Expand structure copy in a unified way.

Currently, the frontend generates the SMOVE ILM for an assignment of a
structure & derived type.  The ILM is then expanded into loads & stores, or
other smove ILI (SMOVE, SMOVEI/SMOVES, XMOVE).  What we want to do is expand to
a new ILI (GSMOVE) that gives us the addresses, nmes, and dtype of the source &
destination.  It is then straightforward for other phases, to make
target-dependent decisions based on size/alignment of the aggregate.

xflag.n: add XBIT(2,0x800000) - expand IM_SMOVE into IL_GSMOVE in exp_rte.c

pgifeat.h:
- define the macro, USE_GSMOVE, as XBIT(2,0x800000)
- minor source cleanup  (comments)

ilitp.n:
Add IL_GSMOVE arlnk arlnk nme nme stc
op1  - the source address.
op2  - the destination address.
nme1 - the names table entry for the source
nme2 - the namse table entry for the destination
stc  - is the dtype of the struct (derived type).

expand.h: prototype for exp_remove_gsmove(void)

expand.c: add prototypes.

exp_rte.c
- expand_smove() - gen IL_GSMOVE if USE_GSMOVE is set
- add exp_remove_gsmove() - called from rmsmove.c; search the ILI blocks
  of the function and if a GSMOVE is found, call _exp_smove() to transform
  the ILI into loads/stores, SMOVEI, SMOVE, ...
- add gsmove_chk_block()

rmsmove.c - call gsmove_chk_block() if USE_GSMOVE is set

Author: tskeith

tools/flang1/flang1exe/pgifeat.h

@@ -47,9 +47,9 @@
 #define PG_LM_PLATFORM
 #define TARGET_64BIT 1
 
-                 /*
-                  * 64-bit CPU targets
-                  */
+/*
+ * 64-bit CPU targets
+ */
 #define TARGET_PTRSIZE 8
 
 #if !defined(TARGET_WIN)
@@ -59,6 +59,8 @@
 #define TARGET_LONGSIZE 4
 #endif
 
+/* Uniform structure assigments/moves ILI */
+#define USE_GSMOVE XBIT(2,0x800000)
 
 /* ETLS and threadprivate related features */
 /* By default, prevent ETLS/TLS threadprivate usage */

tools/flang2/docs/xflag.n

@@ -228,6 +228,8 @@ Fill in BIH_LINENO with line numbers on inlined blocks.
 don't use SMOVEI/SMOVES instead of SMOVE in exp_rte.c
 .XB 0x400000
 F90 pointer optimizations.
+.XB 0x800000
+Use GSMOVE in exp_rte.c
 
 .XF "3:" 
 Used to turn on/off various dual-op/dual-inst/pipelined ops.

tools/flang2/flang2exe/exp_rte.c

@@ -59,6 +59,7 @@ static void cp_byval_mem_arg(int);
 static int allochartmp(int lenili);
 static int block_str_move(STRDESC *, STRDESC *);
 static int getchartmp(int ili);
+static void _exp_smove(int, int, int, int, DTYPE);
 
 void arg_is_refd(int); /* FIXME? move from llassem.c? */
 
@@ -132,6 +133,21 @@ getstrlen64(STRDESC *str)
   return il;
 }
 
+/*
+ * Generating GSMOVE ILI is under XBIT(2,0x800000). When the XBIT is not
+ * set, _exp_smove() will proceed as before; in particular, chk_block() is
+ * called to add terminal ILI to the block current to the expander.   When
+ * the XBIT is set, the GSMOVE ili are transformed sometime after the expander,
+ * but we still want the code in _exp_smove() to do the work. However, we
+ * cannot call chk_block() to add the terminal ILI; we must use 'addilt'.
+ * So, define and use a function pointer, p_chk_block, which calls either
+ * chk_block() or a new local addilit routine, gsmove_chk_block().  In this
+ * case, the current ilt is saved as the file static, gsmove_ilt.
+ */
+static void (*p_chk_block)(int) = chk_block;
+static void gsmove_chk_block(int);
+static int  gsmove_ilt;
+
 /* aux.curr_entry->flags description:
  *     Initialized to 0 by exp_end
  * NA      0x1  -  need to save argument registers  (set by exp_end).
@@ -4202,6 +4218,61 @@ expand_smove(int destilm, int srcilm, DTYPE dtype)
   }
   dest_addr = ILI_OF(destilm);
   src_addr = ILI_OF(srcilm);
+  if (USE_GSMOVE) {
+     int ilix;
+     ilix = ad5ili(IL_GSMOVE, src_addr, dest_addr, src_nme, dest_nme, dtype);
+     chk_block(ilix);
+  }
+  else {
+    _exp_smove(dest_nme, src_nme, dest_addr, src_addr, dtype);
+  }
+}
+
+/** \brief Transform the GSMOVE ILI created by expand_smove()
+  */
+void
+exp_remove_gsmove(void)
+{
+  int bihx, iltx, ilix;
+  p_chk_block = gsmove_chk_block;
+  for (bihx = gbl.entbih; bihx; bihx = BIH_NEXT(bihx)) {
+    int next_ilt;
+    LOGICAL any_gsmove = FALSE;
+    rdilts(bihx);
+    for (iltx = ILT_NEXT(0); iltx; ) {
+      next_ilt = ILT_NEXT(iltx);
+      ilix = ILT_ILIP(iltx);
+      if (ILI_OPC(ilix) == IL_GSMOVE) {
+        int src_addr  = ILI_OPND(ilix, 1);
+        int dest_addr = ILI_OPND(ilix, 2);
+        int src_nme   = ILI_OPND(ilix, 3);
+        int dest_nme  = ILI_OPND(ilix, 4);
+        DTYPE dtype   = ILI_OPND(ilix, 5);
+        any_gsmove = TRUE;
+        gsmove_ilt = iltx;
+        _exp_smove(dest_nme, src_nme, dest_addr, src_addr, dtype);
+        ILT_NEXT(gsmove_ilt) = next_ilt;
+        ILT_PREV(next_ilt) = gsmove_ilt;
+	delilt(iltx);
+      }
+      iltx = next_ilt;
+    }
+    wrilts(bihx);
+    if (DBGBIT(10,2) && any_gsmove) {
+      fprintf(gbl.dbgfil, "\n***** After remove gsmove *****\n");
+      dump_one_block(gbl.dbgfil, bihx, NULL);
+    }
+  }
+  p_chk_block = chk_block;
+}
+
+static void
+_exp_smove(int dest_nme, int src_nme, int dest_addr, int src_addr, DTYPE dtype)
+{
+  ISZ_T n;        /* number of bytes left to copy		*/
+  int i;
+  INT offset; /* number of bytes from begin addr 	*/
+
   n = size_of(dtype);
   offset = 0;
 
@@ -4211,10 +4282,10 @@ expand_smove(int destilm, int srcilm, DTYPE dtype)
   if (n > TEST_BOUND) {
 
     if (XBIT(2, 0x200000)) {
-      chk_block(ad4ili(IL_SMOVE, src_addr, dest_addr, ad_aconi(n / SMOVE_CHUNK),
+      p_chk_block(ad4ili(IL_SMOVE, src_addr, dest_addr, ad_aconi(n / SMOVE_CHUNK),
                        dest_nme));
     } else {
-      chk_block(ad4ili(IL_SMOVEI, ad2ili(IL_SMOVES, src_addr, src_nme),
+      p_chk_block(ad4ili(IL_SMOVEI, ad2ili(IL_SMOVES, src_addr, src_nme),
                        dest_addr, n, dest_nme));
     }
     smove_flag = 1; /* structure move in this function */
@@ -4255,7 +4326,7 @@ expand_smove(int destilm, int srcilm, DTYPE dtype)
         ilip = ad4ili(IL_STKR, ilix, ilip, dest_nme, siz);
       else
         ilip = ad4ili(IL_ST, ilix, ilip, dest_nme, siz);
-      chk_block(ilip);
+      p_chk_block(ilip);
 
       offset += skip;
       n -= skip;
@@ -5317,6 +5388,15 @@ exp_reset_argregs(int ir, int fr)
 {
 }
 
+/**
+ * \brief Create & add an ILT for an ILI when transforming GSMOVE ILI
+ */
+static void
+gsmove_chk_block(int ili)
+{
+  gsmove_ilt = addilt(gsmove_ilt, ili);
+}
+
 /*------------------------------------------------------------------*/
 
 #undef ILM_OPC

tools/flang2/flang2exe/expand.c

@@ -68,7 +68,7 @@ static int efunc(char *);
  * Initialize global data structures
  */
 void
-ds_init()
+ds_init(void)
 {
   int i;
   ili_init();
@@ -94,7 +94,7 @@ ds_init()
 } /* ds_init */
 
 void
-exp_init()
+exp_init(void)
 {
   /*
    * Allocate the space necessary to hold the auxiliary information for ILM
@@ -155,7 +155,7 @@ exp_init()
  * clean up allocated space when the program isn't compiled
  */
 void
-exp_cleanup()
+exp_cleanup(void)
 {
   if (rgsetb.stg_base)
     EXP_FREE(rgsetb);

tools/flang2/flang2exe/expand.h

@@ -267,6 +267,7 @@ void exp_qjsr(char *, int, ILM *, int);
 void exp_estmt(int);
 void exp_label(int);
 void expand_smove(int, int, DTYPE);
+void exp_remove_gsmove(void);
 void exp_load(ILM_OP, ILM *, int);
 void exp_store(ILM_OP, ILM *, int);
 void exp_misc(ILM_OP, ILM *, int);

tools/flang2/flang2exe/pgifeat.h

@@ -47,9 +47,9 @@
 #define PG_LM_PLATFORM
 #define TARGET_64BIT 1
 
-                 /*
-                  * 64-bit CPU targets
-                  */
+/*
+ * 64-bit CPU targets
+ */
 #define TARGET_PTRSIZE 8
 
 #if !defined(TARGET_WIN)
@@ -59,6 +59,8 @@
 #define TARGET_LONGSIZE 4
 #endif
 
+/* Uniform structure assigments/moves ILI */
+#define USE_GSMOVE XBIT(2,0x800000)
 
 /* ETLS and threadprivate related features */
 /* By default, prevent ETLS/TLS threadprivate usage */

tools/flang2/flang2exe/rmsmove.c

@@ -94,10 +94,18 @@ fixup_nme(int nmex, int msize, int offset, int iter)
   return new_nme;
 } /* fixup_nme */
 
+void exp_remove_gsmove(void);
 void
 rm_smove(void)
 {
   int bihx, iltx, ilix, new_acon;
+  /*
+   * First implementation of GSMOVE will be under XBIT(2,0x800000). When this
+   * is the only method, presumably, the code in exp_remove_gsmove() will be
+   * moved to the ensuing loop.
+   */
+  if (USE_GSMOVE)
+    exp_remove_gsmove();
   for (bihx = gbl.entbih; bihx; bihx = BIH_NEXT(bihx)) {
     LOGICAL have_smove = FALSE;
     rdilts(bihx);

tools/flang2/utils/ilitp/aarch64/ilitp.n

@@ -2096,6 +2096,22 @@ Emitted immediately before an SMOVE operation.
 .AT other null trm dom
 .CG CGonly "rep"
 
+.IL GSMOVE arlnk arlnk nme nme stc
+General structure copy.
+A structure store (SMOVE ILM) is expanded into the GSMOVE ILI; a phase, such as
+the ACC CG, needs to have the structure assignment presented as a first-class
+ILI operation. 
+After the last phase requiring GSMOVE, GSMOVE will then be expanded into
+lower level ILI which will be dependent on target, alignment, small vs large,
+etc.
+  \'op1' is the source address.
+  \'op2' is the destination address.
+  \'nme1' is the names table entry for source struct.
+  \'nme2' is the namse table entry for the destination struct.
+  \'stc' is the dtype of the struct.
+.AT other null trm dom
+.CG notCG
+
 .IL SMOVE arlnk arlnk arlnk nme
 Structure copy.
   \'op1' is the source address.

tools/flang2/utils/ilitp/ppc64le/ilitp.n

@@ -2092,6 +2092,22 @@ Emitted immediately before an SMOVE operation.
 .AT other null trm dom
 .CG CGonly "rep"
 
+.IL GSMOVE arlnk arlnk nme nme stc
+General structure copy.
+A structure store (SMOVE ILM) is expanded into the GSMOVE ILI; a phase, such as
+the ACC CG, needs to have the structure assignment presented as a first-class
+ILI operation. 
+After the last phase requiring GSMOVE, GSMOVE will then be expanded into
+lower level ILI which will be dependent on target, alignment, small vs large,
+etc.
+  \'op1' is the source address.
+  \'op2' is the destination address.
+  \'nme1' is the names table entry for source struct.
+  \'nme2' is the namse table entry for the destination struct.
+  \'stc' is the dtype of the struct.
+.AT other null trm dom
+.CG notCG
+
 .IL SMOVE arlnk arlnk arlnk nme
 Structure copy.
   \'op1' is the source address.

tools/flang2/utils/ilitp/x86_64/ilitp.n

@@ -2417,6 +2417,22 @@ Emitted immediately before an SMOVE operation.
 .AT other null trm dom
 .CG CGonly "rep"
 
+.IL GSMOVE arlnk arlnk nme nme stc
+General structure copy.
+A structure store (SMOVE ILM) is expanded into the GSMOVE ILI; a phase, such as
+the ACC CG, needs to have the structure assignment presented as a first-class
+ILI operation. 
+After the last phase requiring GSMOVE, GSMOVE will then be expanded into
+lower level ILI which will be dependent on target, alignment, small vs large,
+etc.
+  \'op1' is the source address.
+  \'op2' is the destination address.
+  \'nme1' is the names table entry for source struct.
+  \'nme2' is the namse table entry for the destination struct.
+  \'stc' is the dtype of the struct.
+.AT other null trm dom
+.CG notCG
+
 .IL SMOVE arlnk arlnk arlnk nme
 Structure copy.
   \'op1' is the source address.