toplev.c (rest_of_compilation): Set bct_p on second call to loop_optimize.

        * toplev.c (rest_of_compilation): Set bct_p on second call to
        loop_optimize.
        * loop.c (loop_optimize, scan_loop, strength_reduce): New argument
        bct_p.
        (strength_reduce): Only call analyze_loop_iterations and
        insert_bct if bct_p set.
        (check_dbra_loop): Fix typo.
        (insert_bct): Use word_mode instead of SImode.
        (instrument_loop_bct): Likewise.  Do not delete iteration count
        condition code generation insn.  Initialize iteration count before
        loop start.
        * rtl.h (loop_optimize): Update prototype.
        * ginclude/va-ppc.h (va_arg): longlong types in overflow area are
        not doubleword aligned.
        * rs6000.c (optimization_options): New function.
        (secondary_reload_class): Only call true_regnum for PSEUDO_REGs.
        * rs6000.h (OPTIMIZATION_OPTIONS): Define.
        (REG_ALLOC_ORDER): Allocate highest numbered condition regsiters
        first; cr1 can be used for FP record condition insns.

From-SVN: r22471

gcc/ChangeLog

+Fri Sep 18 23:50:56 1998  David Edelsohn  <edelsohn@mhpcc.edu>
+
+	* toplev.c (rest_of_compilation): Set bct_p on second call to
+	loop_optimize.
+	* loop.c (loop_optimize, scan_loop, strength_reduce): New argument
+	bct_p.
+	(strength_reduce): Only call analyze_loop_iterations and
+	insert_bct if bct_p set.
+	(check_dbra_loop): Fix typo.
+	(insert_bct): Use word_mode instead of SImode.
+	(instrument_loop_bct): Likewise.  Do not delete iteration count
+	condition code generation insn.  Initialize iteration count before
+	loop start.
+	* rtl.h (loop_optimize): Update prototype.
+
+	* ginclude/va-ppc.h (va_arg): longlong types in overflow area are
+	not doubleword aligned.
+
+	* rs6000.c (optimization_options): New function.
+	(secondary_reload_class): Only call true_regnum for PSEUDO_REGs.
+	* rs6000.h (OPTIMIZATION_OPTIONS): Define.
+	(REG_ALLOC_ORDER): Allocate highest numbered condition regsiters
+	first; cr1 can be used for FP record condition insns.
+
 Fri Sep 18 09:44:55 1998  Nick Clifton  <nickc@cygnus.com>
 
 	* config/m32r/m32r.h (m32r_block_immediate_operand): Add to

gcc/config/rs6000/rs6000.c

@@ -347,6 +347,20 @@ rs6000_override_options (default_cpu)
   SUBTARGET_OVERRIDE_OPTIONS;
 #endif
 }
+
+void
+optimization_options (level, size)
+     int level;
+     int size ATTRIBUTE_UNUSED;
+{
+#if 0
+#ifdef HAIFA
+  /* When optimizing, enable use of BCT instruction.  */
+  if (level >= 1)
+      flag_branch_on_count_reg = 1;
+#endif
+#endif
+}
 
 /* Do anything needed at the start of the asm file.  */
 
@@ -1305,14 +1319,18 @@ function_arg_padding (mode, type)
    
    Windows NT wants anything >= 8 bytes to be double word aligned.
 
-   V.4 wants long longs to be double word aligned.  */
+   V.4 wants long longs to be double word aligned.
+
+   FP emulation: double precision passed, returned, and same alignment
+   as long long.  */
 
 int
 function_arg_boundary (mode, type)
      enum machine_mode mode;
      tree type;
 {
-  if ((DEFAULT_ABI == ABI_V4 || DEFAULT_ABI == ABI_SOLARIS) && mode == DImode)
+  if ((DEFAULT_ABI == ABI_V4 || DEFAULT_ABI == ABI_SOLARIS)
+      && ((mode == DImode) || (TARGET_SOFT_FLOAT && mode == DFmode)))
     return 64;
 
   if (DEFAULT_ABI != ABI_NT || TARGET_64BIT)
@@ -2188,10 +2206,7 @@ secondary_reload_class (class, mode, in)
      enum machine_mode mode ATTRIBUTE_UNUSED;
      rtx in;
 {
-  int regno = true_regnum (in);
-
-  if (regno >= FIRST_PSEUDO_REGISTER)
-    regno = -1;
+  int regno;
 
   /* We can not copy a symbolic operand directly into anything other than
      BASE_REGS for TARGET_ELF.  So indicate that a register from BASE_REGS
@@ -2203,6 +2218,25 @@ secondary_reload_class (class, mode, in)
 	  || GET_CODE (in) == CONST))
     return BASE_REGS;
 
+  if (GET_CODE (in) == REG)
+    {
+      regno = REGNO (in);
+      if (regno >= FIRST_PSEUDO_REGISTER)
+	{
+	  regno = true_regnum (in);
+	  if (regno >= FIRST_PSEUDO_REGISTER)
+	    regno = -1;
+	}
+    }
+  else if (GET_CODE (in) == SUBREG)
+    {
+      regno = true_regnum (in);
+      if (regno >= FIRST_PSEUDO_REGISTER)
+	regno = -1;
+    }
+  else
+    regno = -1;
+
   /* We can place anything into GENERAL_REGS and can put GENERAL_REGS
      into anything.  */
   if (class == GENERAL_REGS || class == BASE_REGS

gcc/config/rs6000/rs6000.h

@@ -473,10 +473,17 @@ extern int rs6000_debug_arg;		/* debug argument handling */
    defined, is executed once just after all the command options have
    been parsed.
 
+   Don't use this macro to turn on various extra optimizations for
+   `-O'.  That is what `OPTIMIZATION_OPTIONS' is for.
+
    On the RS/6000 this is used to define the target cpu type.  */
 
 #define OVERRIDE_OPTIONS rs6000_override_options (TARGET_CPU_DEFAULT)
 
+/* Define this to change the optimizations performed by default.  */
+#define OPTIMIZATION_OPTIONS(LEVEL,SIZE) optimization_options(LEVEL,SIZE)
+
+
 /* Show we can debug even without a frame pointer.  */
 #define CAN_DEBUG_WITHOUT_FP
 
@@ -712,9 +719,10 @@ extern int rs6000_debug_arg;		/* debug argument handling */
 	fp13 - fp2	(not saved; incoming fp arg registers)
 	fp1		(not saved; return value)
  	fp31 - fp14	(saved; order given to save least number)
-	cr1, cr6, cr7	(not saved or special)
+	cr7, cr6	(not saved or special)
+	cr1		(not saved, but used for FP operations)
 	cr0		(not saved, but used for arithmetic operations)
-	cr2, cr3, cr4	(saved)
+	cr4, cr3, cr2	(saved)
         r0		(not saved; cannot be base reg)
 	r9		(not saved; best for TImode)
 	r11, r10, r8-r4	(not saved; highest used first to make less conflict)
@@ -732,7 +740,7 @@ extern int rs6000_debug_arg;		/* debug argument handling */
    33,							\
    63, 62, 61, 60, 59, 58, 57, 56, 55, 54, 53, 52, 51,	\
    50, 49, 48, 47, 46, 					\
-   69, 74, 75, 68, 70, 71, 72,				\
+   75, 74, 69, 68, 72, 71, 70,				\
    0,							\
    9, 11, 10, 8, 7, 6, 5, 4,				\
    3,							\

gcc/ginclude/va-ppc.h

@@ -158,9 +158,6 @@ __extension__ (*({							\
     }									\
   else									\
     {									\
-      if (__va_longlong_p(TYPE) && ((long)__va_overflow(AP) & 4) != 0)	\
-	__va_overflow(AP) += 4;						\
-									\
       __ptr = (TYPE *) (void *) (__va_overflow(AP));			\
       __va_overflow(AP) += __va_size (TYPE) * sizeof (long);		\
     }									\

gcc/loop.c

@@ -310,7 +310,7 @@ static void count_loop_regs_set PROTO((rtx, rtx, varray_type, varray_type,
 				       int *, int)); 
 static void note_addr_stored PROTO((rtx, rtx));
 static int loop_reg_used_before_p PROTO((rtx, rtx, rtx, rtx, rtx));
-static void scan_loop PROTO((rtx, rtx, int));
+static void scan_loop PROTO((rtx, rtx, int, int));
 #if 0
 static void replace_call_address PROTO((rtx, rtx, rtx));
 #endif
@@ -324,7 +324,7 @@ static int rtx_equal_for_loop_p PROTO((rtx, rtx, struct movable *));
 static void add_label_notes PROTO((rtx, rtx));
 static void move_movables PROTO((struct movable *, int, int, rtx, rtx, int));
 static int count_nonfixed_reads PROTO((rtx));
-static void strength_reduce PROTO((rtx, rtx, rtx, int, rtx, rtx, int));
+static void strength_reduce PROTO((rtx, rtx, rtx, int, rtx, rtx, int, int));
 static void find_single_use_in_loop PROTO((rtx, rtx, varray_type));
 static int valid_initial_value_p PROTO((rtx, rtx, int, rtx));
 static void find_mem_givs PROTO((rtx, rtx, int, rtx, rtx));
@@ -440,11 +440,11 @@ init_loop ()
    (or 0 if none should be output).  */
 
 void
-loop_optimize (f, dumpfile, unroll_p)
+loop_optimize (f, dumpfile, unroll_p, bct_p)
      /* f is the first instruction of a chain of insns for one function */
      rtx f;
      FILE *dumpfile;
-     int unroll_p;
+     int unroll_p, bct_p;
 {
   register rtx insn;
   register int i;
@@ -589,7 +589,7 @@ loop_optimize (f, dumpfile, unroll_p)
   for (i = max_loop_num-1; i >= 0; i--)
     if (! loop_invalid[i] && loop_number_loop_ends[i])
       scan_loop (loop_number_loop_starts[i], loop_number_loop_ends[i],
-		 unroll_p);
+		 unroll_p, bct_p);
 
   /* If debugging and unrolling loops, we must replicate the tree nodes
      corresponding to the blocks inside the loop, so that the original one
@@ -643,9 +643,9 @@ next_insn_in_loop (insn, start, end, loop_top)
    write, then we can also mark the memory read as invariant.  */
 
 static void
-scan_loop (loop_start, end, unroll_p)
+scan_loop (loop_start, end, unroll_p, bct_p)
      rtx loop_start, end;
-     int unroll_p;
+     int unroll_p, bct_p;
 {
   register int i;
   rtx p;
@@ -1185,7 +1185,7 @@ scan_loop (loop_start, end, unroll_p)
     {
       the_movables = movables;
       strength_reduce (scan_start, end, loop_top,
-		       insn_count, loop_start, end, unroll_p);
+		       insn_count, loop_start, end, unroll_p, bct_p);
     }
 
   VARRAY_FREE (n_times_set);
@@ -3579,14 +3579,14 @@ static rtx addr_placeholder;
 
 static void
 strength_reduce (scan_start, end, loop_top, insn_count,
-		 loop_start, loop_end, unroll_p)
+		 loop_start, loop_end, unroll_p, bct_p)
      rtx scan_start;
      rtx end;
      rtx loop_top;
      int insn_count;
      rtx loop_start;
      rtx loop_end;
-     int unroll_p;
+     int unroll_p, bct_p;
 {
   rtx p;
   rtx set;
@@ -4106,7 +4106,7 @@ strength_reduce (scan_start, end, loop_top, insn_count,
      the loop.  Unrolling may update part of this information, and the
      correct data will be used for generating the BCT.  */
 #ifdef HAVE_decrement_and_branch_on_count
-  if (HAVE_decrement_and_branch_on_count)
+  if (HAVE_decrement_and_branch_on_count && bct_p)
     analyze_loop_iterations (loop_start, loop_end);
 #endif
 #endif  /* HAIFA */
@@ -4613,7 +4613,7 @@ strength_reduce (scan_start, end, loop_top, insn_count,
 #ifdef HAIFA
   /* instrument the loop with bct insn */
 #ifdef HAVE_decrement_and_branch_on_count
-  if (HAVE_decrement_and_branch_on_count)
+  if (HAVE_decrement_and_branch_on_count && bct_p)
     insert_bct (loop_start, loop_end);
 #endif
 #endif  /* HAIFA */
@@ -6981,7 +6981,7 @@ check_dbra_loop (loop_end, insn_count, loop_start)
 		  /* If we have a decrement_and_branch_on_count, prefer
 		     the NE test, since this will allow that instruction to
 		     be generated.  */
-#if ! defined (HAVE_decrement_and_branch_on_zero) && defined (HAVE_decrement_and_branch_on_count)
+#if ! defined (HAVE_decrement_and_branch_until_zero) && defined (HAVE_decrement_and_branch_on_count)
 		  && (add_val != 1 || ! vtop)
 #endif
 		  && GET_CODE (comparison_value) == CONST_INT
@@ -8189,7 +8189,7 @@ insert_bct (loop_start, loop_end)
 
   /* the only machine mode we work with - is the integer of the size that the
      machine has */
-  enum machine_mode loop_var_mode = SImode;
+  enum machine_mode loop_var_mode = word_mode;
 
   int loop_num = uid_loop_num [INSN_UID (loop_start)];
 
@@ -8284,7 +8284,8 @@ insert_bct (loop_start, loop_end)
   /* try to instrument the loop.  */
 
   /* Handle the simpler case, where the bounds are known at compile time.  */
-  if (GET_CODE (initial_value) == CONST_INT && GET_CODE (comparison_value) == CONST_INT)
+  if (GET_CODE (initial_value) == CONST_INT
+      && GET_CODE (comparison_value) == CONST_INT)
     {
       int n_iterations;
       int increment_value_abs = INTVAL (increment) * increment_direction;
@@ -8459,15 +8460,15 @@ instrument_loop_bct (loop_start, loop_end, loop_num_iterations)
   rtx start_label;
 
   rtx sequence;
-  enum machine_mode loop_var_mode = SImode;
+  enum machine_mode loop_var_mode = word_mode;
 
   if (HAVE_decrement_and_branch_on_count)
     {
       if (loop_dump_stream)
 	fprintf (loop_dump_stream, "Loop: Inserting BCT\n");
 
-      /* eliminate the check on the old variable */
-      delete_insn (PREV_INSN (loop_end));
+      /* Discard original jump to continue loop.  Original compare result
+	 may still be live, so it cannot be discarded explicitly.  */
       delete_insn (PREV_INSN (loop_end));
 
       /* insert the label which will delimit the start of the loop */
@@ -8488,12 +8489,13 @@ instrument_loop_bct (loop_start, loop_end, loop_num_iterations)
 
       sequence = gen_sequence ();
       end_sequence ();
-      emit_insn_after (sequence, loop_start);
+      emit_insn_before (sequence, loop_start);
 
       /* insert new comparison on the count register instead of the
 	 old one, generating the needed BCT pattern (that will be
 	 later recognized by assembly generation phase).  */
-      emit_jump_insn_before (gen_decrement_and_branch_on_count (temp_reg2, start_label),
+      emit_jump_insn_before (gen_decrement_and_branch_on_count (temp_reg2,
+								start_label),
 			     loop_end);
       LABEL_NUSES (start_label)++;
     }
@@ -8935,4 +8937,3 @@ replace_label (x, data)
   return 0;
 }
 
-     

gcc/rtl.h

@@ -1350,7 +1350,7 @@ extern void print_inline_rtx		PROTO ((FILE *, rtx, int));
 /* In loop.c */
 extern void init_loop			PROTO ((void));
 #ifdef BUFSIZ
-extern void loop_optimize		PROTO ((rtx, FILE *, int));
+extern void loop_optimize		PROTO ((rtx, FILE *, int, int));
 #endif
 extern void record_excess_regs		PROTO ((rtx, rtx, rtx *));
 

gcc/toplev.c

@@ -3519,7 +3519,7 @@ rest_of_compilation (decl)
 	     {
 	       /* We only want to perform unrolling once.  */
 	       
-	       loop_optimize (insns, rtl_dump_file, 0);
+	       loop_optimize (insns, rtl_dump_file, 0, 0);
 	       
 	
 	       /* The first call to loop_optimize makes some instructions
@@ -3532,7 +3532,7 @@ rest_of_compilation (decl)
 		  analysis code depends on this information.  */
 	       reg_scan (insns, max_reg_num (), 1);
 	     }
-	   loop_optimize (insns, rtl_dump_file, flag_unroll_loops);
+	   loop_optimize (insns, rtl_dump_file, flag_unroll_loops, 1);
 	 });
       
       /* Dump rtl code after loop opt, if we are doing that.  */