On 01/30/2017 02:18 AM, Thomas Schwinge wrote:
> Hi Cesar!
>
> On Fri, 27 Jan 2017 07:45:52 -0800, Cesar Philippidis
> <ce...@codesourcery.com> wrote:
>> If you take a close look at lower_omp_target, you'll notice that I'm
>> gave reference types special treatment. Specifically, I disabled this
>> optimization on non-INTEGER_TYPE and floating point values, because the
>> nvptx target was having some problems dereferencing boolean-typed
>> pointers. That's something I have on my TODO list to track down later.
>
> Please file an issue as appropriate.
I filed an issue for this internally.
>> As for the performance gains, this optimization resulted in a
>> non-trivial speedup in CloverLeaf running on a Nvidia Pascal board.
>> CloverLeaf is somewhat special in that it consists of a lot of OpenACC
>> offloaded regions which gets called multiple times throughout its
>> execution. Consequently, it is I/O limited. The other benchmarks I ran
>> didn't benefit nearly as much as CloverLeaf. I chose a small data set
>> for CloverLeaf that only ran in 1.3s without the patch, and hence make
>> it even more I/O limited. After the patch, it ran 0.35s faster.
>
> \o/ Yay!
>
>> This patch has been applied to gomp-4_0-branch.
>
> (Not reviewed in detail.)
>
>> --- a/gcc/omp-low.c
>> +++ b/gcc/omp-low.c
>
>> +static tree
>> +convert_from_firstprivate_pointer (tree var, bool is_ref, gimple_seq *gs)
>> +{
>> + tree type = TREE_TYPE (var);
>> + tree new_type = NULL_TREE;
>> + tree tmp = NULL_TREE;
>> + tree inner_type = NULL_TREE;
>
> [...]/source-gcc/gcc/omp-low.c: In function 'tree_node*
> convert_from_firstprivate_pointer(tree, bool, gimple**)':
> [...]/source-gcc/gcc/omp-low.c:16515:8: warning: unused variable
> 'inner_type' [-Wunused-variable]
>
>
>> --- /dev/null
>> +++ b/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90
>
> I see:
>
> {+FAIL: libgomp.oacc-fortran/firstprivate-int.f90
> -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O
> (internal compiler error)+}
> {+FAIL: libgomp.oacc-fortran/firstprivate-int.f90
> -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O 4
> blank line(s) in output+}
> {+FAIL: libgomp.oacc-fortran/firstprivate-int.f90
> -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O (test
> for excess errors)+}
> {+UNRESOLVED: libgomp.oacc-fortran/firstprivate-int.f90
> -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O
> compilation failed to produce executable+}
>
> That's the nvptx offloading compiler configured with
> "--enable-checking=yes,df,fold,rtl":
>
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:
> In function 'MAIN__._omp_fn.1':
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:55:0:
> error: conversion of register to a different size
> VIEW_CONVERT_EXPR<logical(kind=2)>(_17);
>
> _18 = VIEW_CONVERT_EXPR<logical(kind=2)>(_17);
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:55:0:
> error: conversion of register to a different size
> VIEW_CONVERT_EXPR<logical(kind=4)>(_20);
>
> _21 = VIEW_CONVERT_EXPR<logical(kind=4)>(_20);
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:55:0:
> error: conversion of register to a different size
> VIEW_CONVERT_EXPR<logical(kind=8)>(_23);
>
> _24 = VIEW_CONVERT_EXPR<logical(kind=8)>(_23);
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:55:0:
> error: conversion of register to a different size
> VIEW_CONVERT_EXPR<logical(kind=16)>(_26);
>
> _27 = VIEW_CONVERT_EXPR<logical(kind=16)>(_26);
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:55:0:
> internal compiler error: verify_gimple failed
> 0xa67d75 verify_gimple_in_cfg(function*, bool)
> [...]/source-gcc/gcc/tree-cfg.c:5125
> 0x94ebbc execute_function_todo
> [...]/source-gcc/gcc/passes.c:1958
> 0x94f513 execute_todo
> [...]/source-gcc/gcc/passes.c:2010
>
>
> And with "-m32" multilib testing, I see:
>
> {+FAIL: libgomp.oacc-fortran/firstprivate-int.f90
> -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable -O (test for
> excess errors)+}
> {+UNRESOLVED: libgomp.oacc-fortran/firstprivate-int.f90
> -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable -O
> compilation failed to produce executable+}
>
> That is:
>
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:10:18:
> Error: Kind 16 not supported for type INTEGER at (1)
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:16:18:
> Error: Kind 16 not supported for type LOGICAL at (1)
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:115:18:
> Error: Kind 16 not supported for type INTEGER at (1)
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:121:18:
> Error: Kind 16 not supported for type LOGICAL at (1)
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:31:6:
> Error: Symbol 'i16i' at (1) has no IMPLICIT type
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:49:40:
> Error: Symbol 'i16o' at (1) has no IMPLICIT type
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:37:6:
> Error: Symbol 'l16i' at (1) has no IMPLICIT type
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:51:40:
> Error: Symbol 'l16o' at (1) has no IMPLICIT type
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:105:43:
> Error: Symbol 'i16i' at (1) has no IMPLICIT type
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:105:69:
> Error: Symbol 'i16o' at (1) has no IMPLICIT type
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:106:43:
> Error: Symbol 'l16i' at (1) has no IMPLICIT type
>
> [...]/source-gcc/libgomp/testsuite/libgomp.oacc-fortran/firstprivate-int.f90:106:69:
> Error: Symbol 'l16o' at (1) has no IMPLICIT type
I should have been comparing the type size, not precision against the
pointer size. This patch fixes that.
Cesar
2017-02-01 Cesar Philippidis <ce...@codesourcery.com>
gcc/
* omp-low.c (convert_to_firstprivate_pointer): Use TYPE_SIZE instead
of TYPE_PRECISION when determining if a firstprivate variable may be
casted into a pointer.
(convert_from_firstprivate_pointer): Likewise.
(lower_omp_target): Likewise.
diff --git a/gcc/omp-low.c b/gcc/omp-low.c
index 142d928..450d76e 100644
--- a/gcc/omp-low.c
+++ b/gcc/omp-low.c
@@ -16480,7 +16480,7 @@ convert_to_firstprivate_pointer (tree var, gimple_seq *gs)
return fold_convert (pointer_sized_int_node, var);
}
- switch (TYPE_PRECISION (type))
+ switch (tree_to_uhwi (TYPE_SIZE (type)))
{
case 1: case 2: case 4: case 8: new_type = unsigned_char_type_node; break;
case 16: new_type = short_unsigned_type_node; break;
@@ -16520,7 +16520,7 @@ convert_from_firstprivate_pointer (tree var, bool is_ref, gimple_seq *gs)
if (INTEGRAL_TYPE_P (var) || POINTER_TYPE_P (type))
return fold_convert (type, var);
- switch (TYPE_PRECISION (type))
+ switch (tree_to_uhwi (TYPE_SIZE (type)))
{
case 1: case 2: case 4: case 8: new_type = unsigned_char_type_node; break;
case 16: new_type = short_unsigned_type_node; break;
@@ -16732,7 +16732,7 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
&& (TREE_CODE (inner_type) == REAL_TYPE
|| (!is_reference (var) && INTEGRAL_TYPE_P (inner_type))
|| TREE_CODE (inner_type) == INTEGER_TYPE)
- && TYPE_PRECISION (inner_type) <= POINTER_SIZE
+ && tree_to_uhwi (TYPE_SIZE (inner_type)) <= POINTER_SIZE
&& TYPE_PRECISION (inner_type) != 0
&& !maybe_lookup_field_in_outer_ctx (var, ctx))
oacc_firstprivate_int = true;
@@ -17012,7 +17012,7 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
|| (!is_reference (var)
&& INTEGRAL_TYPE_P (inner_type))
|| TREE_CODE (inner_type) == INTEGER_TYPE)
- && TYPE_PRECISION (inner_type) <= POINTER_SIZE
+ && tree_to_uhwi (TYPE_SIZE (inner_type)) <= POINTER_SIZE
&& TYPE_PRECISION (inner_type) != 0
&& !maybe_lookup_field_in_outer_ctx (var, ctx))
{
@@ -17200,7 +17200,7 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
if (is_reference (ovar))
type = TREE_TYPE (type);
if ((INTEGRAL_TYPE_P (type)
- && TYPE_PRECISION (type) <= POINTER_SIZE)
+ && tree_to_uhwi (TYPE_SIZE (type)) <= POINTER_SIZE)
|| TREE_CODE (type) == POINTER_TYPE)
{
tkind = GOMP_MAP_FIRSTPRIVATE_INT;
@@ -17355,7 +17355,7 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
if (is_reference (var))
type = TREE_TYPE (type);
if ((INTEGRAL_TYPE_P (type)
- && TYPE_PRECISION (type) <= POINTER_SIZE)
+ && tree_to_uhwi (TYPE_SIZE (type)) <= POINTER_SIZE)
|| TREE_CODE (type) == POINTER_TYPE)
{
x = build_receiver_ref (var, false, ctx);
