Fix reduction failures.

Signed-off-by: Dominik Adamski <Dominik.Adamski@amd.com>

tools/flang2/flang2exe/exp_ftn.cpp

@@ -4147,7 +4147,11 @@ exp_bran(ILM_OP opc, ILM *ilmp, int curilm)
 /***************************************************************/
 
 void
-exp_misc(ILM_OP opc, ILM *ilmp, int curilm, bool process_expanded)
+exp_misc(ILM_OP opc,
+         ILM *ilmp,
+         int curilm,
+         bool process_expanded,
+         bool is_reduction_kernel)
 {
   int tmp;
   int ilix, listilix;
@@ -4289,11 +4293,11 @@ exp_misc(ILM_OP opc, ILM *ilmp, int curilm, bool process_expanded)
     break;
 
   case IM_ENDF:
-    exp_end(ilmp, curilm, true, process_expanded);
+    exp_end(ilmp, curilm, true, process_expanded, is_reduction_kernel);
     break;
 
   case IM_END:
-    exp_end(ilmp, curilm, false, process_expanded);
+    exp_end(ilmp, curilm, false, process_expanded, is_reduction_kernel);
     break;
 
   case IM_BYVAL:

tools/flang2/flang2exe/exp_ftn.h

@@ -53,7 +53,11 @@ void exp_bran(ILM_OP opc, ILM *ilmp, int curilm);
 /**
    \brief ...
  */
-void exp_misc(ILM_OP opc, ILM *ilmp, int curilm, bool process_expanded = false);
+void exp_misc(ILM_OP opc,
+              ILM *ilmp,
+              int curilm,
+              bool process_expanded = false,
+              bool is_reduction_kernel = false);
 
 /**
    \brief ...

tools/flang2/flang2exe/exp_rte.cpp

@@ -2139,7 +2139,7 @@ exp_alloca(ILM *ilmp)
 static void gen_funcret(finfo_t *);
 
 void
-exp_end(ILM *ilmp, int curilm, bool is_func, bool process_expanded)
+exp_end(ILM *ilmp, int curilm, bool is_func, bool process_expanded, bool is_reduction)
 {
   int tmp;
   int op1;
@@ -2159,7 +2159,7 @@ exp_end(ILM *ilmp, int curilm, bool is_func, bool process_expanded)
   if (flg.omptarget && !is_func) {
     if (XBIT(232, 0x40) && gbl.ompaccel_intarget && !OMPACCFUNCDEVG(gbl.currsub) /*is_gpu_output_file() */  ) {
       OMP_TARGET_MODE mode = ompaccel_tinfo_get(gbl.currsub)->mode;
-      if (!is_SPMD_mode(mode) && !process_expanded) {
+      if (is_reduction || (!is_SPMD_mode(mode) && !process_expanded)) {
         ilix = ll_make_kmpc_target_deinit(mode);
         iltb.callfg = 1;
         chk_block(ilix);

tools/flang2/flang2exe/exp_rte.h

@@ -100,7 +100,11 @@ void exp_cgoto(ILM *ilmp, int curilm);
 /**
    \brief ...
  */
-void exp_end(ILM *ilmp, int curilm, bool is_func, bool process_expanded = false);
+void exp_end(ILM *ilmp,
+             int curilm,
+             bool is_func,
+             bool process_expanded = false,
+             bool is_reduction = false);
 
 /**
    \brief ...

tools/flang2/flang2exe/expand.cpp

@@ -63,6 +63,7 @@ static int efunc(const char *);
 static int create_ref(SPTR sym, int *pnmex, int basenm, int baseilix,
                       int *pclen, int *pmxlen, int *prestype);
 static int jsr2qjsr(int);
+static bool is_reduction_kernel;
 
 SPTR
 eval_ilm_check_if_skip(int ilmx, int *skip_expand = nullptr, int *process_expanded = nullptr);
@@ -218,6 +219,42 @@ parse_im_file(const ILM *ilmp, int *lineno_out, int *findex_out, int *ftag_out)
     *ftag_out = ftag;
 }
 
+bool target_region_contains_reduction()
+{
+  int curr_expb_nilms = expb.nilms;
+  int curr_pos = ilmb.globalilmcount;
+  int curr_ilmpos = get_gilmb_ilmpos();
+  ILM *ilmp;      /* absolute pointer to the ILM */
+  ILM_OP opc;     /* opcode of the ILM		 */
+  int len;
+  do {
+    expb.nilms = rdilms();
+    if (expb.nilms > expb.ilmb.stg_size) {
+      EXP_MORE(expb.ilmb, ILM_AUX, expb.nilms + 100);
+    }
+    for (int ilmx = 0; ilmx < expb.nilms; ilmx += len) {
+      ilmp = (ILM *)(ilmb.ilm_base + ilmx);
+      opc = ILM_OPC(ilmp);
+      len = ilms[opc].oprs + 1; /* length is number of words for the
+                                 * fixed operands and the opcode */
+      if (IM_VAR(opc))
+        len += ILM_OPND(ilmp, 1); /* include the number of
+                                   * variable operands */
+      if (opc == IM_MP_BREDUCTION) {
+	 expb.nilms = curr_expb_nilms;
+	 ilmb.globalilmcount = curr_pos;
+	 set_gilmb_ilmpos(curr_ilmpos);
+         return true;
+      }
+    }
+  }
+  while (opc != IM_ETARGET && opc != IM_END && opc != IM_ENDF);
+  expb.nilms = curr_expb_nilms;
+  ilmb.globalilmcount = curr_pos;
+  set_gilmb_ilmpos(curr_ilmpos);
+  return false;
+}
+
 /***************************************************************/
 /** \brief Expand ILMs to ILIs */
 int
@@ -237,6 +274,7 @@ expand(void)
   static std::map<int, int> process_expanded_map = std::map<int,int>();
   auto it = process_expanded_map.find(gbl.currsub);
   int process_expanded = 0;
+  is_reduction_kernel = false;
 
   // we reset flag because we do not know if we generate initialization
   // function for SPMD kernel (the function with kmpc_parallel_51 call)
@@ -740,9 +778,9 @@ eval_ilm_check_if_skip(int ilmx, int *skip_expand, int *process_expanded)
 
   case IMTY_MISC: /* miscellaneous  */
     if (process_expanded && *process_expanded)
-      exp_misc(opcx, ilmpx, ilmx, true);
+      exp_misc(opcx, ilmpx, ilmx, true, is_reduction_kernel);
     else
-      exp_misc(opcx, ilmpx, ilmx);
+      exp_misc(opcx, ilmpx, ilmx, false, is_reduction_kernel);
     break;
 
   case IMTY_FSTR: /* fortran string */
@@ -787,6 +825,7 @@ eval_ilm_check_if_skip(int ilmx, int *skip_expand, int *process_expanded)
     if (XBIT(232, 0x40) && gbl.ompaccel_intarget && !*process_expanded) {
       //TODO move initialization to separate function
       std::vector<int> allocated_symbols;
+      bool reduction_pragma = target_region_contains_reduction();
       if (is_SPMD_mode(ompaccel_tinfo_get(gbl.currsub)->mode)) {
 	  allocated_symbols = get_allocated_symbols(ompaccel_tinfo_get(gbl.currsub));
       }
@@ -814,8 +853,10 @@ eval_ilm_check_if_skip(int ilmx, int *skip_expand, int *process_expanded)
        * will be located  */
       RFCNTI(target_code_lab);
       exp_label(target_code_lab);
+      if (reduction_pragma)
+        is_reduction_kernel = true;
 
-      if (is_SPMD_mode(ompaccel_tinfo_get(gbl.currsub)->mode)) {
+      if (is_SPMD_mode(ompaccel_tinfo_get(gbl.currsub)->mode) && !reduction_pragma) {
         ilix = ll_make_kmpc_global_thread_num();
         iltb.callfg = 1;
         chk_block(ilix);

tools/flang2/flang2exe/ilmutil.cpp

@@ -2071,3 +2071,13 @@ count_ilms()
   return newnumilms;
 }
 
+int get_gilmb_ilmpos()
+{
+  return gilmb.ilmpos;
+}
+
+void set_gilmb_ilmpos(int val)
+{
+  gilmb.ilmpos = val;
+}
+

tools/flang2/flang2exe/ilmutil.h

@@ -248,4 +248,14 @@ void swap_next_gilm(void);
  */
 void wrilms(int linenum);
 
+/**
+   \brief return gbl.ilmpos value
+ */
+int get_gilmb_ilmpos();
+
+/**
+   \brief set gilmb.ilmpos value
+ */
+void set_gilmb_ilmpos(int value);
+
 #endif // ILMUTIL_H_

tools/shared/expand.h

@@ -357,4 +357,6 @@ void replace_by_zero(ILM_OP opc, ILM *ilmp, int curilm);
  */
 void set_assn(int nme);
 
+bool target_region_contains_reduction();
+
 #endif