Flang OpenMP GPU offload fixes

Author: gklimowicz

include/flang/Error/errmsg-in.n

@@ -1516,3 +1516,4 @@ A DO CONCURRENT variable with LOCAL_INIT locality must have a host variable of t
 .MS S 1203 "OpenMP GPU - Directive target exit data is used, map type [$] cannot be used. It should be one of 'from', 'release', or 'delete'"
 .MS S 1204 "OpenMP GPU - Directive target data is used, map type [$] cannot be used. It should be one of 'to', 'from', 'tofrom', 'alloc', 'release', or 'delete'"
 .MS S 1205 "OpenMP GPU - Directive target enter data is used, map type [$] cannot be used. It should be one of 'to', or 'alloc'"
+.MS W 1206 "OpenMP GPU - Using [map] clause is used with an array section for $ array is not supported, mapping will be done on entire array."

tools/flang1/flang1exe/semsmp.c

@@ -4841,6 +4841,16 @@ semsmp(int rednum, SST *top)
    *	<accel data> ::= <accel data name> ( <accel sub list> ) |
    */
   case ACCEL_DATA1:
+#ifdef OMP_OFFLOAD_LLVM
+    if (SST_IDG(RHS(1)) == S_IDENT || SST_IDG(RHS(1)) == S_DERIVED) {
+      sptr = SST_SYMG(RHS(1));
+    } else {
+      sptr = SST_LSYMG(RHS(1));
+    }
+    error(1206, ERR_Warning, gbl.lineno, sptr ? SYMNAME(sptr) : CNULL, CNULL);
+    goto accel_data2;
+    break;
+#endif
   accel_data1:
     if (SST_IDG(RHS(1)) == S_IDENT || SST_IDG(RHS(1)) == S_DERIVED) {
       sptr = SST_SYMG(RHS(1));

tools/flang2/flang2exe/expsmp.cpp

@@ -2781,7 +2781,7 @@ exp_smp(ILM_OP opc, ILM *ilmp, int curilm)
 #ifdef OMP_OFFLOAD_LLVM
     case IM_MP_MAP:
       if (flg.omptarget) {
-        exp_ompaccel_map(ilmp, curilm);
+        exp_ompaccel_map(ilmp, curilm, outlinedCnt);
       }
       break;
     case IM_MP_TARGETLOOPTRIPCOUNT:

tools/flang2/flang2exe/llassem.cpp

@@ -1029,16 +1029,21 @@ ompaccel_write_sharedvars(void)
 }
 
 static void
-write_tgtrt_statics(SPTR sptr, char *gname, char *typed, int gblsym,
+write_libomptarget_statics(SPTR sptr, char *gname, char *typed, int gblsym,
                     DSRT *dsrtp)
 {
   char *linkage_type;
+
   linkage_type = "internal";
   sprintf(gname, "struct%s", getsname(sptr));
   get_typedef_ag(gname, typed);
   free(typed);
   gblsym = find_ag(gname);
   typed = AG_TYPENAME(gblsym);
+
+  process_ftn_dtype_struct(DTYPEG(sptr), typed, false);
+  write_struct_defs();
+
 #ifdef WEAKG
   if (WEAKG(sptr))
     linkage_type = "weak";
@@ -1055,20 +1060,25 @@ write_tgtrt_statics(SPTR sptr, char *gname, char *typed, int gblsym,
   fputc('\n', ASMFIL);
 }
 
-static bool isLibomptargetInit = false;
+static bool isOmptargetInitialized = false;
+
 void
 write_libomtparget(void)
 {
-  if (isLibomptargetInit)
-    return;
-  fprintf(ASMFIL, "\n; OpenMP GPU Offload Init\n\
-  @.omp_offloading.img_end.nvptx64-nvidia-cuda = external constant i8 \n\
-  @.omp_offloading.img_start.nvptx64-nvidia-cuda = external constant i8 \n\
-  @.omp_offloading.entries_end = external constant %%struct.__tgt_offload_entry \n\
-  @.omp_offloading.entries_begin = external constant %%struct.__tgt_offload_entry \n\
-  @.omp_offloading.device_images = internal unnamed_addr constant [1 x %%struct.__tgt_device_image] [%%struct.__tgt_device_image { i8* @.omp_offloading.img_start.nvptx64-nvidia-cuda, i8* @.omp_offloading.img_end.nvptx64-nvidia-cuda, %%struct.__tgt_offload_entry* @.omp_offloading.entries_begin, %%struct.__tgt_offload_entry* @.omp_offloading.entries_end }], align 8\n\
-  @.omp_offloading.descriptor_ = internal constant %%struct.__tgt_bin_desc { i64 1, %%struct.__tgt_device_image* getelementptr inbounds ([1 x %%struct.__tgt_device_image], [1 x %%struct.__tgt_device_image]* @.omp_offloading.device_images, i32 0, i32 0), %%struct.__tgt_offload_entry* @.omp_offloading.entries_begin, %%struct.__tgt_offload_entry* @.omp_offloading.entries_end }, align 8\n\n");
-  isLibomptargetInit = true;
+  /* These structs should be created just right after the first target region. */
+  if (!isOmptargetInitialized) {
+    if(!strcmp(SYMNAME(gbl.currsub), "ompaccel.register"))
+    {
+      fprintf(ASMFIL, "\n; OpenMP GPU Offload Init\n\
+@.omp_offloading.img_end.nvptx64-nvidia-cuda = external constant i8 \n\
+@.omp_offloading.img_start.nvptx64-nvidia-cuda = external constant i8 \n\
+@.omp_offloading.entries_end = external constant %%struct.__tgt_offload_entry_ \n\
+@.omp_offloading.entries_begin = external constant %%struct.__tgt_offload_entry_ \n\
+@.omp_offloading.device_images = internal unnamed_addr constant [1 x %%struct.__tgt_device_image] [%%struct.__tgt_device_image { i8* @.omp_offloading.img_start.nvptx64-nvidia-cuda, i8* @.omp_offloading.img_end.nvptx64-nvidia-cuda, %%struct.__tgt_offload_entry_* @.omp_offloading.entries_begin, %%struct.__tgt_offload_entry_* @.omp_offloading.entries_end }], align 8\n\
+@.omp_offloading.descriptor_ = internal constant %%struct.__tgt_bin_desc { i64 1, %%struct.__tgt_device_image* getelementptr inbounds ([1 x %%struct.__tgt_device_image], [1 x %%struct.__tgt_device_image]* @.omp_offloading.device_images, i32 0, i32 0), %%struct.__tgt_offload_entry_* @.omp_offloading.entries_begin, %%struct.__tgt_offload_entry_* @.omp_offloading.entries_end }, align 8\n\n");
+      isOmptargetInitialized = true;
+    }
+  }
 }
 
 #endif
@@ -1617,7 +1627,7 @@ write_statics(void)
     typed = get_struct_from_dsrt(sptr, dsrtp, SIZEG(sptr), &align8, true, 0);
 #ifdef OMP_OFFLOAD_LLVM
     if (OMPACCSTRUCTG(sptr)) {
-      write_tgtrt_statics(sptr, gname, typed, gblsym, dsrtp);
+      write_libomptarget_statics(sptr, gname, typed, gblsym, dsrtp);
       count--;
       continue;
     }
@@ -3715,13 +3725,6 @@ getsname(SPTR sptr)
       sprintf(name, "%s", SYMNAME(sptr));
       p = name;
     }
-#ifdef OMP_OFFLOAD_LLVM
-    if (gbl.isnvvmcodegen && STYPEG(sptr) == ST_PROC &&
-        strncmp(SYMNAME(sptr), "omp_get_", 8) == 0) {
-      sprintf(name, "%s", SYMNAME(sptr));
-      return name;
-    }
-#endif
     else if (gbl.internal && CONTAINEDG(sptr)) {
       p = name;
       if (gbl.outersub) {

tools/flang2/flang2exe/main.cpp

@@ -194,6 +194,8 @@ process_input(char *argv0, bool *need_cuda_constructor)
       interr("main: malloc_verify failsA", errno, ERR_Fatal);
 #endif
     reinit();
+
+
 #if DEBUG & sun
   if (DBGBIT(7, 4))
     if (malloc_verify() != 1)
@@ -243,12 +245,11 @@ process_input(char *argv0, bool *need_cuda_constructor)
     } else {
 #ifdef OMP_OFFLOAD_LLVM
       if (flg.omptarget) {
+        init_test();
         ompaccel_initsyms();
-        if (ompaccel_tinfo_has(gbl.currsub))
-          gbl.inomptarget = true;
-        else
-          gbl.inomptarget = false;
-        ompaccel_create_globalctor();
+        gbl.inomptarget = ompaccel_tinfo_has(gbl.currsub);
+        if(!DTYPEG(gbl.currsub))
+          ompaccel_create_globalctor();
         ompaccel_create_reduction_wrappers();
       }
 #endif
@@ -319,9 +320,6 @@ process_input(char *argv0, bool *need_cuda_constructor)
 #ifdef OMP_OFFLOAD_LLVM
         if (DBGBIT(61, 1) && flg.omptarget)
           dumpomptarget(ompaccel_tinfo_current_get());
-        /* todo ompaccel we do not have host fallback mechanism.
-         * all codes are in the target region will be generated ONLY for device.
-         */
         if (flg.omptarget && ompaccel_tinfo_has(gbl.currsub))
           gbl.isnvvmcodegen = true;
 #endif