[gcc/devel/omp/gcc-11] openmp: Improve expand_omp_atomic_pipeline

[gcc/devel/omp/gcc-11] openmp: Improve expand_omp_atomic_pipeline

[Tobias Burnus]

https://gcc.gnu.org/g:afe01c57c1c71a0e26a2676860eb3101281d80b5

commit afe01c57c1c71a0e26a2676860eb3101281d80b5
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Fri Sep 3 10:15:17 2021 +0200

    openmp: Improve expand_omp_atomic_pipeline
    
    When __atomic_* builtins were introduced, omp-expand.c (omp-low.c
    at that point) has been adjusted in several spots so that it uses
    the atomic builtins instead of sync builtins, but
    expand_omp_atomic_pipeline has not because the __atomic_compare_exchange_*
    APIs take address of the argument, so it kept using __sync_val_compare_swap_*.
    That means it always uses seq_cst though.
    This patch changes it to use the ATOMIC_COMPARE_EXCHANGE ifn which gimple-fold
    folds __atomic_compare_exchange_* into - that ifn also passes expected
    directly.
    
    2021-09-03  Jakub Jelinek  <jakub@redhat.com>
    
            * omp-expand.c (expand_omp_atomic_pipeline): Use
            IFN_ATOMIC_COMPARE_EXCHANGE instead of
            BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_? so that memory order
            can be provided.
    
    (cherry picked from commit 090f0d78f194e3cda23fe904016db77ea36c38fa)

Diff:
---
 gcc/ChangeLog.omp | 78 ++++++++++++++++++++++++++++++-------------------------
 gcc/omp-expand.c  | 21 ++++++++++++---
 2 files changed, 61 insertions(+), 38 deletions(-)

diff --git a/gcc/ChangeLog.omp b/gcc/ChangeLog.omp
index 59678d62af4..b4231380da4 100644
--- a/gcc/ChangeLog.omp
+++ b/gcc/ChangeLog.omp
@@ -1,49 +1,57 @@
+2021-09-03  Tobias Burnus  <tobias@codesourcery.com>
+
+	2021-09-03  Jakub Jelinek  <jakub@redhat.com>
+
+	* omp-expand.c (expand_omp_atomic_pipeline): Use
+	IFN_ATOMIC_COMPARE_EXCHANGE instead of
+	BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_? so that memory order
+	can be provided.
+
 2021-09-02  Marcel Vollweiler  <marcel@codesourcery.com>
 
 	Backported from master:
 	2021-09-02  Marcel Vollweiler  <marcel@codesourcery.com>
 
-        * gfortran.dg/gomp/target-device-ancestor-4.f90: Comment out dg-final to avoid
-        UNRESOLVED.
+	* gfortran.dg/gomp/target-device-ancestor-4.f90: Comment out dg-final to avoid
+	UNRESOLVED.
 
 2021-08-31  Marcel Vollweiler  <marcel@codesourcery.com>
 
 	Backported from master:
 	2021-08-31  Marcel Vollweiler  <marcel@codesourcery.com>
 
-        * c-parser.c (c_parser_omp_clause_device): Parse device-modifiers 'device_num'
-        and 'ancestor' in 'target device' clauses.
-        * parser.c (cp_parser_omp_clause_device): Parse device-modifiers 'device_num'
-        and 'ancestor' in 'target device' clauses.
-        * semantics.c (finish_omp_clauses): Error handling. Constant device ids must
-        evaluate to '1' if 'ancestor' is used.
-        * gfortran.h: Add variable for 'ancestor' in struct gfc_omp_clauses.
-        * openmp.c (gfc_match_omp_clauses): Parse device-modifiers 'device_num'
-        and 'ancestor' in 'target device' clauses.
-        * trans-openmp.c (gfc_trans_omp_clauses): Set OMP_CLAUSE_DEVICE_ANCESTOR.
-        * gimplify.c (gimplify_scan_omp_clauses): Error handling. 'ancestor' only
-        allowed on target constructs and only with particular other clauses.
-        * omp-expand.c (expand_omp_target): Output of 'sorry, not supported' if
-        'ancestor' is used.
-        * omp-low.c (check_omp_nesting_restrictions): Error handling. No nested OpenMP
-        structs when 'ancestor' is used.
-        (scan_omp_1_stmt): No usage of OpenMP runtime routines in a target region when
-        'ancestor' is used.
-        * tree-pretty-print.c (dump_omp_clause): Append 'ancestor'.
-        * tree.h (OMP_CLAUSE_DEVICE_ANCESTOR): Define macro.
-        * c-c++-common/gomp/target-device-1.c: New test.
-        * c-c++-common/gomp/target-device-2.c: New test.
-        * c-c++-common/gomp/target-device-ancestor-1.c: New test.
-        * c-c++-common/gomp/target-device-ancestor-2.c: New test.
-        * c-c++-common/gomp/target-device-ancestor-3.c: New test.
-        * c-c++-common/gomp/target-device-ancestor-4.c: New test.
-        * gfortran.dg/gomp/target-device-1.f90: New test.
-        * gfortran.dg/gomp/target-device-2.f90: New test.
-        * gfortran.dg/gomp/target-device-ancestor-1.f90: New test.
-        * gfortran.dg/gomp/target-device-ancestor-2.f90: New test.
-        * gfortran.dg/gomp/target-device-ancestor-3.f90: New test.
-        * gfortran.dg/gomp/target-device-ancestor-4.f90: New test.
-
+	* c-parser.c (c_parser_omp_clause_device): Parse device-modifiers 'device_num'
+	and 'ancestor' in 'target device' clauses.
+	* parser.c (cp_parser_omp_clause_device): Parse device-modifiers 'device_num'
+	and 'ancestor' in 'target device' clauses.
+	* semantics.c (finish_omp_clauses): Error handling. Constant device ids must
+	evaluate to '1' if 'ancestor' is used.
+	* gfortran.h: Add variable for 'ancestor' in struct gfc_omp_clauses.
+	* openmp.c (gfc_match_omp_clauses): Parse device-modifiers 'device_num'
+	and 'ancestor' in 'target device' clauses.
+	* trans-openmp.c (gfc_trans_omp_clauses): Set OMP_CLAUSE_DEVICE_ANCESTOR.
+	* gimplify.c (gimplify_scan_omp_clauses): Error handling. 'ancestor' only
+	allowed on target constructs and only with particular other clauses.
+	* omp-expand.c (expand_omp_target): Output of 'sorry, not supported' if
+	'ancestor' is used.
+	* omp-low.c (check_omp_nesting_restrictions): Error handling. No nested OpenMP
+	structs when 'ancestor' is used.
+	(scan_omp_1_stmt): No usage of OpenMP runtime routines in a target region when
+	'ancestor' is used.
+	* tree-pretty-print.c (dump_omp_clause): Append 'ancestor'.
+	* tree.h (OMP_CLAUSE_DEVICE_ANCESTOR): Define macro.
+	* c-c++-common/gomp/target-device-1.c: New test.
+	* c-c++-common/gomp/target-device-2.c: New test.
+	* c-c++-common/gomp/target-device-ancestor-1.c: New test.
+	* c-c++-common/gomp/target-device-ancestor-2.c: New test.
+	* c-c++-common/gomp/target-device-ancestor-3.c: New test.
+	* c-c++-common/gomp/target-device-ancestor-4.c: New test.
+	* gfortran.dg/gomp/target-device-1.f90: New test.
+	* gfortran.dg/gomp/target-device-2.f90: New test.
+	* gfortran.dg/gomp/target-device-ancestor-1.f90: New test.
+	* gfortran.dg/gomp/target-device-ancestor-2.f90: New test.
+	* gfortran.dg/gomp/target-device-ancestor-3.f90: New test.
+	* gfortran.dg/gomp/target-device-ancestor-4.f90: New test.
 
 2021-08-23  Tobias Burnus  <tobias@codesourcery.com>
 
diff --git a/gcc/omp-expand.c b/gcc/omp-expand.c
index 9cd75a91649..6c538ebfbb1 100644
--- a/gcc/omp-expand.c
+++ b/gcc/omp-expand.c
@@ -8857,8 +8857,6 @@ expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb,
   edge e;
   enum built_in_function fncode;
 
-  /* ??? We need a non-pointer interface to __atomic_compare_exchange in
-     order to use the RELAXED memory model effectively.  */
   fncode = (enum built_in_function)((int)BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_N
 				    + index + 1);
   cmpxchg = builtin_decl_explicit (fncode);
@@ -8875,6 +8873,15 @@ expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb,
   /* Load the initial value, replacing the GIMPLE_OMP_ATOMIC_LOAD.  */
   si = gsi_last_nondebug_bb (load_bb);
   gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD);
+  location_t loc = gimple_location (gsi_stmt (si));
+  enum omp_memory_order omo = gimple_omp_atomic_memory_order (gsi_stmt (si));
+  enum memmodel imo = omp_memory_order_to_memmodel (omo);
+  tree mo = build_int_cst (NULL, imo);
+  if (imo == MEMMODEL_RELEASE)
+    imo = MEMMODEL_RELAXED;
+  else if (imo == MEMMODEL_ACQ_REL)
+    imo = MEMMODEL_ACQUIRE;
+  tree fmo = build_int_cst (NULL, imo);
 
   /* For floating-point values, we'll need to view-convert them to integers
      so that we can perform the atomic compare and swap.  Simplify the
@@ -8971,7 +8978,15 @@ expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb,
 				  GSI_SAME_STMT);
 
   /* Build the compare&swap statement.  */
-  new_storedi = build_call_expr (cmpxchg, 3, iaddr, loadedi, storedi);
+  tree ctype = build_complex_type (itype);
+  int flag = int_size_in_bytes (itype);
+  new_storedi = build_call_expr_internal_loc (loc, IFN_ATOMIC_COMPARE_EXCHANGE,
+					      ctype, 6, iaddr, loadedi,
+					      storedi,
+					      build_int_cst (integer_type_node,
+							     flag),
+					      mo, fmo);
+  new_storedi = build1 (REALPART_EXPR, itype, new_storedi);
   new_storedi = force_gimple_operand_gsi (&si,
 					  fold_convert (TREE_TYPE (loadedi),
 							new_storedi),