[nvptx] Force vl32 if calling vector-partitionable routines

With PTX_MAX_VECTOR_LENGTH set to larger than PTX_WARP_SIZE, routines can be
called from offloading regions with vector-size set to larger than warp size.
OTOH, vector-partitionable routines assume warp-sized vector length.

Detect if we're calling a vector-partitionable routine from an offloading
region, and if so, fall back to warp-sized vector length in that region.

2019-01-07  Tom de Vries  <tdevries@suse.de>

	PR target/85486
	* config/nvptx/nvptx.c (has_vector_partitionable_routine_calls_p): New
	function.
	(nvptx_goacc_validate_dims): Force vl32 if calling vector-partitionable
	routines.

From-SVN: r267640

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 3c7bb497cf4c59f59e5584ebf1a78498babd1df7..ae2e85cae6aca343827d2e7fbabadcd21e1d5d20 100644 (file)
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,11 @@
+2019-01-07  Tom de Vries  <tdevries@suse.de>
+
+       PR target/85486
+       * config/nvptx/nvptx.c (has_vector_partitionable_routine_calls_p): New
+       function.
+       (nvptx_goacc_validate_dims): Force vl32 if calling vector-partitionable
+       routines.
+
 2019-01-07  Jakub Jelinek  <jakub@redhat.com>
 
        * config/i386/sse.md (vec_extract<mode><ssehalfvecmodelower>): Use

diff --git a/gcc/config/nvptx/nvptx.c b/gcc/config/nvptx/nvptx.c
index 5a4b38de5225b19b55a5a15a10158cbabe404c40..7fdc285b6f8eba0c16c337a3619523be5d75aed1 100644 (file)
--- a/gcc/config/nvptx/nvptx.c
+++ b/gcc/config/nvptx/nvptx.c
@@ -59,6 +59,7 @@
 #include "builtins.h"
 #include "omp-general.h"
 #include "omp-low.h"
+#include "omp-offload.h"
 #include "gomp-constants.h"
 #include "dumpfile.h"
 #include "internal-fn.h"
@@ -5496,6 +5497,40 @@ nvptx_apply_dim_limits (int dims[])
     dims[GOMP_DIM_VECTOR] = PTX_WARP_SIZE;
 }
 
+/* Return true if FNDECL contains calls to vector-partitionable routines.  */
+
+static bool
+has_vector_partitionable_routine_calls_p (tree fndecl)
+{
+  if (!fndecl)
+    return false;
+
+  basic_block bb;
+  FOR_EACH_BB_FN (bb, DECL_STRUCT_FUNCTION (fndecl))
+    for (gimple_stmt_iterator i = gsi_start_bb (bb); !gsi_end_p (i);
+        gsi_next_nondebug (&i))
+      {
+       gimple *stmt = gsi_stmt (i);
+       if (gimple_code (stmt) != GIMPLE_CALL)
+         continue;
+
+       tree callee = gimple_call_fndecl (stmt);
+       if (!callee)
+         continue;
+
+       tree attrs  = oacc_get_fn_attrib (callee);
+       if (attrs == NULL_TREE)
+         return false;
+
+       int partition_level = oacc_fn_attrib_level (attrs);
+       bool seq_routine_p = partition_level == GOMP_DIM_MAX;
+       if (!seq_routine_p)
+         return true;
+      }
+
+  return false;
+}
+
 /* As nvptx_goacc_validate_dims, but does not return bool to indicate whether
    DIMS has changed.  */
 
@@ -5611,6 +5646,16 @@ nvptx_goacc_validate_dims_1 (tree decl, int dims[], int fn_level)
     old_dims[i] = dims[i];
 
   const char *vector_reason = NULL;
+  if (offload_region_p && has_vector_partitionable_routine_calls_p (decl))
+    {
+      if (dims[GOMP_DIM_VECTOR] > PTX_WARP_SIZE)
+       {
+         vector_reason = G_("using vector_length (%d) due to call to"
+                            " vector-partitionable routine, ignoring %d");
+         dims[GOMP_DIM_VECTOR] = PTX_WARP_SIZE;
+       }
+    }
+
   if (dims[GOMP_DIM_VECTOR] == 0)
     {
       vector_reason = G_("using vector_length (%d), ignoring runtime setting");