Merge pull request #127 from ThePortlandGroup/nv_stage

Pull 2017-07-06T07-00 Recent NVIDIA Changes

Author: sscalpone

runtime/include/mthdecls.h

@@ -74,7 +74,7 @@ typedef struct {
    single precision versions of the math.h functions, in which case the
    single precision versions should be used:  */
 
-#if defined(WIN64) || defined(SOLARIS8664)
+#if defined(WIN64)
 
 #define ACOSF acos
 #define ASINF asin
@@ -101,8 +101,8 @@ typedef struct {
 #define FLOORF floor
 #define FMODF fmod
 #define HYPOTF hypot
-#define COPYSIGNF copysignf
-#define COPYSIGN copysign
+#define COPYSIGNF _copysignf
+#define COPYSIGN _copysign
 #define ERFF erf
 #define ERFCF erfc
 #define GAMMAF tgamma
@@ -132,18 +132,12 @@ typedef struct {
  * Also, with MSOT8, hypot is depcrecated, and will not link
  * correctly. Need to use _hypot.
  */
-#if defined(WIN64)
 #define POWF __mth_i_dpowd
-#define COPYSIGNF _copysignf
-#define COPYSIGN _copysign
 #ifdef __PGI_TOOLS9
 #define hypot _hypot
 #endif
-#else
-#define POWF pow
-#endif
 
-#else
+#else		/* #if defined (WIN64) */
 #define ACOSF acosf
 #define ASINF asinf
 #define ATANF atanf
@@ -249,17 +243,7 @@ typedef struct {
 #define BESSEL_Y1 y1
 #define BESSEL_YN yn
 #endif
-#endif
-
-/*
-#if defined(INTERIX86)
-extern double HYPOTF();
-#endif
-*/
-
-#if defined(SOLARIS8664)
-extern double hypot(double, double);
-#endif
+#endif		/* #if defined (WIN64) */
 
 /*  declarations for math functions */
 

tools/flang2/flang2exe/cgmain.c

@@ -5491,27 +5491,6 @@ find_load_cse(int ilix, OPERAND *load_op, LL_Type *llt)
   return NULL;
 }
 
-static LOGICAL
-openmp_atomic_ld(int ilix)
-{
-  ATOMIC_INFO info;
-  switch (ILI_OPC(ilix)) {
-  case IL_ATOMICLDI:
-  case IL_ATOMICLDKR:
-  case IL_ATOMICLDA:
-  case IL_ATOMICLDSP:
-  case IL_ATOMICLDDP:
-    break;
-  default:
-    return FALSE;
-  }
-  info = atomic_info(ilix);
-
-  if (info.origin == AORG_OPENMP)
-    return TRUE;
-  else
-    return FALSE;
-}
 
 /**
    \brief return new operand of type OT_TMP as result of loading \p load_op
@@ -5536,7 +5515,7 @@ make_load(int ilix, OPERAND *load_op, LL_Type *rslt_type, MSZ msz,
          0, ERR_Fatal);
 
   cse_op = NULL;
-  if (ENABLE_CSE_OPT && !openmp_atomic_ld(ilix)) {
+  if (ENABLE_CSE_OPT && !is_omp_atomic_ld(ilix)) {
     operand = find_load_cse(ilix, load_op, rslt_type);
     if (operand != NULL) {
       const int bits = ll_type_int_bits(operand->ll_type);

tools/flang2/flang2exe/expatomics.c

@@ -2563,13 +2563,15 @@ _exp_mp_atomic_read(int stc, const datainfo* o, DTYPE dtype,
       o->dtype == DT_DCMPLX) {
 
     tmp_sptr = mkatomictemp(o->dtype);
-    nme[0] = addnme(NT_VAR, tmp_sptr, 0, (INT)0);
     size = zsize_of(dtype);
     size_ili = ad_icon(size);
+    ADDRTKNP(tmp_sptr, 1);
+    loc_of(nme[0]);
     result = ll_make_atomic_load(size_ili, opnd[0], 
                                  ad_acon(tmp_sptr, 0), opnd[2]);
     iltb.callfg = 1;
     chk_block(result);
+    nme[0] = addnme(NT_VAR, tmp_sptr, 0, (INT)0);
     result = ad3ili(o->ld, 
                     ad_acon(tmp_sptr, 0), 
                     nme[0], 
@@ -2650,6 +2652,8 @@ _exp_mp_atomic_write(int stc, const datainfo* o, DTYPE dtype,
                    nme[1], mem_size(DTY(dtype)));
     chk_block(result);
     size_ili = ad_icon(size);
+    ADDRTKNP(tmp_sptr, 1);
+    loc_of(nme[0]);
     result = ll_make_atomic_store(size_ili, opnd[0], 
                                   ad_acon(tmp_sptr, 0), opnd[2]);
     iltb.callfg = 1;
@@ -2802,6 +2806,7 @@ exp_mp_atomic_update(ILM *ilmp)
     expected_sptr = mkatomictemp(o->dtype);
     stc = atomic_encode_rmw(mem_size(DTY(dtype)), 
                             SS_PROCESS, AORG_OPENMP, aop);
+    loc_of(nme[0]);
     result = ad5ili(o->atomicrmw, opnd[1], opnd[0], nme[0], stc, opnd[2]);
 
     result = ad4ili(o->st, result, ad_acon(expected_sptr, 0), 
@@ -2844,6 +2849,7 @@ exp_mp_atomic_update(ILM *ilmp)
       result = ad4ili(o->st, desired_val, ad_acon(desired_sptr, 0), 
                       addnme(NT_VAR, desired_sptr, 0, (INT)0),
                       mem_size(DTY(dtype)));
+      ASSNP(desired_sptr, 1);
       chk_block(result);
     }
     /* do compare exchange */
@@ -2852,6 +2858,7 @@ exp_mp_atomic_update(ILM *ilmp)
       size_ili = ad_icon(size);
       ADDRTKNP(expected_sptr, 1);
       ADDRTKNP(desired_sptr, 1);
+      loc_of(nme[0]);
       result = ll_make_atomic_compare_xchg(size_ili, opnd[0], 
                                          ad_acon(expected_sptr, 0), 
                                          ad_acon(desired_sptr, 0), 
@@ -2889,6 +2896,7 @@ exp_mp_atomic_update(ILM *ilmp)
                     addnme(NT_VAR, desired_sptr, 0, (INT)0),
                     msz);
       stc = atomic_encode(msz, SS_PROCESS, AORG_OPENMP);
+      loc_of(nme[0]);
       cmpxchg = ad_cmpxchg(o->cmpxchg, desired_val, opnd[0], nme[0],
                            stc, expected_val, ad_icon(0), opnd[2],
                            ad_icon(0));
@@ -3054,6 +3062,7 @@ exp_mp_atomic_capture(ILM *ilmp)
     expected_sptr = mkatomictemp(o->dtype);
     stc = atomic_encode_rmw(mem_size(DTY(cpt.dtype)), 
                             SS_PROCESS, AORG_OPENMP, cpt.aop);
+    loc_of(cpt.x_nme);
     result = ad5ili(o->atomicrmw, cpt.expr, cpt.x, cpt.x_nme, stc, 
                     cpt.mem_order);
 
@@ -3111,6 +3120,7 @@ exp_mp_atomic_capture(ILM *ilmp)
       size_ili = ad_icon(size);
       ADDRTKNP(expected_sptr, 1);
       ADDRTKNP(desired_sptr, 1);
+      loc_of(cpt.x_nme);
       result = ll_make_atomic_compare_xchg(size_ili, cpt.x, 
                                          ad_acon(expected_sptr, 0), 
                                          ad_acon(desired_sptr, 0), 
@@ -3166,6 +3176,7 @@ exp_mp_atomic_capture(ILM *ilmp)
                     addnme(NT_VAR, desired_sptr, 0, (INT)0),
                     msz);
       stc = atomic_encode(msz, SS_PROCESS, AORG_OPENMP);
+      loc_of(cpt.x_nme);
       cmpxchg = ad_cmpxchg(o->cmpxchg, desired_val, cpt.x, cpt.x_nme,
                            stc, expected_val, ad_icon(0), cpt.mem_order,
                            ad_icon(0));

tools/flang2/flang2exe/expsmp.c

@@ -2211,7 +2211,18 @@ exp_smp(ILM_OP opc, ILM *ilmp, int curilm)
     break;
 
   case IM_MP_ATOMIC:
+    if (ll_ilm_is_rewriting())
+      break;
+    wr_block();
+    cr_block();
+    bihb.csfg = BIH_CS(expb.curbih) = TRUE;
+    break;
   case IM_MP_ENDATOMIC:
+    if (ll_ilm_is_rewriting())
+      break;
+    wr_block();
+    cr_block();
+    bihb.csfg = BIH_CS(expb.curbih) = FALSE;
     break;
 
   case IM_MP_ATOMICREAD:

tools/flang2/flang2exe/ili.h

@@ -588,6 +588,8 @@ int atomic_encode_rmw(MSZ msz, SYNC_SCOPE scope, ATOMIC_ORIGIN origin,
                       ATOMIC_RMW_OP op);
 MEMORY_ORDER memory_order(int ilix);
 ATOMIC_INFO atomic_info(int ilix);
+extern LOGICAL is_omp_atomic_ld(int);
+extern LOGICAL is_omp_atomic_st(int);
 
 ATOMIC_INFO atomic_decode(int encoding);
 int atomic_info_index(ILI_OP opc);

tools/flang2/flang2exe/iliutil.c

@@ -13029,3 +13029,48 @@ memory_order(int ilix)
   }
   return memory_order_from_operand(ILI_OPND(ilix, i));
 }
+
+
+LOGICAL
+is_omp_atomic_ld(int ilix)
+{
+  ATOMIC_INFO info;
+  switch (ILI_OPC(ilix)) {
+  case IL_ATOMICLDI:
+  case IL_ATOMICLDKR:
+  case IL_ATOMICLDA:
+  case IL_ATOMICLDSP:
+  case IL_ATOMICLDDP:
+    break;
+  default:
+    return FALSE;
+  }
+  info = atomic_info(ilix);
+
+  if (info.origin == AORG_OPENMP)
+    return TRUE;
+  else
+    return FALSE;
+}
+
+LOGICAL
+is_omp_atomic_st(int ilix)
+{
+  ATOMIC_INFO info;
+  switch (ILI_OPC(ilix)) {
+  case IL_ATOMICSTI:
+  case IL_ATOMICSTKR:
+  case IL_ATOMICSTA:
+  case IL_ATOMICSTSP:
+  case IL_ATOMICSTDP:
+    break;
+  default:
+    return FALSE;
+  }
+  info = atomic_info(ilix);
+
+  if (info.origin == AORG_OPENMP)
+    return TRUE;
+  else
+    return FALSE;
+}