common.opt (flra-remat): New.

2014-11-12  Vladimir Makarov  <vmakarov@redhat.com>

	* common.opt (flra-remat): New.
	* opts.c (default_options_table): Add entry for flra_remat.
	* timevar_def (TV_LRA_REMAT): New.
	* doc/invoke.texi (-flra-remat): Add description of the new
	option.
	* doc/passes.texi (-flra-remat): Remove lra-equivs.c and
	lra-saves.c.  Add lra-remat.c.
	* Makefile.in (OBJS): Add lra-remat.o.
	* lra-remat.c: New file.
	* lra.c: Add info about the rematerialization pass in the top
	comment.
	(collect_non_operand_hard_regs, add_regs_to_insn_regno_info):
	Process unallocatable regs too.
	(lra_constraint_new_insn_uid_start): Remove.
	(lra): Add code for calling rematerialization sub-pass.
	* lra-int.h (lra_constraint_new_insn_uid_start): Remove.
	(lra_constrain_insn, lra_remat): New prototypes.
	(lra_eliminate_regs_1): Add parameter.
	* lra-lives.c (make_hard_regno_born, make_hard_regno_dead):
	Process unallocatable hard regs too.
	(process_bb_lives): Ditto.
	* lra-spills.c (remove_pseudos): Add argument to
	lra_eliminate_regs_1 call.
	* lra-eliminations.c (lra_eliminate_regs_1): Add parameter.  Use it
	for sp offset calculation.
	(lra_eliminate_regs): Add argument for lra_eliminate_regs_1 call.
	(eliminate_regs_in_insn): Add parameter.  Use it for sp offset
	calculation.
	(process_insn_for_elimination): Add argument for
	eliminate_regs_in_insn call.
	* lra-constraints.c (get_equiv_with_elimination):  Add argument
	for lra_eliminate_regs_1 call.
	(process_addr_reg): Add parameter.  Use it.
	(process_address_1): Ditto.  Add argument for process_addr_reg
	call.
	(process_address): Ditto.
	(curr_insn_transform): Add parameter.  Use it.  Add argument for
	process_address calls.
	(lra_constrain_insn): New function.
	(lra_constraints): Add argument for curr_insn_transform call.

From-SVN: r217458

gcc/ChangeLog

+2014-11-12  Vladimir Makarov  <vmakarov@redhat.com>
+
+	* common.opt (flra-remat): New.
+	* opts.c (default_options_table): Add entry for flra_remat.
+	* timevar_def (TV_LRA_REMAT): New.
+	* doc/invoke.texi (-flra-remat): Add description of the new
+	option.
+	* doc/passes.texi (-flra-remat): Remove lra-equivs.c and
+	lra-saves.c.  Add lra-remat.c.
+	* Makefile.in (OBJS): Add lra-remat.o.
+	* lra-remat.c: New file.
+	* lra.c: Add info about the rematerialization pass in the top
+	comment.
+	(collect_non_operand_hard_regs, add_regs_to_insn_regno_info):
+	Process unallocatable regs too.
+	(lra_constraint_new_insn_uid_start): Remove.
+	(lra): Add code for calling rematerialization sub-pass.
+	* lra-int.h (lra_constraint_new_insn_uid_start): Remove.
+	(lra_constrain_insn, lra_remat): New prototypes.
+	(lra_eliminate_regs_1): Add parameter.
+	* lra-lives.c (make_hard_regno_born, make_hard_regno_dead):
+	Process unallocatable hard regs too.
+	(process_bb_lives): Ditto.
+	* lra-spills.c (remove_pseudos): Add argument to
+	lra_eliminate_regs_1 call.
+	* lra-eliminations.c (lra_eliminate_regs_1): Add parameter.  Use it
+	for sp offset calculation.
+	(lra_eliminate_regs): Add argument for lra_eliminate_regs_1 call.
+	(eliminate_regs_in_insn): Add parameter.  Use it for sp offset
+	calculation.
+	(process_insn_for_elimination): Add argument for
+	eliminate_regs_in_insn call.
+	* lra-constraints.c (get_equiv_with_elimination):  Add argument
+	for lra_eliminate_regs_1 call.
+	(process_addr_reg): Add parameter.  Use it.
+	(process_address_1): Ditto.  Add argument for process_addr_reg
+	call.
+	(process_address): Ditto.
+	(curr_insn_transform): Add parameter.  Use it.  Add argument for
+	process_address calls.
+	(lra_constrain_insn): New function.
+	(lra_constraints): Add argument for curr_insn_transform call.
+
 2014-11-13  Manuel López-Ibáñez  <manu@gcc.gnu.org>
 
 	* opts-global.c (postpone_unknown_option_warning): Fix spelling.

gcc/Makefile.in

@@ -1304,6 +1304,7 @@ OBJS = \
 	lra-constraints.o \
 	lra-eliminations.o \
 	lra-lives.o \
+	lra-remat.o \
 	lra-spills.o \
 	lto-cgraph.o \
 	lto-streamer.o \

gcc/common.opt

@@ -1551,6 +1551,10 @@ floop-optimize
 Common Ignore
 Does nothing.  Preserved for backward compatibility.
 
+flra-remat
+Common Report Var(flag_lra_remat) Optimization
+Do CFG-sensitive rematerialization in LRA
+
 flto
 Common
 Enable link-time optimization.

gcc/doc/invoke.texi

@@ -392,7 +392,7 @@ Objective-C and Objective-C++ Dialects}.
 -fisolate-erroneous-paths-dereference -fisolate-erroneous-paths-attribute @gol
 -fivopts -fkeep-inline-functions -fkeep-static-consts -flive-range-shrinkage @gol
 -floop-block -floop-interchange -floop-strip-mine -floop-nest-optimize @gol
--floop-parallelize-all -flto -flto-compression-level @gol
+-floop-parallelize-all -flra-remat -flto -flto-compression-level @gol
 -flto-partition=@var{alg} -flto-report -flto-report-wpa -fmerge-all-constants @gol
 -fmerge-constants -fmodulo-sched -fmodulo-sched-allow-regmoves @gol
 -fmove-loop-invariants -fno-branch-count-reg @gol
@@ -7183,6 +7183,7 @@ also turns on the following optimization flags:
 -fipa-sra @gol
 -fipa-icf @gol
 -fisolate-erroneous-paths-dereference @gol
+-flra-remat @gol
 -foptimize-sibling-calls @gol
 -foptimize-strlen @gol
 -fpartial-inlining @gol
@@ -7811,6 +7812,14 @@ Control the verbosity of the dump file for the integrated register allocator.
 The default value is 5.  If the value @var{n} is greater or equal to 10,
 the dump output is sent to stderr using the same format as @var{n} minus 10.
 
+@item -flra-remat
+@opindex fcaller-saves
+Enable CFG-sensitive rematerialization in LRA.  Instead of loading
+values of spilled pseudos, LRA tries to rematerialize (recalculate)
+values if it is profitable.
+
+Enabled at levels @option{-O2}, @option{-O3}, @option{-Os}.
+
 @item -fdelayed-branch
 @opindex fdelayed-branch
 If supported for the target machine, attempt to reorder instructions

gcc/doc/passes.texi

@@ -911,10 +911,10 @@ Source files are @file{reload.c} and @file{reload1.c}, plus the header
 This pass is a modern replacement of the reload pass.  Source files
 are @file{lra.c}, @file{lra-assign.c}, @file{lra-coalesce.c},
 @file{lra-constraints.c}, @file{lra-eliminations.c},
-@file{lra-equivs.c}, @file{lra-lives.c}, @file{lra-saves.c},
-@file{lra-spills.c}, the header @file{lra-int.h} used for
-communication between them, and the header @file{lra.h} used for
-communication between LRA and the rest of compiler.
+@file{lra-lives.c}, @file{lra-remat.c}, @file{lra-spills.c}, the
+header @file{lra-int.h} used for communication between them, and the
+header @file{lra.h} used for communication between LRA and the rest of
+compiler.
 
 Unlike the reload pass, intermediate LRA decisions are reflected in
 RTL as much as possible.  This reduces the number of target-dependent

gcc/lra-eliminations.c

@@ -298,7 +298,8 @@ get_elimination (rtx reg)
    a change in the offset between the eliminable register and its
    substitution if UPDATE_P, or the full offset if FULL_P, or
    otherwise zero.  If FULL_P, we also use the SP offsets for
-   elimination to SP.
+   elimination to SP.  If UPDATE_P, use UPDATE_SP_OFFSET for updating
+   offsets of register elimnable to SP.
 
    MEM_MODE is the mode of an enclosing MEM.  We need this to know how
    much to adjust a register for, e.g., PRE_DEC.  Also, if we are
@@ -311,7 +312,8 @@ get_elimination (rtx reg)
    sp offset.  */
 rtx
 lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
-		      bool subst_p, bool update_p, bool full_p)
+		      bool subst_p, bool update_p,
+		      HOST_WIDE_INT update_sp_offset, bool full_p)
 {
   enum rtx_code code = GET_CODE (x);
   struct lra_elim_table *ep;
@@ -346,7 +348,10 @@ lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
 	  rtx to = subst_p ? ep->to_rtx : ep->from_rtx;
 
 	  if (update_p)
-	    return plus_constant (Pmode, to, ep->offset - ep->previous_offset);
+	    return plus_constant (Pmode, to,
+				  ep->offset - ep->previous_offset
+				  + (ep->to_rtx == stack_pointer_rtx
+				     ? update_sp_offset : 0));
 	  else if (full_p)
 	    return plus_constant (Pmode, to,
 				  ep->offset
@@ -373,7 +378,10 @@ lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
 		return gen_rtx_PLUS (Pmode, to, XEXP (x, 1));
 
 	      offset = (update_p
-			? ep->offset - ep->previous_offset : ep->offset);
+			? ep->offset - ep->previous_offset
+			+ (ep->to_rtx == stack_pointer_rtx
+			   ? update_sp_offset : 0)
+			: ep->offset);
 	      if (full_p && insn != NULL_RTX && ep->to_rtx == stack_pointer_rtx)
 		offset -= lra_get_insn_recog_data (insn)->sp_offset;
 	      if (CONST_INT_P (XEXP (x, 1))
@@ -402,9 +410,11 @@ lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
 
       {
 	rtx new0 = lra_eliminate_regs_1 (insn, XEXP (x, 0), mem_mode,
-					 subst_p, update_p, full_p);
+					 subst_p, update_p,
+					 update_sp_offset, full_p);
 	rtx new1 = lra_eliminate_regs_1 (insn, XEXP (x, 1), mem_mode,
-					 subst_p, update_p, full_p);
+					 subst_p, update_p,
+					 update_sp_offset, full_p);
 
 	if (new0 != XEXP (x, 0) || new1 != XEXP (x, 1))
 	  return form_sum (new0, new1);
@@ -423,11 +433,12 @@ lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
 	  rtx to = subst_p ? ep->to_rtx : ep->from_rtx;
 
 	  if (update_p)
-	    return
-	      plus_constant (Pmode,
-			     gen_rtx_MULT (Pmode, to, XEXP (x, 1)),
-			     (ep->offset - ep->previous_offset)
-			     * INTVAL (XEXP (x, 1)));
+	    return plus_constant (Pmode,
+				  gen_rtx_MULT (Pmode, to, XEXP (x, 1)),
+				  (ep->offset - ep->previous_offset
+				   + (ep->to_rtx == stack_pointer_rtx
+				      ? update_sp_offset : 0))
+				  * INTVAL (XEXP (x, 1)));
 	  else if (full_p)
 	    {
 	      HOST_WIDE_INT offset = ep->offset;
@@ -459,10 +470,12 @@ lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
     case LE:	   case LT:	  case LEU:    case LTU:
       {
 	rtx new0 = lra_eliminate_regs_1 (insn, XEXP (x, 0), mem_mode,
-					 subst_p, update_p, full_p);
+					 subst_p, update_p, 
+					 update_sp_offset, full_p);
 	rtx new1 = XEXP (x, 1)
 		   ? lra_eliminate_regs_1 (insn, XEXP (x, 1), mem_mode,
-					   subst_p, update_p, full_p) : 0;
+					   subst_p, update_p,
+					   update_sp_offset, full_p) : 0;
 
 	if (new0 != XEXP (x, 0) || new1 != XEXP (x, 1))
 	  return gen_rtx_fmt_ee (code, GET_MODE (x), new0, new1);
@@ -475,7 +488,8 @@ lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
       if (XEXP (x, 0))
 	{
 	  new_rtx = lra_eliminate_regs_1 (insn, XEXP (x, 0), mem_mode,
-					  subst_p, update_p, full_p);
+					  subst_p, update_p,
+					  update_sp_offset, full_p);
 	  if (new_rtx != XEXP (x, 0))
 	    {
 	      /* If this is a REG_DEAD note, it is not valid anymore.
@@ -484,7 +498,8 @@ lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
 	      if (REG_NOTE_KIND (x) == REG_DEAD)
 		return (XEXP (x, 1)
 			? lra_eliminate_regs_1 (insn, XEXP (x, 1), mem_mode,
-						subst_p, update_p, full_p)
+						subst_p, update_p,
+						update_sp_offset, full_p)
 			: NULL_RTX);
 
 	      x = alloc_reg_note (REG_NOTE_KIND (x), new_rtx, XEXP (x, 1));
@@ -501,7 +516,8 @@ lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
       if (XEXP (x, 1))
 	{
 	  new_rtx = lra_eliminate_regs_1 (insn, XEXP (x, 1), mem_mode,
-					  subst_p, update_p, full_p);
+					  subst_p, update_p,
+					  update_sp_offset, full_p);
 	  if (new_rtx != XEXP (x, 1))
 	    return
 	      gen_rtx_fmt_ee (GET_CODE (x), GET_MODE (x),
@@ -528,8 +544,8 @@ lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
 	  && XEXP (XEXP (x, 1), 0) == XEXP (x, 0))
 	{
 	  rtx new_rtx = lra_eliminate_regs_1 (insn, XEXP (XEXP (x, 1), 1),
-					      mem_mode,
-					      subst_p, update_p, full_p);
+					      mem_mode, subst_p, update_p,
+					      update_sp_offset, full_p);
 
 	  if (new_rtx != XEXP (XEXP (x, 1), 1))
 	    return gen_rtx_fmt_ee (code, GET_MODE (x), XEXP (x, 0),
@@ -553,14 +569,16 @@ lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
     case PARITY:
     case BSWAP:
       new_rtx = lra_eliminate_regs_1 (insn, XEXP (x, 0), mem_mode,
-				      subst_p, update_p, full_p);
+				      subst_p, update_p,
+				      update_sp_offset, full_p);
       if (new_rtx != XEXP (x, 0))
 	return gen_rtx_fmt_e (code, GET_MODE (x), new_rtx);
       return x;
 
     case SUBREG:
       new_rtx = lra_eliminate_regs_1 (insn, SUBREG_REG (x), mem_mode,
-				      subst_p, update_p, full_p);
+				      subst_p, update_p,
+				      update_sp_offset, full_p);
 
       if (new_rtx != SUBREG_REG (x))
 	{
@@ -598,12 +616,12 @@ lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
 	replace_equiv_address_nv
 	(x,
 	 lra_eliminate_regs_1 (insn, XEXP (x, 0), GET_MODE (x),
-			       subst_p, update_p, full_p));
+			       subst_p, update_p, update_sp_offset, full_p));
 
     case USE:
       /* Handle insn_list USE that a call to a pure function may generate.  */
       new_rtx = lra_eliminate_regs_1 (insn, XEXP (x, 0), VOIDmode,
-				      subst_p, update_p, full_p);
+				      subst_p, update_p, update_sp_offset, full_p);
       if (new_rtx != XEXP (x, 0))
 	return gen_rtx_USE (GET_MODE (x), new_rtx);
       return x;
@@ -624,7 +642,8 @@ lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
       if (*fmt == 'e')
 	{
 	  new_rtx = lra_eliminate_regs_1 (insn, XEXP (x, i), mem_mode,
-					  subst_p, update_p, full_p);
+					  subst_p, update_p,
+					  update_sp_offset, full_p);
 	  if (new_rtx != XEXP (x, i) && ! copied)
 	    {
 	      x = shallow_copy_rtx (x);
@@ -638,7 +657,8 @@ lra_eliminate_regs_1 (rtx_insn *insn, rtx x, machine_mode mem_mode,
 	  for (j = 0; j < XVECLEN (x, i); j++)
 	    {
 	      new_rtx = lra_eliminate_regs_1 (insn, XVECEXP (x, i, j), mem_mode,
-					      subst_p, update_p, full_p);
+					      subst_p, update_p,
+					      update_sp_offset, full_p);
 	      if (new_rtx != XVECEXP (x, i, j) && ! copied_vec)
 		{
 		  rtvec new_v = gen_rtvec_v (XVECLEN (x, i),
@@ -665,7 +685,7 @@ rtx
 lra_eliminate_regs (rtx x, machine_mode mem_mode,
 		    rtx insn ATTRIBUTE_UNUSED)
 {
-  return lra_eliminate_regs_1 (NULL, x, mem_mode, true, false, true);
+  return lra_eliminate_regs_1 (NULL, x, mem_mode, true, false, 0, true);
 }
 
 /* Stack pointer offset before the current insn relative to one at the
@@ -850,13 +870,15 @@ remove_reg_equal_offset_note (rtx insn, rtx what)
 
    If REPLACE_P is false, just update the offsets while keeping the
    base register the same.  If FIRST_P, use the sp offset for
-   elimination to sp.  Attach the note about used elimination for
-   insns setting frame pointer to update elimination easy (without
-   parsing already generated elimination insns to find offset
-   previously used) in future.  */
+   elimination to sp.  Otherwise, use UPDATE_SP_OFFSET for this.
+   Attach the note about used elimination for insns setting frame
+   pointer to update elimination easy (without parsing already
+   generated elimination insns to find offset previously used) in
+   future.  */
 
-static void
-eliminate_regs_in_insn (rtx_insn *insn, bool replace_p, bool first_p)
+void
+eliminate_regs_in_insn (rtx_insn *insn, bool replace_p, bool first_p,
+			HOST_WIDE_INT update_sp_offset)
 {
   int icode = recog_memoized (insn);
   rtx old_set = single_set (insn);
@@ -986,8 +1008,13 @@ eliminate_regs_in_insn (rtx_insn *insn, bool replace_p, bool first_p)
 	  if (! replace_p)
 	    {
 	      offset += (ep->offset - ep->previous_offset);
-	      if (first_p && ep->to_rtx == stack_pointer_rtx)
-		offset -= lra_get_insn_recog_data (insn)->sp_offset;
+	      if (ep->to_rtx == stack_pointer_rtx)
+		{
+		  if (first_p)
+		    offset -= lra_get_insn_recog_data (insn)->sp_offset;
+		  else
+		    offset += update_sp_offset;
+		}
 	      offset = trunc_int_for_mode (offset, GET_MODE (plus_cst_src));
 	    }
 
@@ -1061,7 +1088,7 @@ eliminate_regs_in_insn (rtx_insn *insn, bool replace_p, bool first_p)
 	  substed_operand[i]
 	    = lra_eliminate_regs_1 (insn, *id->operand_loc[i], VOIDmode,
 				    replace_p, ! replace_p && ! first_p,
-				    first_p);
+				    update_sp_offset, first_p);
 	  if (substed_operand[i] != orig_operand[i])
 	    validate_p = true;
 	}
@@ -1349,7 +1376,7 @@ lra_eliminate_reg_if_possible (rtx *loc)
 static void
 process_insn_for_elimination (rtx_insn *insn, bool final_p, bool first_p)
 {
-  eliminate_regs_in_insn (insn, final_p, first_p);
+  eliminate_regs_in_insn (insn, final_p, first_p, 0);
   if (! final_p)
     {
       /* Check that insn changed its code.  This is a case when a move

gcc/lra-int.h

@@ -328,7 +328,6 @@ extern bitmap_head lra_inheritance_pseudos;
 extern bitmap_head lra_split_regs;
 extern bitmap_head lra_subreg_reload_pseudos;
 extern bitmap_head lra_optional_reload_pseudos;
-extern int lra_constraint_new_insn_uid_start;
 
 /* lra-constraints.c: */
 
@@ -339,6 +338,7 @@ extern int lra_constraint_iter;
 extern bool lra_risky_transformations_p;
 extern int lra_inheritance_iter;
 extern int lra_undo_inheritance_iter;
+extern bool lra_constrain_insn (rtx_insn *);
 extern bool lra_constraints (bool);
 extern void lra_constraints_init (void);
 extern void lra_constraints_finish (void);
@@ -389,13 +389,17 @@ extern bool lra_need_for_spills_p (void);
 extern void lra_spill (void);
 extern void lra_final_code_change (void);
 
+/* lra-remat.c:  */
+
+extern bool lra_remat (void);
 
 /* lra-elimination.c: */
 
 extern void lra_debug_elim_table (void);
 extern int lra_get_elimination_hard_regno (int);
-extern rtx lra_eliminate_regs_1 (rtx_insn *, rtx, machine_mode, bool,
-				 bool, bool);
+extern rtx lra_eliminate_regs_1 (rtx_insn *, rtx, machine_mode,
+				 bool, bool, HOST_WIDE_INT, bool);
+extern void eliminate_regs_in_insn (rtx_insn *insn, bool, bool, HOST_WIDE_INT);
 extern void lra_eliminate (bool, bool);
 
 extern void lra_eliminate_reg_if_possible (rtx *);

gcc/lra-lives.c

@@ -252,8 +252,7 @@ make_hard_regno_born (int regno)
   unsigned int i;
 
   lra_assert (regno < FIRST_PSEUDO_REGISTER);
-  if (TEST_HARD_REG_BIT (lra_no_alloc_regs, regno)
-      || TEST_HARD_REG_BIT (hard_regs_live, regno))
+  if (TEST_HARD_REG_BIT (hard_regs_live, regno))
     return;
   SET_HARD_REG_BIT (hard_regs_live, regno);
   sparseset_set_bit (start_living, regno);
@@ -267,8 +266,7 @@ static void
 make_hard_regno_dead (int regno)
 {
   lra_assert (regno < FIRST_PSEUDO_REGISTER);
-  if (TEST_HARD_REG_BIT (lra_no_alloc_regs, regno)
-      || ! TEST_HARD_REG_BIT (hard_regs_live, regno))
+  if (! TEST_HARD_REG_BIT (hard_regs_live, regno))
     return;
   sparseset_set_bit (start_dying, regno);
   CLEAR_HARD_REG_BIT (hard_regs_live, regno);
@@ -662,7 +660,6 @@ process_bb_lives (basic_block bb, int &curr_point)
   sparseset_clear (pseudos_live_through_setjumps);
   REG_SET_TO_HARD_REG_SET (hard_regs_live, reg_live_out);
   AND_COMPL_HARD_REG_SET (hard_regs_live, eliminable_regset);
-  AND_COMPL_HARD_REG_SET (hard_regs_live, lra_no_alloc_regs);
   EXECUTE_IF_SET_IN_BITMAP (reg_live_out, FIRST_PSEUDO_REGISTER, j, bi)
     mark_pseudo_live (j, curr_point);
 

gcc/lra-spills.c

@@ -445,7 +445,7 @@ remove_pseudos (rtx *loc, rtx_insn *insn)
 	{
 	  rtx x = lra_eliminate_regs_1 (insn, pseudo_slots[i].mem,
 					GET_MODE (pseudo_slots[i].mem),
-					false, false, true);
+					0, false, false, true);
 	  *loc = x != pseudo_slots[i].mem ? x : copy_rtx (x);
 	}
       return;

gcc/lra.c

@@ -37,6 +37,7 @@ along with GCC; see the file COPYING3.	If not see
        generated;
      o Some pseudos might be spilled to assign hard registers to
        new reload pseudos;
+     o Recalculating spilled pseudo values (rematerialization);
      o Changing spilled pseudos to stack memory or their equivalences;
      o Allocation stack memory changes the address displacement and
        new iteration is needed.
@@ -57,19 +58,26 @@ along with GCC; see the file COPYING3.	If not see
  -----------      |              ----------------                   |
                   |                      |                          |
                   |                      V         New              |
-         ----------------    No    ------------  pseudos   -------------------
-        | Spilled pseudo | change |Constraints:| or insns | Inheritance/split |
-        |    to memory   |<-------|    RTL     |--------->|  transformations  |
-        |  substitution  |        | transfor-  |          |    in EBB scope   |
-         ----------------         |  mations   |           -------------------
-                |                   ------------
-                V
-    -------------------------
-   | Hard regs substitution, |
-   |  devirtalization, and   |------> Finish
-   | restoring scratches got |
-   |         memory          |
-    -------------------------
+                  |                 ------------  pseudos   -------------------
+                  |                |Constraints:| or insns | Inheritance/split |
+                  |                |    RTL     |--------->|  transformations  |
+                  |                | transfor-  |          |    in EBB scope   |
+                  | substi-        |  mations   |           -------------------
+                  | tutions         ------------
+                  |                     | No change
+          ----------------              V
+         | Spilled pseudo |      -------------------
+         |    to memory   |<----| Rematerialization |
+         |  substitution  |      -------------------
+          ----------------        
+                  | No susbtitions
+                  V                
+      -------------------------
+     | Hard regs substitution, |
+     |  devirtalization, and   |------> Finish
+     | restoring scratches got |
+     |         memory          |
+      -------------------------
 
    To speed up the process:
      o We process only insns affected by changes on previous
@@ -849,38 +857,38 @@ collect_non_operand_hard_regs (rtx *x, lra_insn_recog_data_t data,
     {
       if ((regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER)
 	return list;
+      /* Process all regs even unallocatable ones as we need info
+	 about all regs for rematerialization pass.  */
       for (last = regno + hard_regno_nregs[regno][mode];
 	   regno < last;
 	   regno++)
-	if (! TEST_HARD_REG_BIT (lra_no_alloc_regs, regno)
-	    || TEST_HARD_REG_BIT (eliminable_regset, regno))
-	  {
-	    for (curr = list; curr != NULL; curr = curr->next)
-	      if (curr->regno == regno && curr->subreg_p == subreg_p
-		  && curr->biggest_mode == mode)
-		{
-		  if (curr->type != type)
-		    curr->type = OP_INOUT;
-		  if (curr->early_clobber != early_clobber)
-		    curr->early_clobber = true;
-		  break;
-		}
-	    if (curr == NULL)
+	{
+	  for (curr = list; curr != NULL; curr = curr->next)
+	    if (curr->regno == regno && curr->subreg_p == subreg_p
+		&& curr->biggest_mode == mode)
 	      {
-		/* This is a new hard regno or the info can not be
-		   integrated into the found structure.	 */
+		if (curr->type != type)
+		  curr->type = OP_INOUT;
+		if (curr->early_clobber != early_clobber)
+		  curr->early_clobber = true;
+		break;
+	      }
+	  if (curr == NULL)
+	    {
+	      /* This is a new hard regno or the info can not be
+		 integrated into the found structure.	 */
 #ifdef STACK_REGS
-		early_clobber
-		  = (early_clobber
-		     /* This clobber is to inform popping floating
-			point stack only.  */
-		     && ! (FIRST_STACK_REG <= regno
-			   && regno <= LAST_STACK_REG));
+	      early_clobber
+		= (early_clobber
+		   /* This clobber is to inform popping floating
+		      point stack only.  */
+		   && ! (FIRST_STACK_REG <= regno
+			 && regno <= LAST_STACK_REG));
 #endif
-		list = new_insn_reg (data->insn, regno, type, mode, subreg_p,
-				     early_clobber, list);
-	      }
-	  }
+	      list = new_insn_reg (data->insn, regno, type, mode, subreg_p,
+				   early_clobber, list);
+	    }
+	}
       return list;
     }
   switch (code)
@@ -1456,10 +1464,8 @@ add_regs_to_insn_regno_info (lra_insn_recog_data_t data, rtx x, int uid,
   if (REG_P (x))
     {
       regno = REGNO (x);
-      if (regno < FIRST_PSEUDO_REGISTER
-	  && TEST_HARD_REG_BIT (lra_no_alloc_regs, regno)
-	  && ! TEST_HARD_REG_BIT (eliminable_regset, regno))
-	return;
+      /* Process all regs even unallocatable ones as we need info about
+	 all regs for rematerialization pass.  */
       expand_reg_info ();
       if (bitmap_set_bit (&lra_reg_info[regno].insn_bitmap, uid))
 	{
@@ -2152,9 +2158,6 @@ bitmap_head lra_optional_reload_pseudos;
    pass.  */
 bitmap_head lra_subreg_reload_pseudos;
 
-/* First UID of insns generated before a new spill pass.  */
-int lra_constraint_new_insn_uid_start;
-
 /* File used for output of LRA debug information.  */
 FILE *lra_dump_file;
 
@@ -2252,7 +2255,6 @@ lra (FILE *f)
   lra_curr_reload_num = 0;
   push_insns (get_last_insn (), NULL);
   /* It is needed for the 1st coalescing.  */
-  lra_constraint_new_insn_uid_start = get_max_uid ();
   bitmap_initialize (&lra_inheritance_pseudos, &reg_obstack);
   bitmap_initialize (&lra_split_regs, &reg_obstack);
   bitmap_initialize (&lra_optional_reload_pseudos, &reg_obstack);
@@ -2345,12 +2347,21 @@ lra (FILE *f)
 	  lra_create_live_ranges (lra_reg_spill_p);
 	  live_p = true;
 	}
+      /* Now we know what pseudos should be spilled.  Try to
+	 rematerialize them first.  */
+      if (lra_remat ())
+	{
+	  /* We need full live info -- see the comment above.  */
+	  lra_create_live_ranges (lra_reg_spill_p);
+	  live_p = true;
+	  if (! lra_need_for_spills_p ())
+	    break;
+	}
       lra_spill ();
       /* Assignment of stack slots changes elimination offsets for
 	 some eliminations.  So update the offsets here.  */
       lra_eliminate (false, false);
       lra_constraint_new_regno_start = max_reg_num ();
-      lra_constraint_new_insn_uid_start = get_max_uid ();
       lra_assignment_iter_after_spill = 0;
     }
   restore_scratches ();

gcc/opts.c

@@ -500,6 +500,7 @@ static const struct default_options default_options_table[] =
     { OPT_LEVELS_2_PLUS, OPT_fipa_icf, NULL, 1 },
     { OPT_LEVELS_2_PLUS, OPT_fisolate_erroneous_paths_dereference, NULL, 1 },
     { OPT_LEVELS_2_PLUS, OPT_fuse_caller_save, NULL, 1 },
+    { OPT_LEVELS_2_PLUS, OPT_flra_remat, NULL, 1 },
 
     /* -O3 optimizations.  */
     { OPT_LEVELS_3_PLUS, OPT_ftree_loop_distribute_patterns, NULL, 1 },

gcc/timevar.def

@@ -237,6 +237,7 @@ DEFTIMEVAR (TV_LRA_INHERITANCE	     , "LRA reload inheritance")
 DEFTIMEVAR (TV_LRA_CREATE_LIVE_RANGES, "LRA create live ranges")
 DEFTIMEVAR (TV_LRA_ASSIGN	     , "LRA hard reg assignment")
 DEFTIMEVAR (TV_LRA_COALESCE	     , "LRA coalesce pseudo regs")
+DEFTIMEVAR (TV_LRA_REMAT	     , "LRA rematerialization")
 DEFTIMEVAR (TV_RELOAD		     , "reload")
 DEFTIMEVAR (TV_RELOAD_CSE_REGS       , "reload CSE regs")
 DEFTIMEVAR (TV_GCSE_AFTER_RELOAD     , "load CSE after reload")