argument, must be a 5-bit immediate' error

> when we try to splat a non-constant value.

> 

> V2 updates include..

>  Use the gimple_convert() helper.

>  Use the build_vector_from_val() helper.

>  whitespace fix-ups.

> Those changes actually simplify the code here significantly, which is good. :-)

> 

> Per comments from last iteration, regarding the early exit for

> arg0 != INTEGER_CST, if I remove the early exit when arg0 is not constant,

> the code appears to handle things OK, and the generated code looks

> reasonable, but this is a behavior change with respect to when folding is

> disabled, which is to error out with an ("argument 1 must be a 5-bit signed

> literal") error.  For that reason, I'm planning to leave that check in place.

> I'll defer to Bill/Segher/David, etc on whether it would be appropriate to

> change that behavior.


RIght -- the long-documented behavior of these built-ins requires a constant
value.  Until such time as we change that documentation, if we feel it's a
good idea to do so, I would like to error out for non-constant values.

Thanks,
Bill

> 

> This intrinsic only takes INTegral types, so no need to check for FLOAT_CST.

> (The other vec_splat* intrinsics will deal more with floats, those are later

> in my queue).

> 

> Regtests across assorted power systems look clean.

> OK for trunk?

> 

> Thanks,

> -Will

> 

> [gcc]

> 

> 2017-12-08  Will Schmidt  <will_schmidt@vnet.ibm.com>

> 

> 	* config/rs6000/rs6000.c (rs6000_gimple_fold_builtin): Add support for

> 	early folding of splat_u{8,16,32}.

> 

> [testsuite]

> 

> 2017-12-08  Will Schmidt  <will_schmidt@vnet.ibm.com>

> 

> 	* gcc.target/powerpc/fold-vec-splat-misc-invalid.c: New.

> 

> diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c

> index 045a014..8972d80 100644

> --- a/gcc/config/rs6000/rs6000.c

> +++ b/gcc/config/rs6000/rs6000.c

> @@ -16614,10 +16614,32 @@ rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)

>     case VSX_BUILTIN_CMPLE_2DI:

>     case VSX_BUILTIN_CMPLE_U2DI:

>       fold_compare_helper (gsi, LE_EXPR, stmt);

>       return true;

> 

> +    /* flavors of vec_splat_[us]{8,16,32}.  */

> +    case ALTIVEC_BUILTIN_VSPLTISB:

> +    case ALTIVEC_BUILTIN_VSPLTISH:

> +    case ALTIVEC_BUILTIN_VSPLTISW:

> +      {

> +	 arg0 = gimple_call_arg (stmt, 0);

> +	 lhs = gimple_call_lhs (stmt);

> +	 /* Only fold the vec_splat_*() if arg0 is constant.  */

> +	 if (TREE_CODE (arg0) != INTEGER_CST)

> +	   return false;

> +	 gimple_seq stmts = NULL;

> +	 location_t loc = gimple_location (stmt);

> +	 tree splat_value = gimple_convert (&stmts, loc,

> +					    TREE_TYPE (TREE_TYPE (lhs)), arg0);

> +	 gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);

> +	 tree splat_tree = build_vector_from_val (TREE_TYPE (lhs), splat_value);

> +	 g = gimple_build_assign (lhs, splat_tree);

> +	 gimple_set_location (g, gimple_location (stmt));

> +	 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;

> diff --git a/gcc/testsuite/gcc.target/powerpc/fold-vec-splat-misc-invalid.c b/gcc/testsuite/gcc.target/powerpc/fold-vec-splat-misc-invalid.c

> new file mode 100644

> index 0000000..20f5b05

> --- /dev/null

> +++ b/gcc/testsuite/gcc.target/powerpc/fold-vec-splat-misc-invalid.c

> @@ -0,0 +1,33 @@

> +/* Verify that overloaded built-ins for vec_splat_s8 and vec_splat_s16

> +   generate errors as expected when we attempt to use invalid inputs.  */

> +

> +/* { dg-do compile } */

> +/* { dg-require-effective-target powerpc_vsx_ok } */

> +/* { dg-options "-mvsx -O2" } */

> +

> +#include <altivec.h>

> +

> +vector signed short

> +testss_1 (unsigned int ui)

> +{

> +  return vec_splat_s16 (ui);/* { dg-error "argument 1 must be a 5-bit signed literal" } */

> +}

> +

> +vector unsigned short

> +testss_2 (signed int si)

> +{

> +  return vec_splat_u16 (si);/* { dg-error "argument 1 must be a 5-bit signed literal" } */

> +}

> +

> +vector signed char

> +testsc_1 (unsigned int ui)

> +{

> +  return vec_splat_s8 (ui); /* { dg-error "argument 1 must be a 5-bit signed literal" } */

> +}

> +

> +vector unsigned char

> +testsc_2 (signed int si)

> +{

> +  return vec_splat_u8 (si);/* { dg-error "argument 1 must be a 5-bit signed literal" } */

> +}

> +

> 

>
Segher Boessenkool Dec. 11, 2017, 4:47 p.m. | #2
Hi!

On Fri, Dec 08, 2017 at 11:08:26AM -0600, Will Schmidt wrote:
> Add support for gimple folding of splat_u{8,16,32}.

> Testcase coverage is primarily handled by existing tests

> testsuite/gcc.target/powerpc/fold-vec-splat_*.c

>     

> One new test added to verify we continue to receive

> an 'invalid argument, must be a 5-bit immediate' error

> when we try to splat a non-constant value.

>     

> V2 updates include..

>   Use the gimple_convert() helper.

>   Use the build_vector_from_val() helper.

>   whitespace fix-ups.

> Those changes actually simplify the code here significantly, which is good. :-)


:-)

> 2017-12-08  Will Schmidt  <will_schmidt@vnet.ibm.com>

>     

> 	* config/rs6000/rs6000.c (rs6000_gimple_fold_builtin): Add support for

> 	early folding of splat_u{8,16,32}.

>     

> [testsuite]

> 

> 2017-12-08  Will Schmidt  <will_schmidt@vnet.ibm.com>

> 

> 	* gcc.target/powerpc/fold-vec-splat-misc-invalid.c: New.



> +    /* flavors of vec_splat_[us]{8,16,32}.  */

> +    case ALTIVEC_BUILTIN_VSPLTISB:

> +    case ALTIVEC_BUILTIN_VSPLTISH:

> +    case ALTIVEC_BUILTIN_VSPLTISW:

> +      {

> +	 arg0 = gimple_call_arg (stmt, 0);


The indent here is wrong (should be two spaces, is three).

Looks fine otherwise.  Okay for trunk with that fixed.  Thanks!


Segher

Patch

diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
index 045a014..8972d80 100644
--- a/gcc/config/rs6000/rs6000.c
+++ b/gcc/config/rs6000/rs6000.c
@@ -16614,10 +16614,32 @@  rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)
     case VSX_BUILTIN_CMPLE_2DI:
     case VSX_BUILTIN_CMPLE_U2DI:
       fold_compare_helper (gsi, LE_EXPR, stmt);
       return true;
 
+    /* flavors of vec_splat_[us]{8,16,32}.  */
+    case ALTIVEC_BUILTIN_VSPLTISB:
+    case ALTIVEC_BUILTIN_VSPLTISH:
+    case ALTIVEC_BUILTIN_VSPLTISW:
+      {
+	 arg0 = gimple_call_arg (stmt, 0);
+	 lhs = gimple_call_lhs (stmt);
+	 /* Only fold the vec_splat_*() if arg0 is constant.  */
+	 if (TREE_CODE (arg0) != INTEGER_CST)
+	   return false;
+	 gimple_seq stmts = NULL;
+	 location_t loc = gimple_location (stmt);
+	 tree splat_value = gimple_convert (&stmts, loc,
+					    TREE_TYPE (TREE_TYPE (lhs)), arg0);
+	 gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
+	 tree splat_tree = build_vector_from_val (TREE_TYPE (lhs), splat_value);
+	 g = gimple_build_assign (lhs, splat_tree);
+	 gimple_set_location (g, gimple_location (stmt));
+	 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;
diff --git a/gcc/testsuite/gcc.target/powerpc/fold-vec-splat-misc-invalid.c b/gcc/testsuite/gcc.target/powerpc/fold-vec-splat-misc-invalid.c
new file mode 100644
index 0000000..20f5b05
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/fold-vec-splat-misc-invalid.c
@@ -0,0 +1,33 @@ 
+/* Verify that overloaded built-ins for vec_splat_s8 and vec_splat_s16
+   generate errors as expected when we attempt to use invalid inputs.  */
+
+/* { dg-do compile } */
+/* { dg-require-effective-target powerpc_vsx_ok } */
+/* { dg-options "-mvsx -O2" } */
+
+#include <altivec.h>
+
+vector signed short
+testss_1 (unsigned int ui)
+{
+  return vec_splat_s16 (ui);/* { dg-error "argument 1 must be a 5-bit signed literal" } */
+}
+
+vector unsigned short
+testss_2 (signed int si)
+{
+  return vec_splat_u16 (si);/* { dg-error "argument 1 must be a 5-bit signed literal" } */
+}
+
+vector signed char
+testsc_1 (unsigned int ui)
+{
+  return vec_splat_s8 (ui); /* { dg-error "argument 1 must be a 5-bit signed literal" } */
+}
+
+vector unsigned char
+testsc_2 (signed int si)
+{
+  return vec_splat_u8 (si);/* { dg-error "argument 1 must be a 5-bit signed literal" } */
+}
+