re PR target/79131 (ICE: in extract_constrain_insn, at recog.c:2213, big-endian ARM)

2017-01-26  Vladimir Makarov  <vmakarov@redhat.com>

	PR target/79131
	* lra-assigns.c (setup_live_pseudos_and_spill_after_risky): Take
	endianess for subregs into account.
	* lra-constraints.c (lra_constraints): Do risky transformations
	always on the first iteration.
	* lra-lives.c (check_pseudos_live_through_calls): Add arg
	last_call_used_reg_set.
	(process_bb_lives): Define and use last_call_used_reg_set.
	* lra.c (lra): Always continue after lra_constraints on the first
	iteration.

2017-01-26  Vladimir Makarov  <vmakarov@redhat.com>

	PR target/79131
	* gcc.target/arm/pr79131.c: New.

From-SVN: r244942

7 files changed, 85 insertions, 21 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 903e966..d09061e 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,16 @@
+2017-01-26  Vladimir Makarov  <vmakarov@redhat.com>
+
+	PR target/79131
+	* lra-assigns.c (setup_live_pseudos_and_spill_after_risky): Take
+	endianess for subregs into account.
+	* lra-constraints.c (lra_constraints): Do risky transformations
+	always on the first iteration.
+	* lra-lives.c (check_pseudos_live_through_calls): Add arg
+	last_call_used_reg_set.
+	(process_bb_lives): Define and use last_call_used_reg_set.
+	* lra.c (lra): Always continue after lra_constraints on the first
+	iteration.
+
 2017-01-26  Kirill Yukhin  <kirill.yukhin@gmail.com>
 
 	* gcc.target/i386/avx512bw-kshiftlq-2.c: Use unsigned long long
diff --git a/gcc/lra-assigns.c b/gcc/lra-assigns.c
index 8e21a49..74a3195 100644
--- a/gcc/lra-assigns.c
+++ b/gcc/lra-assigns.c
@@ -1179,9 +1179,19 @@ setup_live_pseudos_and_spill_after_risky_transforms (bitmap
 	    /* If it is multi-register pseudos they should start on
 	       the same hard register.	*/
 	    || hard_regno != reg_renumber[conflict_regno])
-	  add_to_hard_reg_set (&conflict_set,
-			       lra_reg_info[conflict_regno].biggest_mode,
-			       reg_renumber[conflict_regno]);
+	  {
+	    int conflict_hard_regno = reg_renumber[conflict_regno];
+	    machine_mode biggest_mode = lra_reg_info[conflict_regno].biggest_mode;
+	    int biggest_nregs = hard_regno_nregs[conflict_hard_regno][biggest_mode];
+	    int nregs_diff = (biggest_nregs
+			      - (hard_regno_nregs
+				 [conflict_hard_regno]
+				 [PSEUDO_REGNO_MODE (conflict_regno)]));
+	    add_to_hard_reg_set (&conflict_set,
+				 biggest_mode,
+				 conflict_hard_regno
+				 - (WORDS_BIG_ENDIAN ? nregs_diff : 0));
+	  }
       if (! overlaps_hard_reg_set_p (conflict_set, mode, hard_regno))
 	{
 	  update_lives (regno, false);
diff --git a/gcc/lra-constraints.c b/gcc/lra-constraints.c
index 7b0d2f4..2586913 100644
--- a/gcc/lra-constraints.c
+++ b/gcc/lra-constraints.c
@@ -4511,7 +4511,10 @@ lra_constraints (bool first_p)
       && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
     lra_risky_transformations_p = true;
   else
-    lra_risky_transformations_p = false;
+    /* On the first iteration we should check IRA assignment
+       correctness.  In rare cases, the assignments can be wrong as
+       early clobbers operands are ignored in IRA.  */
+    lra_risky_transformations_p = first_p;
   new_insn_uid_start = get_max_uid ();
   new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
   /* Mark used hard regs for target stack size calulations.  */
diff --git a/gcc/lra-lives.c b/gcc/lra-lives.c
index 010d4e6..f019438 100644
--- a/gcc/lra-lives.c
+++ b/gcc/lra-lives.c
@@ -560,10 +560,11 @@ lra_setup_reload_pseudo_preferenced_hard_reg (int regno,
 }
 
 /* Check that REGNO living through calls and setjumps, set up conflict
-   regs, and clear corresponding bits in PSEUDOS_LIVE_THROUGH_CALLS and
-   PSEUDOS_LIVE_THROUGH_SETJUMPS.  */
+   regs using LAST_CALL_USED_REG_SET, and clear corresponding bits in
+   PSEUDOS_LIVE_THROUGH_CALLS and PSEUDOS_LIVE_THROUGH_SETJUMPS.  */
 static inline void
-check_pseudos_live_through_calls (int regno)
+check_pseudos_live_through_calls (int regno,
+				  HARD_REG_SET last_call_used_reg_set)
 {
   int hr;
 
@@ -571,7 +572,7 @@ check_pseudos_live_through_calls (int regno)
     return;
   sparseset_clear_bit (pseudos_live_through_calls, regno);
   IOR_HARD_REG_SET (lra_reg_info[regno].conflict_hard_regs,
-		    call_used_reg_set);
+		    last_call_used_reg_set);
 
   for (hr = 0; hr < FIRST_PSEUDO_REGISTER; hr++)
     if (HARD_REGNO_CALL_PART_CLOBBERED (hr, PSEUDO_REGNO_MODE (regno)))
@@ -604,11 +605,13 @@ process_bb_lives (basic_block bb, int &curr_point, bool dead_insn_p)
   rtx_insn *next;
   rtx link, *link_loc;
   bool need_curr_point_incr;
-
+  HARD_REG_SET last_call_used_reg_set;
+  
   reg_live_out = df_get_live_out (bb);
   sparseset_clear (pseudos_live);
   sparseset_clear (pseudos_live_through_calls);
   sparseset_clear (pseudos_live_through_setjumps);
+  CLEAR_HARD_REG_SET (last_call_used_reg_set);
   REG_SET_TO_HARD_REG_SET (hard_regs_live, reg_live_out);
   AND_COMPL_HARD_REG_SET (hard_regs_live, eliminable_regset);
   EXECUTE_IF_SET_IN_BITMAP (reg_live_out, FIRST_PSEUDO_REGISTER, j, bi)
@@ -795,7 +798,8 @@ process_bb_lives (basic_block bb, int &curr_point, bool dead_insn_p)
 	    need_curr_point_incr
 	      |= mark_regno_live (reg->regno, reg->biggest_mode,
 				  curr_point);
-	    check_pseudos_live_through_calls (reg->regno);
+	    check_pseudos_live_through_calls (reg->regno,
+					      last_call_used_reg_set);
 	  }
 
       for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
@@ -831,15 +835,27 @@ process_bb_lives (basic_block bb, int &curr_point, bool dead_insn_p)
 
       if (call_p)
 	{
-	  if (flag_ipa_ra)
+	  if (! flag_ipa_ra)
+	    COPY_HARD_REG_SET(last_call_used_reg_set, call_used_reg_set);
+	  else
 	    {
 	      HARD_REG_SET this_call_used_reg_set;
 	      get_call_reg_set_usage (curr_insn, &this_call_used_reg_set,
 				      call_used_reg_set);
 
+	      bool flush = (! hard_reg_set_empty_p (last_call_used_reg_set)
+			    && ! hard_reg_set_equal_p (last_call_used_reg_set,
+						       this_call_used_reg_set));
+
 	      EXECUTE_IF_SET_IN_SPARSESET (pseudos_live, j)
-		IOR_HARD_REG_SET (lra_reg_info[j].actual_call_used_reg_set,
-				  this_call_used_reg_set);
+		{
+		  IOR_HARD_REG_SET (lra_reg_info[j].actual_call_used_reg_set,
+				    this_call_used_reg_set);
+		  if (flush)
+		    check_pseudos_live_through_calls
+		      (j, last_call_used_reg_set);
+		}
+	      COPY_HARD_REG_SET(last_call_used_reg_set, this_call_used_reg_set);
 	    }
 
 	  sparseset_ior (pseudos_live_through_calls,
@@ -866,7 +882,8 @@ process_bb_lives (basic_block bb, int &curr_point, bool dead_insn_p)
 	    need_curr_point_incr
 	      |= mark_regno_live (reg->regno, reg->biggest_mode,
 				  curr_point);
-	    check_pseudos_live_through_calls (reg->regno);
+	    check_pseudos_live_through_calls (reg->regno,
+					      last_call_used_reg_set);
 	  }
 
       for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
@@ -1009,7 +1026,7 @@ process_bb_lives (basic_block bb, int &curr_point, bool dead_insn_p)
       if (sparseset_cardinality (pseudos_live_through_calls) == 0)
 	break;
       if (sparseset_bit_p (pseudos_live_through_calls, j))
-	check_pseudos_live_through_calls (j);
+	check_pseudos_live_through_calls (j, last_call_used_reg_set);
     }
 
   if (need_curr_point_incr)
diff --git a/gcc/lra.c b/gcc/lra.c
index da3b955..9be0053 100644
--- a/gcc/lra.c
+++ b/gcc/lra.c
@@ -2295,7 +2295,7 @@ void
 lra (FILE *f)
 {
   int i;
-  bool live_p, scratch_p, inserted_p;
+  bool live_p, inserted_p;
 
   lra_dump_file = f;
 
@@ -2332,7 +2332,6 @@ lra (FILE *f)
   lra_constraint_new_regno_start = lra_new_regno_start = max_reg_num ();
   lra_bad_spill_regno_start = INT_MAX;
   remove_scratches ();
-  scratch_p = lra_constraint_new_regno_start != max_reg_num ();
 
   /* A function that has a non-local label that can reach the exit
      block via non-exceptional paths must save all call-saved
@@ -2372,11 +2371,12 @@ lra (FILE *f)
       for (;;)
 	{
 	  /* We should try to assign hard registers to scratches even
-	     if there were no RTL transformations in
-	     lra_constraints.  */
+	     if there were no RTL transformations in lra_constraints.
+	     Also we should check IRA assignments on the first
+	     iteration as they can be wrong because of early clobbers
+	     operands which are ignored in IRA.  */
 	  if (! lra_constraints (lra_constraint_iter == 0)
-	      && (lra_constraint_iter > 1
-		  || (! scratch_p && ! caller_save_needed)))
+	      && lra_constraint_iter > 1)
 	    break;
 	  /* Constraint transformations may result in that eliminable
 	     hard regs become uneliminable and pseudos which use them
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index bd07034..8796bb5 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,8 @@
+2017-01-26  Vladimir Makarov  <vmakarov@redhat.com>
+
+	PR target/79131
+	* gcc.target/arm/pr79131.c: New.
+
 2017-01-26  Bin Cheng  <bin.cheng@arm.com>
 
 	* gcc.target/aarch64/ldp_vec_64_1.c: Xfail.
diff --git a/gcc/testsuite/gcc.target/arm/pr79131.c b/gcc/testsuite/gcc.target/arm/pr79131.c
new file mode 100644
index 0000000..3bd9e0f
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/pr79131.c
@@ -0,0 +1,16 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -mbig-endian" } */
+
+long long a;
+enum { NILFS_SEGMENT_USAGE_ACTIVE, NILFS_SEGMENT_USAGE_DIRTY } b;
+void nilfs_sufile_mod_counter(long long p1) {
+  long c = p1;
+  unsigned d = __builtin_bswap64(a);
+  a = __builtin_bswap64(d + c);
+}
+void nilfs_sufile_do_free() {
+  int e, f;
+  e = __builtin_bswap32(b) & 1 << NILFS_SEGMENT_USAGE_DIRTY;
+  f = e;
+  nilfs_sufile_mod_counter(f ? -1 : 0);
+}