* Automatic generation of inline long call sequences when needed.

	* pa.h (TARGET_LONG_CALLS): Delete.  Replace all uses of with
	TARGET_PORTABLE_RUNTIME.
	(TARGET_MILLICODE_LONG_CALLS): New target flag.
	* pa.c (output_function_prologue): Keep track of the total number
	of code bytes emitted for each source file.
	(output_call): Handle TARGET_PORTABLE_RUNTIME and millicode calls
	for TARGET_MILLICODE_LONG_CALLS.  Emit an inline long-call if
	needed.  If emitting an inline long-call, perform argument relocations
	before the call if they are needed, unfill the delay slot of the
	call if necessary.
	* pa.md (define_delay for millicode): Disable the delay slot if
	TARGET_MILLICODE_LONG_CALLS.
	(millicode insns and indirect calls): Properly compute length for
	both TARGET_PORTABLE_RUNTIME and TARGET_MILLICODE_LONG_CALLS.
	(call_internal_symref): Properly compute the length when more
	than 240000 bytes of code have already been output.  Take
	TARGET_MILLICODE_LONG_CALLS into account in the length
	computation.

From-SVN: r8422

gcc/config/pa/pa.c

@@ -52,6 +52,11 @@ static int gr_saved, fr_saved;
 
 static rtx find_addr_reg ();
 
+/* Keep track of the number of bytes we have output in the CODE subspaces
+   during this compilation so we'll know when to emit inline long-calls.  */
+
+unsigned int total_code_bytes;
+
 /* Return non-zero only if OP is a register of mode MODE,
    or CONST0_RTX.  */
 int
@@ -72,7 +77,7 @@ call_operand_address (op, mode)
      rtx op;
      enum machine_mode mode;
 {
-  return (CONSTANT_P (op) && ! TARGET_LONG_CALLS);
+  return (CONSTANT_P (op) && ! TARGET_PORTABLE_RUNTIME);
 }
 
 /* Return 1 if X contains a symbolic expression.  We know these
@@ -2044,6 +2049,19 @@ output_function_prologue (file, size)
   if (profile_flag)
     ASM_GENERATE_INTERNAL_LABEL (hp_profile_label_name, "LP",
 				 hp_profile_labelno);
+
+  if (insn_addresses)
+    {
+      unsigned int old_total = total_code_bytes;
+
+      total_code_bytes += insn_addresses[INSN_UID (get_last_insn())];
+      total_code_bytes += FUNCTION_BOUNDARY /BITS_PER_UNIT;
+
+      /* Be prepared to handle overflows.  */
+      total_code_bytes = old_total > total_code_bytes ? -1 : total_code_bytes;
+    }
+  else
+    total_code_bytes = -1;
 }
 
 void
@@ -3760,9 +3778,9 @@ output_movb (operands, insn, which_alternative, reverse_comparison)
    RETURN_POINTER is the register which will hold the return address.
    %r2 for most calls, %r31 for millicode calls. 
 
-   When TARGET_LONG_CALLS is true, output_call is only called for
-   millicode calls.  In addition, no delay slots are available when
-   TARGET_LONG_CALLS is true.  */
+   When TARGET_MILLICODE_LONG_CALLS is true, then we have to assume
+   that two instruction sequences must be used to reach the millicode
+   routines (including dyncall!).  */
 
 char *
 output_call (insn, call_dest, return_pointer)
@@ -3775,21 +3793,124 @@ output_call (insn, call_dest, return_pointer)
   rtx xoperands[4];
   rtx seq_insn;
 
-  /* Handle common case -- empty delay slot or no jump in the delay slot.  */
-  if (dbr_sequence_length () == 0
+  /* Handle long millicode calls for mod, div, and mul.  */
+  if (TARGET_PORTABLE_RUNTIME
+      || (TARGET_MILLICODE_LONG_CALLS && REGNO (return_pointer) == 31))
+    {
+      xoperands[0] = call_dest;
+      xoperands[1] = return_pointer;
+      output_asm_insn ("ldil L%%%0,%%r29", xoperands);
+      output_asm_insn ("ldo R%%%0(%%r29),%%r29", xoperands);
+      output_asm_insn ("blr 0,%r1\n\tbv,n 0(%%r29)\n\tnop", xoperands);
+      return "";
+    }
+
+  /* Handle common case -- empty delay slot or no jump in the delay slot,
+     and we're sure that the branch will reach the beginning of the $CODE$
+     subspace.  */
+  if ((dbr_sequence_length () == 0
+       && get_attr_length (insn) == 8)
       || (dbr_sequence_length () != 0
-	  && GET_CODE (NEXT_INSN (insn)) != JUMP_INSN))
+	  && GET_CODE (NEXT_INSN (insn)) != JUMP_INSN
+	  && get_attr_length (insn) == 4))
     {
       xoperands[0] = call_dest;
       xoperands[1] = return_pointer;
-      if (TARGET_LONG_CALLS)
+      output_asm_insn ("bl %0,%r1%#", xoperands);
+      return "";
+    }
+
+  /* This call may not reach the beginning of the $CODE$ subspace.  */
+  if (get_attr_length (insn) > 8)
+    {
+      int delay_insn_deleted = 0;
+      rtx xoperands[2];
+      rtx link;
+
+      /* We need to emit an inline long-call branch.  Furthermore,
+	 because we're changing a named function call into an indirect
+	 function call well after the parameters have been set up, we
+	 need to make sure any FP args appear in both the integer
+	 and FP registers.  Also, we need move any delay slot insn
+	 out of the delay slot -- Yuk!  */
+      if (dbr_sequence_length () != 0
+	  && GET_CODE (NEXT_INSN (insn)) != JUMP_INSN)
 	{
-	  output_asm_insn ("ldil L%%%0,%%r29", xoperands);
-	  output_asm_insn ("ldo R%%%0(%%r29),%%r29", xoperands);
-	  output_asm_insn ("blr 0,%r1\n\tbv,n 0(%%r29)\n\tnop", xoperands);
+	  /* A non-jump insn in the delay slot.  By definition we can
+	  emit this insn before the call (and in fact before argument
+	  relocating.  */
+	  final_scan_insn (NEXT_INSN (insn), asm_out_file, optimize, 0, 0);
+
+	  /* Now delete the delay insn.  */
+	  PUT_CODE (NEXT_INSN (insn), NOTE);
+	  NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED;
+	  NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0;
+	  delay_insn_deleted = 1;
+	}
+
+      /* Now copy any FP arguments into integer registers.  */
+      for (link = CALL_INSN_FUNCTION_USAGE (insn); link; link = XEXP (link, 1))
+	{
+	  int arg_mode, regno;
+	  rtx use = XEXP (link, 0);
+	  if (! (GET_CODE (use) == USE
+		 && GET_CODE (XEXP (use, 0)) == REG
+		 && FUNCTION_ARG_REGNO_P (REGNO (XEXP (use, 0)))))
+	    continue;
+
+	  arg_mode = GET_MODE (XEXP (use, 0));
+	  regno = REGNO (XEXP (use, 0));
+	  /* Is it a floating point register?  */
+	  if (regno >= 32 && regno <= 39)
+	    {
+	      /* Copy from the FP register into an integer register
+		 (via memory).  */
+	      if (arg_mode == SFmode)
+		{
+		  xoperands[0] = XEXP (use, 0);
+		  xoperands[1] = gen_rtx (REG, SImode, 26 - (regno - 32) / 2);
+		  output_asm_insn ("fstws %0,-16(%%sr0,%%r30)", xoperands);
+		  output_asm_insn ("ldw -16(%%sr0,%%r30),%1", xoperands);
+		}
+	      else
+		{
+		  xoperands[0] = XEXP (use, 0);
+		  xoperands[1] = gen_rtx (REG, DImode, 25 - (regno - 34) / 2);
+		  output_asm_insn ("fstds %0,-16(%%sr0,%%r30)", xoperands);
+		  output_asm_insn ("ldw -12(%%sr0,%%r30),%R1", xoperands);
+		  output_asm_insn ("ldw -16(%%sr0,%%r30),%1", xoperands);
+		}
+		
+	    }
+	}
+
+      /* Now emit the inline long-call.  */
+      xoperands[0] = call_dest;
+      output_asm_insn ("ldil L%%%0,%%r22\n\tldo R%%%0(%%r22),%%r22", xoperands);
+
+      /* If TARGET_MILLICODE_LONG_CALLS, then we must use a long-call sequence
+	 to call dyncall!  */
+      if (TARGET_MILLICODE_LONG_CALLS)
+	{
+	  output_asm_insn ("ldil L%%$$dyncall,%%r31", xoperands);
+	  output_asm_insn ("ldo R%%$$dyncall(%%r31),%%r31", xoperands);
+	  output_asm_insn ("blr 0,%%r2\n\tbv,n 0(%%r31)\n\tnop", xoperands);
 	}
       else
-	output_asm_insn ("bl %0,%r1%#", xoperands);
+	output_asm_insn ("bl $$dyncall,%%r31\n\tcopy %%r31,%%r2", xoperands);
+
+      /* If we had a jump in the call's delay slot, output it now.  */
+      if (dbr_sequence_length () != 0
+	  && !delay_insn_deleted)
+	{
+	  xoperands[0] = XEXP (PATTERN (NEXT_INSN (insn)), 1);
+	  output_asm_insn ("b,n %0", xoperands);
+
+	  /* Now delete the delay insn.  */
+	  PUT_CODE (NEXT_INSN (insn), NOTE);
+	  NOTE_LINE_NUMBER (NEXT_INSN (insn)) = NOTE_INSN_DELETED;
+	  NOTE_SOURCE_FILE (NEXT_INSN (insn)) = 0;
+	}
       return "";
     }
 

gcc/config/pa/pa.h

@@ -28,6 +28,9 @@ enum cmp_type				/* comparison type */
   CMP_MAX				/* max comparison type */
 };
 
+/* For long call handling.  */
+extern unsigned int total_code_bytes;
+
 /* Print subsidiary information on the compiler version in use.  */
 
 #define TARGET_VERSION fprintf (stderr, " (hppa)");
@@ -57,13 +60,18 @@ extern int target_flags;
 /* Allow unconditional jumps in the delay slots of call instructions.  */
 #define TARGET_JUMP_IN_DELAY (target_flags & 8)
 
-/* Force all function calls to indirect addressing via a register.  This
-   avoids lossage when the function is very far away from the current PC.
+/* In rare cases, a millicode call via "bl" can not be turned into
+   a millicode call using "ble" (when SHLIB_INFO subspace is very large).
+
+   This option forces just millicode calls to use inline long-calls
+   This is far more efficient than the old long-call option which forced
+   every function to be called indirectly (as is still the case for
+   TARGET_PORTABLE_RUNTIME).
 
    ??? What about simple jumps, they can suffer from the same problem.
    Would require significant surgery in pa.md.  */
 
-#define TARGET_LONG_CALLS (target_flags & 16)
+#define TARGET_MILLICODE_LONG_CALLS (target_flags & 16)
 
 /* Disable indexed addressing modes.  */
 
@@ -73,7 +81,8 @@ extern int target_flags;
    HP wants everyone to use for ELF objects.  If at all possible you want
    to avoid this since it's a performance loss for non-prototyped code.
 
-   Note TARGET_PORTABLE_RUNTIME also implies TARGET_LONG_CALLS.  */
+   Note TARGET_PORTABLE_RUNTIME also forces all calls to use inline
+   long-call stubs which is quite expensive.  */
 
 #define TARGET_PORTABLE_RUNTIME (target_flags & 64)
 
@@ -100,8 +109,8 @@ extern int target_flags;
    {"no-fast-indirect-calls", -4},\
    {"jump-in-delay", 8},	\
    {"no-jump-in-delay", -8},	\
-   {"long-calls", 16},		\
-   {"no-long-calls", -16},	\
+   {"millicode-long-calls", 16},\
+   {"no-millicode-long-calls", -16},\
    {"disable-indexing", 32},	\
    {"no-disable-indexing", -32},\
    {"portable-runtime", 64+16},\
@@ -832,9 +841,7 @@ struct hppa_args {int words, nargs_prototype; };
    The caller must make a distinction between calls to explicitly named
    functions and calls through pointers to functions -- the conventions
    are different!  Calls through pointers to functions only use general
-   registers for the first four argument words.  Note the indirect function
-   calling conventions are in effect during TARGET_LONG_CALLS, but 
-   current_call_is_indirect will not be set in such situations. 
+   registers for the first four argument words.
 
    Of course all this is different for the portable runtime model
    HP wants everyone to use for ELF.  Ugh.  Here's a quick description
@@ -869,12 +876,12 @@ struct hppa_args {int words, nargs_prototype; };
       || !FLOAT_MODE_P (MODE) || (CUM).nargs_prototype > 0)		\
       ? gen_rtx (REG, (MODE),						\
 		 (FUNCTION_ARG_SIZE ((MODE), (TYPE)) > 1		\
-		  ? (((!(current_call_is_indirect || TARGET_LONG_CALLS)	\
+		  ? (((!current_call_is_indirect 			\
 		       || TARGET_PORTABLE_RUNTIME)			\
 		      && (MODE) == DFmode)				\
 		     ? ((CUM).words ? 38 : 34)				\
 		     : ((CUM).words ? 23 : 25))				\
-		  : (((!(current_call_is_indirect || TARGET_LONG_CALLS)	\
+		  : (((!current_call_is_indirect			\
 		       || TARGET_PORTABLE_RUNTIME)			\
 		      && (MODE) == SFmode)				\
 		     ? (32 + 2 * (CUM).words)				\

gcc/config/pa/pa.md

@@ -93,10 +93,11 @@
   [(eq_attr "in_call_delay" "true") (nil) (nil)])
 
 ;; millicode call delay slot description.  Note it disallows delay slot
-;; when TARGET_LONG_CALLS is true.
+;; when TARGET_PORTABLE_RUNTIME or TARGET_MILLICODE_LONG_CALLS is true.
 (define_delay (eq_attr "type" "milli")
   [(and (eq_attr "in_call_delay" "true")
-	(eq (symbol_ref "TARGET_LONG_CALLS") (const_int 0)))
+	(and (eq (symbol_ref "TARGET_PORTABLE_RUNTIME") (const_int 0))
+	     (eq (symbol_ref "TARGET_MILLICODE_LONG_CALLS") (const_int 0))))
    (nil) (nil)])
 
 ;; Unconditional branch, return and other similar instructions.
@@ -2268,10 +2269,13 @@
   ""
   "* return output_mul_insn (0, insn);"
   [(set_attr "type" "milli")
-   (set (attr "length") (if_then_else (ne (symbol_ref "TARGET_LONG_CALLS")
-					  (const_int 0))
-				      (const_int 4)
-				      (const_int 24)))])
+   (set (attr "length")
+     (if_then_else (and (eq (symbol_ref "TARGET_PORTABLE_RUNTIME")
+			    (const_int 0))
+			(eq (symbol_ref "TARGET_MILLICODE_LONG_CALLS")
+			    (const_int 0)))
+		   (const_int 4)
+		   (const_int 24)))])
 
 ;;; Division and mod.
 (define_expand "divsi3"
@@ -2318,10 +2322,13 @@
   "*
    return output_div_insn (operands, 0, insn);"
   [(set_attr "type" "milli")
-   (set (attr "length") (if_then_else (ne (symbol_ref "TARGET_LONG_CALLS")
-					  (const_int 0))
-				      (const_int 4)
-				      (const_int 24)))])
+   (set (attr "length")
+     (if_then_else (and (eq (symbol_ref "TARGET_PORTABLE_RUNTIME")
+			    (const_int 0))
+			(eq (symbol_ref "TARGET_MILLICODE_LONG_CALLS")
+			    (const_int 0)))
+		   (const_int 4)
+		   (const_int 24)))])
 
 (define_expand "udivsi3"
   [(set (reg:SI 26) (match_operand:SI 1 "move_operand" ""))
@@ -2367,10 +2374,13 @@
   "*
    return output_div_insn (operands, 1, insn);"
   [(set_attr "type" "milli")
-   (set (attr "length") (if_then_else (ne (symbol_ref "TARGET_LONG_CALLS")
-					  (const_int 0))
-				      (const_int 4)
-				      (const_int 24)))])
+   (set (attr "length")
+     (if_then_else (and (eq (symbol_ref "TARGET_PORTABLE_RUNTIME")
+			    (const_int 0))
+			(eq (symbol_ref "TARGET_MILLICODE_LONG_CALLS")
+			    (const_int 0)))
+		   (const_int 4)
+		   (const_int 24)))])
 
 (define_expand "modsi3"
   [(set (reg:SI 26) (match_operand:SI 1 "move_operand" ""))
@@ -2412,10 +2422,13 @@
   "*
   return output_mod_insn (0, insn);"
   [(set_attr "type" "milli")
-   (set (attr "length") (if_then_else (ne (symbol_ref "TARGET_LONG_CALLS")
-					  (const_int 0))
-				      (const_int 4)
-				      (const_int 24)))])
+   (set (attr "length")
+     (if_then_else (and (eq (symbol_ref "TARGET_PORTABLE_RUNTIME")
+			    (const_int 0))
+			(eq (symbol_ref "TARGET_MILLICODE_LONG_CALLS")
+			    (const_int 0)))
+		   (const_int 4)
+		   (const_int 24)))])
 
 (define_expand "umodsi3"
   [(set (reg:SI 26) (match_operand:SI 1 "move_operand" ""))
@@ -2457,10 +2470,13 @@
   "*
   return output_mod_insn (1, insn);"
   [(set_attr "type" "milli")
-   (set (attr "length") (if_then_else (ne (symbol_ref "TARGET_LONG_CALLS")
-					  (const_int 0))
-				      (const_int 4)
-				      (const_int 24)))])
+   (set (attr "length")
+     (if_then_else (and (eq (symbol_ref "TARGET_PORTABLE_RUNTIME")
+			    (const_int 0))
+			(eq (symbol_ref "TARGET_MILLICODE_LONG_CALLS")
+			    (const_int 0)))
+		   (const_int 4)
+		   (const_int 24)))])
 
 ;;- and instructions
 ;; We define DImode `and` so with DImode `not` we can get
@@ -3143,7 +3159,7 @@
   rtx op;
   rtx call_insn;
 
-  if (TARGET_LONG_CALLS)
+  if (TARGET_PORTABLE_RUNTIME)
     op = force_reg (SImode, XEXP (operands[0], 0));
   else
     op = XEXP (operands[0], 0);
@@ -3185,14 +3201,21 @@
 	 (match_operand 1 "" "i"))
    (clobber (reg:SI 2))
    (use (const_int 0))]
-  "! TARGET_LONG_CALLS"
+  "! TARGET_PORTABLE_RUNTIME"
   "*
 {
   output_arg_descriptor (insn);
   return output_call (insn, operands[0], gen_rtx (REG, SImode, 2));
 }"
   [(set_attr "type" "call")
-   (set_attr "length" "4")])
+   (set (attr "length")
+      (if_then_else (lt (plus (symbol_ref "total_code_bytes") (pc))
+			(const_int 240000))
+		    (const_int 4)
+		    (if_then_else (ne (symbol_ref "TARGET_MILLICODE_LONG_CALLS")
+				      (const_int 0))
+				  (const_int 64)
+				  (const_int 52))))])
 
 (define_insn "call_internal_reg"
   [(call (mem:SI (match_operand:SI 0 "register_operand" "r"))
@@ -3206,16 +3229,19 @@
     return \"blr 0,%%r2\;bv,n 0(%r0)\;ldo 4(%%r2),%%r2\";
 
   /* Yuk!  bl may not be able to reach $$dyncall.  */
-  if (TARGET_LONG_CALLS)
+  if (TARGET_PORTABLE_RUNTIME || TARGET_MILLICODE_LONG_CALLS)
     return \"copy %r0,%%r22\;ldil L%%$$dyncall,%%r31\;ldo R%%$$dyncall(%%r31),%%r31\;blr 0,%%r2\;bv,n 0(%%r31)\;nop\";
   else
     return \"copy %r0,%%r22\;.CALL\\tARGW0=GR\;bl $$dyncall,%%r31\;copy %%r31,%%r2\";
 }"
   [(set_attr "type" "dyncall")
-   (set (attr "length") (if_then_else (ne (symbol_ref "TARGET_LONG_CALLS")
-					  (const_int 0))
-				      (const_int 12)
-				      (const_int 24)))])
+   (set (attr "length")
+     (if_then_else (and (ne (symbol_ref "TARGET_PORTABLE_RUNTIME")
+			    (const_int 0))
+			(ne (symbol_ref "TARGET_MILLICODE_LONG_CALLS")
+			    (const_int 0)))
+		   (const_int 12)
+		   (const_int 24)))])
 
 (define_expand "call_value"
   [(parallel [(set (match_operand 0 "" "")
@@ -3228,7 +3254,7 @@
   rtx op;
   rtx call_insn;
 
-  if (TARGET_LONG_CALLS)
+  if (TARGET_PORTABLE_RUNTIME)
     op = force_reg (SImode, XEXP (operands[1], 0));
   else
     op = XEXP (operands[1], 0);
@@ -3275,14 +3301,21 @@
    (clobber (reg:SI 2))
    (use (const_int 0))]
   ;;- Don't use operand 1 for most machines.
-  "! TARGET_LONG_CALLS"
+  "! TARGET_PORTABLE_RUNTIME"
   "*
 {
   output_arg_descriptor (insn);
   return output_call (insn, operands[1], gen_rtx (REG, SImode, 2));
 }"
   [(set_attr "type" "call")
-   (set_attr "length" "4")])
+   (set (attr "length")
+      (if_then_else (lt (plus (symbol_ref "total_code_bytes") (pc))
+			(const_int 240000))
+		    (const_int 4)
+		    (if_then_else (ne (symbol_ref "TARGET_MILLICODE_LONG_CALLS")
+				      (const_int 0))
+				  (const_int 64)
+				  (const_int 52))))])
 
 (define_insn "call_value_internal_reg"
   [(set (match_operand 0 "" "=rf")
@@ -3297,16 +3330,19 @@
     return \"blr 0,%%r2\;bv,n 0(%r1)\;ldo 4(%%r2),%%r2\";
 
   /* Yuk!  bl may not be able to reach $$dyncall.  */
-  if (TARGET_LONG_CALLS)
+  if (TARGET_PORTABLE_RUNTIME || TARGET_MILLICODE_LONG_CALLS)
     return \"copy %r1,%%r22\;ldil L%%$$dyncall,%%r31\;ldo R%%$$dyncall(%%r31),%%r31\;blr 0,%%r2\;bv,n 0(%%r31)\;nop\";
   else
     return \"copy %r1,%%r22\;.CALL\\tARGW0=GR\;bl $$dyncall,%%r31\;copy %%r31,%%r2\";
 }"
   [(set_attr "type" "dyncall")
-   (set (attr "length") (if_then_else (ne (symbol_ref "TARGET_LONG_CALLS")
-					  (const_int 0))
-				      (const_int 12)
-				      (const_int 24)))])
+   (set (attr "length")
+     (if_then_else (and (ne (symbol_ref "TARGET_PORTABLE_RUNTIME")
+			    (const_int 0))
+			(ne (symbol_ref "TARGET_MILLICODE_LONG_CALLS")
+			    (const_int 0)))
+		   (const_int 12)
+		   (const_int 24)))])
 
 ;; Call subroutine returning any type.