Merge pull request #303 from ThePortlandGroup/nv_stage

Pull 2017-11-10T16-05 Recent NVIDIA Changes
flang-compiler · Nov 11, 2017 · 2b20297 · 2b20297
2 parents 7d8327e + 88e109c
commit 2b20297
Show file tree

Hide file tree

Showing 14 changed files with 79 additions and 21 deletions.
diff --git a/runtime/CMakeLists.txt b/runtime/CMakeLists.txt
@@ -21,6 +21,7 @@ add_definitions(
  -DMAXCPUSL=8 
  -DMAXCPUSR=8 
  -DTARGET_LINUX 
+ -DTARGET_LLVM 
  -DLINUX 
  -DPGF90 
  -DPGFLANG

diff --git a/runtime/flang/initpar.c b/runtime/flang/initpar.c
@@ -852,9 +852,7 @@ f90_compiled_arg()
 
 
 void
-#if defined(TARGET_LLVM)
 __attribute__((constructor))
-#endif
 f90_compiled()
 {
 #ifndef TARGET_LINUX_ARM
@@ -872,12 +870,3 @@ f90_compiled()
   }
 }
 
-#ifndef PGLANG
-/* Wrapper function to maintain compatibility with previous PGI products */
-void
-pgf90_compiled()
-{
-  f90_compiled();
-#endif
-}
-
diff --git a/tools/flang1/flang1exe/dump.c b/tools/flang1/flang1exe/dump.c
@@ -2748,7 +2748,7 @@ dsym(int sptr)
     PARAMVALP(0, 0);
 #ifdef PDALNG
     putnzint("pdaln", PDALNG(0));
-    PDALNP(0, 0);
+    b4P(0, 0);
 #endif
     putnzint("socptr", SOCPTRG(0));
     SOCPTRP(0, 0);

diff --git a/tools/flang1/flang1exe/exterf.c b/tools/flang1/flang1exe/exterf.c
@@ -41,7 +41,6 @@
 
 #include "dpm_out.h"
 
-#define MOD_CMN_IDX(a, c) (((a) << 1) | (c))
 #define COMPILER_OWNED_MODULE XBIT(58,0x100000)
 
 /* ------------------------------------------------------------------ */

diff --git a/tools/flang1/flang1exe/lowersym.c b/tools/flang1/flang1exe/lowersym.c
@@ -536,7 +536,7 @@ fill_fixed_array_dtype(int dtype)
   if (mlpyr > 0) {
     ADD_NUMELM(dtype) = mk_cnst(lower_getiszcon(mlpyr));
   } else {
-    ADD_NUMELM(dtype) = 0;
+    ADD_NUMELM(dtype) = astb.bnd.zero;
   }
 } /* fill_fixed_array_dtype */
 

diff --git a/tools/flang1/flang1exe/semfunc.c b/tools/flang1/flang1exe/semfunc.c
@@ -6173,6 +6173,7 @@ ref_pd(SST *stktop, ITEM *list)
     shape1 = A_SHAPEG(ARG_AST(0));
     count = SHD_NDIM(shape1); /* rank of array arg */
     argt_count = count * 2 + 2;
+    adjarr = 0;
     asumsz = 0;
     assumshp = 0;
     arg1 = ARG_AST(0);

diff --git a/tools/flang1/utils/prstab/prstab.h b/tools/flang1/utils/prstab/prstab.h
@@ -52,7 +52,7 @@
 #define MAXPROD 4000
 #define MAXLST 24000
 #define MAXBAS 44000
-#define MAXTRN 28000
+#define MAXTRN 40000
 #define MAXRED 10000
 #define MAXSCR 6000
 #define MAXHASH 2048

diff --git a/tools/flang2/flang2exe/exputil.c b/tools/flang2/flang2exe/exputil.c
@@ -384,14 +384,14 @@ chk_block(int newili)
   }
 }
 
-/** \brief Like chk_block, but suppress CAN_THROW flag. 
- *   
+/** \brief Like chk_block, but suppress CAN_THROW flag.
+ *
  * When a call can throw and defines two result registers, we have an ad-hoc
  * rule that only the second store is marked as "can throw".  This utility
  * is useful for ensuring that the first store is not marked "can throw".
  */
 void
-chk_block_suppress_throw(int newili) 
+chk_block_suppress_throw(int newili)
 {
   chk_block(newili);
   ILT_SET_CAN_THROW(expb.curilt, 0);
@@ -1181,10 +1181,10 @@ void
 exp_add_copy(int lhssptr, int rhssptr)
 {
   int rhsacon, lhsacon, rhsnme, lhsnme, rhsld, lhsst, sz;
-  ILI_OP rhsopc, lhsopc; 
+  ILI_OP rhsopc, lhsopc;
   MSZ msz;
   if (lhssptr == rhssptr)
-    return;  
+    return;
   rhsacon = ad_acon(rhssptr, 0);
   sz = size_of(DTYPEG(rhssptr));
   if (sz == 8) {

diff --git a/tools/flang2/flang2exe/ilmutil.c b/tools/flang2/flang2exe/ilmutil.c
@@ -1140,6 +1140,12 @@ _dumpilms(ILM_T *ilm_base, int check)
         case PR_ACCENDSCALARREG:
           s = "ACCENDSCALARREG";
           break;
+        case PR_ACCSERIAL:
+          s = "ACCSERIAL";
+          break;
+        case PR_ACCENDSERIAL:
+          s = "ACCENDSERIAL";
+          break;
         case PR_ACCPARCONSTRUCT:
           s = "ACCPARCONSTRUCT";
           break;
@@ -1191,6 +1197,9 @@ _dumpilms(ILM_T *ilm_base, int check)
         case PR_ACCWAITDIR:
           s = "ACCWAITDIR";
           break;
+        case PR_ACCSLOOP:
+          s = "ACCSLOOP";
+          break;
         case PR_ACCKLOOP:
           s = "ACCKLOOP";
           break;

diff --git a/tools/flang2/flang2exe/outliner.c b/tools/flang2/flang2exe/outliner.c
@@ -1015,6 +1015,15 @@ clone_uplevel(int uplevel_sptr, int uplevel_stblk_sptr)
     ilix = ad2ili(IL_LDA, ili, addnme(NT_VAR, arg, 0, (INT)0));
   }
 
+    /* set alignment of last argument for GPU "align 8". */
+    if (DTY(uplevel_dtype) == TY_STRUCT)
+      DTY(uplevel_dtype + 4) = 7;
+
+    if (DTY(DTYPEG(new_uplevel)) == TY_STRUCT)
+      DTY(DTYPEG(new_uplevel) + 4) = 7;
+
+
+
   /* For C we have a homed argument, a pointer to a pointer to an uplevel.
    * This will dereference the pointer, we do not need to do this for Fortran.
    */

diff --git a/tools/flang2/utils/ilitp/aarch64/ilitp.n b/tools/flang2/utils/ilitp/aarch64/ilitp.n
@@ -4277,11 +4277,26 @@ End a block of code to run as a scalar kernel on the accelerator
 .AT other null trm
 .CG notCG
 
+.IL ACCSERIAL lnk
+Start a block of code to run as a serial kernel on the accelerator
+.AT other null trm
+.CG notCG
+
+.IL ACCENDSERIAL
+End a block of code to run as a serial kernel on the accelerator
+.AT other null trm
+.CG notCG
+
 .IL ACCELLP lnk
 The following loop is to be targeted for the accelerator
 .AT other null trm
 .CG notCG
 
+.IL ACCSLOOP lnk
+The following loop in a serial region is to be targeted for the accelerator
+.AT other null trm
+.CG notCG
+
 .IL ACCKLOOP lnk
 The following loop in a kernels region is to be targeted for the accelerator
 .AT other null trm

diff --git a/tools/flang2/utils/ilitp/ppc64le/ilitp.n b/tools/flang2/utils/ilitp/ppc64le/ilitp.n
@@ -4274,11 +4274,26 @@ End a block of code to run as a scalar kernel on the accelerator
 .AT other null trm
 .CG notCG
 
+.IL ACCSERIAL lnk
+Start a block of code to run as a serial kernel on the accelerator
+.AT other null trm
+.CG notCG
+
+.IL ACCENDSERIAL
+End a block of code to run as a serial kernel on the accelerator
+.AT other null trm
+.CG notCG
+
 .IL ACCELLP lnk
 The following loop is to be targeted for the accelerator
 .AT other null trm
 .CG notCG
 
+.IL ACCSLOOP lnk
+The following loop in a serial region is to be targeted for the accelerator
+.AT other null trm
+.CG notCG
+
 .IL ACCKLOOP lnk
 The following loop in a kernels region is to be targeted for the accelerator
 .AT other null trm

diff --git a/tools/flang2/utils/ilitp/x86_64/ilitp.n b/tools/flang2/utils/ilitp/x86_64/ilitp.n
@@ -5129,12 +5129,30 @@ Link to a list of clauses.
 .AT other null trm
 .CG notCG
 
+.IL ACCSERIAL lnk
+Start a block of code to run as a serial kernel on the accelerator.
+Link to a list of clauses.
+.AT other null trm
+.CG notCG
+
+.IL ACCENDSERIAL
+End a block of code to run as a serial kernel on the accelerator.
+Link to a list of clauses.
+.AT other null trm
+.CG notCG
+
 .IL ACCELLP lnk
 The following loop is to be targeted for the accelerator.
 Link to a list of clauses.
 .AT other null trm
 .CG notCG
 
+.IL ACCSLOOP lnk
+The following loop in a serial region is to be targeted for the accelerator.
+Link to a list of clauses.
+.AT other null trm
+.CG notCG
+
 .IL ACCKLOOP lnk
 The following loop in a kernels region is to be targeted for the accelerator.
 Link to a list of clauses.

diff --git a/tools/shared/pragma.h b/tools/shared/pragma.h
@@ -166,8 +166,10 @@ typedef enum {
   PR_ACCSETDIR = 116,          /* accelerator set directive */
   PR_ACCUSEDEVICEIFP = 117,    /* accelerator use device clause combined with if present */
   PR_ACCNO_CREATE = 118,       /* no_create clause */
+  PR_ACCSERIAL = 119,          /* accelerator serial construct */
+  PR_ACCENDSERIAL = 120,       /* end accelerator serial construct */
+  PR_ACCSLOOP = 121,           /* loop in accelerator serial region */
 } PR_PRAGMA;
 
 /* Ignore data movement pragmas */
 #define ACC_DATAMOVEMENT_DISABLED XBIT(195, 0x400)
-