lib2funcs.asm (__outline_prologue): New "function".

        * lib2funcs.asm (__outline_prologue): New "function".
        (__outline_epilogue): New "function".
        * pa.h (TARGET_SPACE): Define.
        (target_flags): Add -mspace and -mno-space.  Enable/disable
        space saving optimizations.
        (FRAME_POINTER_REQUIRED): Frame pointers are always required
        when generating out of line prologues and epilogues.
        * pa.c (compute_frame_size): Handle out of line prologues/epilogues.
        (hppa_expand_prologue): If optimizing for space, emit an out of
        line prologue.
        * pa.c (compute_frame_size): Handle out of line prologues/epilogues.
        (hppa_expand_prologue): If optimizing for space, emit an out of
        line prologue.
        (hppa_expand_epilogue): Similarly.
        (override_options): Optimizing for space is not compatable with
        either profiling or PIC code generation.
        * pa.md (outline_prologue_call): New pattern.
        (outline_epilogue_call): Likewise.

From-SVN: r11438

gcc/config/pa/lib2funcs.asm

@@ -24,6 +24,8 @@
 	.SPACE $TEXT$
 	.SUBSPA $LIT$,QUAD=0,ALIGN=8,ACCESS=44
 	.SUBSPA $CODE$,QUAD=0,ALIGN=8,ACCESS=44,CODE_ONLY
+	.SUBSPA $MILLICODE$,QUAD=0,ALIGN=8,ACCESS=44,SORT=8
+
 	.IMPORT $$dyncall,MILLICODE
 ; gcc_compiled.:
 	.SPACE $TEXT$
@@ -61,3 +63,251 @@ L$foo
 	be,n 0(%sr0,%rp)
 	.EXIT
 	.PROCEND
+
+	.SPACE $TEXT$
+	.SUBSPA $MILLICODE$
+
+; This is an out-of-line prologue.
+;
+; It performs the following operations:
+;
+;	* Saves the return pointer at sp - 20
+;
+;	* Creates a new stack frame (sp'), size of the frame is passed in %r21
+;
+;	* The old stack pointer is saved at sp
+;
+;	* Saves grs (passed in low 16 bits of %r22 into the stack frame
+;	at sp' + local_fsize (passed in %r19).
+;
+;	* Saves frs (passed in high 16 bits of %r22) into the stack
+;	frame at sp' + local_fsize (passed in %r19).
+;
+;	* Sets up a frame pointer (in %r3).
+;
+;	* Returns to the instruction _immediately_ after the call to
+;	this function.
+
+	.align 32
+	.NSUBSPA $MILLICODE$,QUAD=0,ALIGN=8,ACCESS=44,CODE_ONLY
+	.EXPORT __outline_prologue,MILLICODE
+__outline_prologue
+	.PROC
+	.CALLINFO FRAME=0,NO_CALLS
+	.ENTRY
+	copy %r30,%r20
+
+	; Subtract 4 from our return pointer so that we return to
+	; the right location.
+        ldo -4(%r31),%r31
+
+	; Save off %r2
+	stw %r2,-20(0,%r30)
+
+	; Make our new frame.
+	add %r21,%r30,%r30
+
+	; Save our old stack pointer.
+	stw %r20,0(0,%r20)
+
+	; Add in local_fsize to our frame pointer so we do register
+	; saves into the right place
+	add %r20,%r19,%r20
+
+	; %r22 tells us what registers we need to save.  The upper half
+	; is for fp registers, the lower half for integer registers.
+	; We put the lower half in %r1 and the upper half into %r22
+	; for later use.
+	extru %r22,31,16,%r1
+	extrs %r22,15,16,%r22
+
+	; %r1 now olds a value 0-18 which corresponds to the number
+	; of grs we need to save.  We need to reverse that value so
+	; we can just into the table and straight-line execute to the
+	; end of the gr saves.
+	comb,= %r0,%r1,L$0000
+	subi 18,%r1,%r1
+	blr,n %r1,%r0
+	b,n L$0000
+	stws,ma %r18,4(0,%r20)
+	nop
+	stws,ma %r17,4(0,%r20)
+	nop
+	stws,ma %r16,4(0,%r20)
+	nop
+	stws,ma %r15,4(0,%r20)
+	nop
+	stws,ma %r14,4(0,%r20)
+	nop
+	stws,ma %r13,4(0,%r20)
+	nop
+	stws,ma %r12,4(0,%r20)
+	nop
+	stws,ma %r11,4(0,%r20)
+	nop
+	stws,ma %r10,4(0,%r20)
+	nop
+	stws,ma %r9,4(0,%r20)
+	nop
+	stws,ma %r8,4(0,%r20)
+	nop
+	stws,ma %r7,4(0,%r20)
+	nop
+	stws,ma %r6,4(0,%r20)
+	nop
+	stws,ma %r5,4(0,%r20)
+	nop
+	stws,ma %r4,4(0,%r20)
+	nop
+	stws,ma %r3,4(0,%r20)
+	nop
+L$0000
+	; All gr saves are done.  Align the temporary frame pointer and
+	; do the fr saves.
+	ldo 7(%r20),%r20
+	depi 0,31,3,%r20
+
+	comb,= %r0,%r22,L$0001
+	subi 21,%r22,%r22
+	blr,n %r22,%r0
+	b,n L$0001
+	fstws,ma %fr21,8(0,%r20)
+	nop
+	fstws,ma %fr20,8(0,%r20)
+	nop
+	fstws,ma %fr19,8(0,%r20)
+	nop
+	fstws,ma %fr18,8(0,%r20)
+	nop
+	fstws,ma %fr17,8(0,%r20)
+	nop
+	fstws,ma %fr16,8(0,%r20)
+	nop
+	fstws,ma %fr15,8(0,%r20)
+	nop
+	fstws,ma %fr14,8(0,%r20)
+	nop
+	fstws,ma %fr13,8(0,%r20)
+	nop
+	fstws,ma %fr12,8(0,%r20)
+	nop
+L$0001
+	; Return, setting up a frame pointer in the delay slot
+	bv 0(%r31)
+	sub %r30,%r21,%r3
+
+	.EXIT
+	.PROCEND
+
+; This is an out-of-line epilogue.  It's operation is basically the reverse
+; of the out-of-line prologue.
+
+	.align 32
+	.NSUBSPA $MILLICODE$,QUAD=0,ALIGN=8,ACCESS=44,CODE_ONLY
+	.EXPORT __outline_epilogue,MILLICODE
+__outline_epilogue
+	.PROC
+	.CALLINFO FRAME=0,NO_CALLS
+	.ENTRY
+	; Make a copy of our frame pointer into %r20
+	copy %r3,%r20
+
+	; Subtract 4 from our return pointer so that we return to
+	; the right location.
+        ldo -4(%r31),%r31
+
+	; Reload %r2
+	; First save off %r2
+	ldw -20(0,%r20),%r2
+
+	; Load our old stack pointer, save it in %r21.
+	ldw 0(0,%r20),%r21
+
+	; Add in local_fsize (%r19) to the frame pointer to find
+	; the saved registers.
+	add %r20,%r19,%r20
+
+	; %r22 tells us what registers we need to restore.  The upper half
+	; is for fp registers, the lower half for integer registers.
+	; We put the lower half in %r1 and the upper half into %r22
+	; for later use.
+	extru %r22,31,16,%r1
+	extrs %r22,15,16,%r22
+
+	; %r1 now olds a value 0-18 which corresponds to the number
+	; of grs we need to restore.  We need to reverse that value so
+	; we can just into the table and straight-line execute to the
+	; end of the gr restore.
+	comb,= %r0,%r1,L$0002
+	subi 18,%r1,%r1
+	blr,n %r1,%r0
+	b,n L$0002
+	ldws,ma 4(0,%r20),%r18
+	nop
+	ldws,ma 4(0,%r20),%r17
+	nop
+	ldws,ma 4(0,%r20),%r16
+	nop
+	ldws,ma 4(0,%r20),%r15
+	nop
+	ldws,ma 4(0,%r20),%r14
+	nop
+	ldws,ma 4(0,%r20),%r13
+	nop
+	ldws,ma 4(0,%r20),%r12
+	nop
+	ldws,ma 4(0,%r20),%r11
+	nop
+	ldws,ma 4(0,%r20),%r10
+	nop
+	ldws,ma 4(0,%r20),%r9
+	nop
+	ldws,ma 4(0,%r20),%r8
+	nop
+	ldws,ma 4(0,%r20),%r7
+	nop
+	ldws,ma 4(0,%r20),%r6
+	nop
+	ldws,ma 4(0,%r20),%r5
+	nop
+	ldws,ma 4(0,%r20),%r4
+	nop
+	ldws,ma 4(0,%r20),%r3
+	nop
+L$0002
+	; All gr restore are done.  Align the temporary frame pointer and
+	; do the fr restore.
+	ldo 7(%r20),%r20
+	depi 0,31,3,%r20
+
+	comb,= %r0,%r22,L$0003
+	subi 21,%r22,%r22
+	blr,n %r22,%r0
+	b,n L$0003
+	fldws,ma 8(0,%r20),%fr21
+	nop
+	fldws,ma 8(0,%r20),%fr20
+	nop
+	fldws,ma 8(0,%r20),%fr19
+	nop
+	fldws,ma 8(0,%r20),%fr18
+	nop
+	fldws,ma 8(0,%r20),%fr17
+	nop
+	fldws,ma 8(0,%r20),%fr16
+	nop
+	fldws,ma 8(0,%r20),%fr15
+	nop
+	fldws,ma 8(0,%r20),%fr14
+	nop
+	fldws,ma 8(0,%r20),%fr13
+	nop
+	fldws,ma 8(0,%r20),%fr12
+	nop
+L$0003
+	; Return and deallocate our frame.
+	bv 0(%r31)
+	copy %r21,%r30
+	.EXIT
+	.PROCEND
+

gcc/config/pa/pa.c

@@ -113,6 +113,12 @@ override_options ()
     {
       warning ("PIC code generation is not compatable with profiling\n");
     }
+
+  if (TARGET_SPACE && (flag_pic || profile_flag))
+    {
+      warning ("Out of line entry/exit sequences are not compatable\n");
+      warning ("with PIC or profiling\n");
+    }
 }
 
 
@@ -2039,8 +2045,24 @@ compute_frame_size (size, fregs_live)
   for (i = 18; i >= 4; i--)
     {
       if (regs_ever_live[i])
+	{
+	  /* For out of line prologues/epilogues we only need to
+	     compute the highest register number to save and
+	     allocate space for all the callee saved registers
+	     with a lower number.  */
+	  if (TARGET_SPACE)
+	    {
+	      fsize += 4 * (i - 3);
+	      break;
+	    }
 	  fsize += 4;
 	}
+    }
+
+  /* We always save %r3, make room for it.  */
+  if (TARGET_SPACE)
+    fsize += 8;
+
   /* If we don't have a frame pointer, the register normally used for that
      purpose is saved just like other registers, not in the "frame marker".  */
   if (! frame_pointer_needed)
@@ -2053,9 +2075,19 @@ compute_frame_size (size, fregs_live)
   for (i = 66; i >= 48; i -= 2)
     if (regs_ever_live[i] || regs_ever_live[i + 1])
       {
-	fsize += 8;
 	if (fregs_live)
 	  *fregs_live = 1;
+
+	/* For out of line prologues/epilogues we only need to
+	   compute the highest register number to save and
+	   allocate space for all the callee saved registers
+	   with a lower number.  */
+        if (TARGET_SPACE)
+	  {
+	    fsize += 4 * (i - 46);
+	    break;
+	  }
+	fsize += 8;
       }
 
   fsize += current_function_outgoing_args_size;
@@ -2148,6 +2180,47 @@ hppa_expand_prologue()
   tmpreg = gen_rtx (REG, SImode, 1);
   size_rtx = GEN_INT (actual_fsize);
 
+  /* Handle out of line prologues and epilogues.  */
+  if (TARGET_SPACE)
+    {
+      rtx operands[2];
+      int saves = 0;
+
+      /* Put the local_fisze into %r19.  */
+      operands[0] = gen_rtx (REG, SImode, 19);
+      operands[1] = GEN_INT (local_fsize);
+      emit_move_insn (operands[0], operands[1]);
+
+      /* Put the stack size into %r21.  */
+      operands[0] = gen_rtx (REG, SImode, 21);
+      operands[1] = size_rtx;
+      emit_move_insn (operands[0], operands[1]);
+
+      /* Put the register save info into %r22.  */
+      for (i = 18; i >= 3; i--)
+	if (regs_ever_live[i] && ! call_used_regs[i])
+	  {
+	    saves = i;
+            break;
+	  }
+	  
+      for (i = 66; i >= 48; i -= 2)
+	if (regs_ever_live[i] || regs_ever_live[i + 1])
+	  {
+	    saves |= ((i/2 - 12 ) << 16);
+	    break;
+	  }
+
+      operands[0] = gen_rtx (REG, SImode, 22);
+      operands[1] = GEN_INT (saves);
+      emit_move_insn (operands[0], operands[1]);
+
+      /* Now call the out-of-line prologue.  */
+      emit_insn (gen_outline_prologue_call ());
+      emit_insn (gen_blockage ());
+      return;     
+    }
+
   /* Save RP first.  The calling conventions manual states RP will
      always be stored into the caller's frame at sp-20.  */
   if (regs_ever_live[2] || profile_flag)
@@ -2416,6 +2489,43 @@ hppa_expand_epilogue ()
   int offset,i;
   int merge_sp_adjust_with_load  = 0;
 
+  /* Handle out of line prologues and epilogues.  */
+  if (TARGET_SPACE)
+    {
+      int saves = 0;
+      rtx operands[2];
+
+      /* Put the register save info into %r22.  */
+      for (i = 18; i >= 3; i--)
+	if (regs_ever_live[i] && ! call_used_regs[i])
+	  {
+	    saves = i;
+            break;
+	  }
+	  
+      for (i = 66; i >= 48; i -= 2)
+	if (regs_ever_live[i] || regs_ever_live[i + 1])
+	  {
+	    saves |= ((i/2 - 12 ) << 16);
+	    break;
+	  }
+
+      emit_insn (gen_blockage ());
+
+      /* Put the local_fisze into %r19.  */
+      operands[0] = gen_rtx (REG, SImode, 19);
+      operands[1] = GEN_INT (local_fsize);
+      emit_move_insn (operands[0], operands[1]);
+
+      operands[0] = gen_rtx (REG, SImode, 22);
+      operands[1] = GEN_INT (saves);
+      emit_move_insn (operands[0], operands[1]);
+
+      /* Now call the out-of-line epilogue.  */
+      emit_insn (gen_outline_epilogue_call ());
+      return;
+    }
+
   /* We will use this often.  */
   tmpreg = gen_rtx (REG, SImode, 1);
 

gcc/config/pa/pa.h

@@ -78,6 +78,10 @@ extern int target_flags;
 /* Allow unconditional jumps in the delay slots of call instructions.  */
 #define TARGET_JUMP_IN_DELAY (target_flags & 8)
 
+/* Optimize for space.  Currently this only turns on out of line
+   prologues and epilogues.  */
+#define TARGET_SPACE (target_flags & 16)
+
 /* Disable indexed addressing modes.  */
 
 #define TARGET_DISABLE_INDEXING (target_flags & 32)
@@ -119,9 +123,11 @@ extern int target_flags;
    {"disable-fpregs", 2},	\
    {"no-disable-fpregs", -2},	\
    {"no-space-regs", 4},	\
-   {"space-regs", -4},\
+   {"space-regs", -4},		\
    {"jump-in-delay", 8},	\
    {"no-jump-in-delay", -8},	\
+   {"space", 16},		\
+   {"no-space", -16},		\
    {"disable-indexing", 32},	\
    {"no-disable-indexing", -32},\
    {"portable-runtime", 64},	\
@@ -554,9 +560,11 @@ do {								\
 /* Base register for access to local variables of the function.  */
 #define FRAME_POINTER_REGNUM 3
 
-/* Value should be nonzero if functions must have frame pointers. */
-#define FRAME_POINTER_REQUIRED (current_function_calls_alloca)
-
+/* Value should be nonzero if functions must have frame pointers.
+   All functions have frame pointers when optimizing for space
+   (for now).
+#define FRAME_POINTER_REQUIRED \
+  (current_function_calls_alloca || TARGET_SPACE)
 
 /* C statement to store the difference between the frame pointer
    and the stack pointer values immediately after the function prologue.

gcc/config/pa/pa.md

@@ -4837,3 +4837,65 @@
   "mfsp %%sr0,%4\;ldsid (0,%2),%3\;mtsp %3,%%sr0\;fic 0(%%sr0,%0)\;fic 0(%%sr0,%1)\;sync\;mtsp %4,%%sr0\;nop\;nop\;nop\;nop\;nop\;nop"
   [(set_attr "type" "multi")
    (set_attr "length" "52")])
+
+;; An out-of-line prologue.
+(define_insn "outline_prologue_call"
+  [(unspec_volatile [(const_int 0)] 0)
+   (clobber (reg:SI 31))
+   (clobber (reg:SI 22))
+   (clobber (reg:SI 21))
+   (clobber (reg:SI 20))
+   (clobber (reg:SI 19))
+   (clobber (reg:SI 1))]
+  ""
+  "*
+{
+  /* Must import the magic millicode routine.  */
+  output_asm_insn (\".IMPORT __outline_prologue,MILLICODE\", NULL);
+
+  /* The out-of-line prologue will make sure we return to the right
+     instruction.  */
+  if (TARGET_PORTABLE_RUNTIME)
+    {
+      output_asm_insn (\"ldil L'__outline_prologue,%%r31\", NULL);
+      output_asm_insn (\"ble,n R'__outline_prologue(%%sr0,%%r31)\", NULL);
+    }
+  else
+    output_asm_insn (\"bl,n __outline_prologue,%%r31\", NULL);
+  return \"\";
+}"
+  [(set_attr "type" "multi")
+   (set_attr "length" "8")])
+
+;; An out-of-line epilogue.
+(define_insn "outline_epilogue_call"
+  [(unspec_volatile [(const_int 1)] 0)
+   (use (reg:SI 29))
+   (use (reg:SI 28))
+   (clobber (reg:SI 31))
+   (clobber (reg:SI 22))
+   (clobber (reg:SI 21))
+   (clobber (reg:SI 20))
+   (clobber (reg:SI 19))
+   (clobber (reg:SI 2))
+   (clobber (reg:SI 1))]
+  ""
+  "*
+{
+  /* Must import the magic millicode routine.  */
+  output_asm_insn (\".IMPORT __outline_epilogue,MILLICODE\", NULL);
+
+  /* The out-of-line prologue will make sure we return to the right
+     instruction.  */
+  if (TARGET_PORTABLE_RUNTIME)
+    {
+      output_asm_insn (\"ldil L'__outline_epilogue,%%r31\", NULL);
+      output_asm_insn (\"ble,n R'__outline_epilogue(%%sr0,%%r31)\", NULL);
+    }
+  else
+    output_asm_insn (\"bl,n __outline_epilogue,%%r31\", NULL);
+  return \"\";
+}"
+  [(set_attr "type" "multi")
+   (set_attr "length" "8")])
+