gcc/0036-rtl-ifcvt-introduce-rtl-ifcvt-enchancements.patch

From 4cae948c1c00ad7a59f0f234f809fbd9a0208eb4 Mon Sep 17 00:00:00 2001
From: vchernon <chernonog.vyacheslav@huawei.com>
Date: Wed, 28 Feb 2024 23:05:12 +0800
Subject: [PATCH 02/18] [rtl-ifcvt] introduce rtl ifcvt enchancements     new
 option:       -fifcvt-allow-complicated-cmps:         allows ifcvt to deal
 with complicated cmps like

        cmp reg1 (reg2 + reg3)

        can increase compilation time
    new param:
      -param=ifcvt-allow-register-renaming=[0,1,2]
        1 : allows ifcvt to rename registers in then and else bb
        2 : allows to rename registers in condition and else/then bb
        can increase compilation time and register pressure
---
 gcc/common.opt                                |   4 +
 gcc/ifcvt.cc                                  | 291 +++++++++++++++---
 gcc/params.opt                                |   4 +
 .../gcc.c-torture/execute/ifcvt-renaming-1.c  |  35 +++
 gcc/testsuite/gcc.dg/ifcvt-6.c                |  27 ++
 5 files changed, 311 insertions(+), 50 deletions(-)
 create mode 100644 gcc/testsuite/gcc.c-torture/execute/ifcvt-renaming-1.c
 create mode 100644 gcc/testsuite/gcc.dg/ifcvt-6.c

diff --git a/gcc/common.opt b/gcc/common.opt
index c7c6bc256..aa00fb7b0 100644
--- a/gcc/common.opt
+++ b/gcc/common.opt
@@ -3691,4 +3691,8 @@ fipa-ra
 Common Var(flag_ipa_ra) Optimization
 Use caller save register across calls if possible.
 
+fifcvt-allow-complicated-cmps
+Common Var(flag_ifcvt_allow_complicated_cmps) Optimization
+Allow RTL if-conversion pass to deal with complicated cmps (can increase compilation time).
+
 ; This comment is to ensure we retain the blank line above.
diff --git a/gcc/ifcvt.cc b/gcc/ifcvt.cc
index 2c1eba312..584db7b55 100644
--- a/gcc/ifcvt.cc
+++ b/gcc/ifcvt.cc
@@ -886,7 +886,9 @@ noce_emit_store_flag (struct noce_if_info *if_info, rtx x, int reversep,
     }
 
   /* Don't even try if the comparison operands or the mode of X are weird.  */
-  if (cond_complex || !SCALAR_INT_MODE_P (GET_MODE (x)))
+  if (!flag_ifcvt_allow_complicated_cmps
+      && (cond_complex
+	  || !SCALAR_INT_MODE_P (GET_MODE (x))))
     return NULL_RTX;
 
   return emit_store_flag (x, code, XEXP (cond, 0),
@@ -1965,7 +1967,8 @@ insn_valid_noce_process_p (rtx_insn *insn, rtx cc)
   /* Currently support only simple single sets in test_bb.  */
   if (!sset
       || !noce_operand_ok (SET_DEST (sset))
-      || contains_ccmode_rtx_p (SET_DEST (sset))
+      || (!flag_ifcvt_allow_complicated_cmps
+	  && contains_ccmode_rtx_p (SET_DEST (sset)))
       || !noce_operand_ok (SET_SRC (sset)))
     return false;
 
@@ -1979,13 +1982,17 @@ insn_valid_noce_process_p (rtx_insn *insn, rtx cc)
    in this function.  */
 
 static bool
-bbs_ok_for_cmove_arith (basic_block bb_a, basic_block bb_b, rtx to_rename)
+bbs_ok_for_cmove_arith (basic_block bb_a,
+			basic_block bb_b,
+			rtx to_rename,
+			bitmap conflict_regs)
 {
   rtx_insn *a_insn;
   bitmap bba_sets = BITMAP_ALLOC (&reg_obstack);
-
+  bitmap intersections = BITMAP_ALLOC (&reg_obstack);
   df_ref def;
   df_ref use;
+  rtx_insn *last_a = last_active_insn (bb_a, FALSE);
 
   FOR_BB_INSNS (bb_a, a_insn)
     {
@@ -1995,18 +2002,15 @@ bbs_ok_for_cmove_arith (basic_block bb_a, basic_block bb_b, rtx to_rename)
       rtx sset_a = single_set (a_insn);
 
       if (!sset_a)
-	{
-	  BITMAP_FREE (bba_sets);
-	  return false;
-	}
+	goto end_cmove_arith_check_and_fail;
       /* Record all registers that BB_A sets.  */
       FOR_EACH_INSN_DEF (def, a_insn)
-	if (!(to_rename && DF_REF_REG (def) == to_rename))
+	if (!(to_rename && DF_REF_REG (def) == to_rename && a_insn == last_a))
 	  bitmap_set_bit (bba_sets, DF_REF_REGNO (def));
     }
 
+  bitmap_and (intersections, df_get_live_in (bb_b), bba_sets);
   rtx_insn *b_insn;
-
   FOR_BB_INSNS (bb_b, b_insn)
     {
       if (!active_insn_p (b_insn))
@@ -2015,10 +2019,7 @@ bbs_ok_for_cmove_arith (basic_block bb_a, basic_block bb_b, rtx to_rename)
       rtx sset_b = single_set (b_insn);
 
       if (!sset_b)
-	{
-	  BITMAP_FREE (bba_sets);
-	  return false;
-	}
+	goto end_cmove_arith_check_and_fail;
 
       /* Make sure this is a REG and not some instance
 	 of ZERO_EXTRACT or SUBREG or other dangerous stuff.
@@ -2030,25 +2031,34 @@ bbs_ok_for_cmove_arith (basic_block bb_a, basic_block bb_b, rtx to_rename)
       if (MEM_P (SET_DEST (sset_b)))
 	gcc_assert (rtx_equal_p (SET_DEST (sset_b), to_rename));
       else if (!REG_P (SET_DEST (sset_b)))
-	{
-	  BITMAP_FREE (bba_sets);
-	  return false;
-	}
+	goto end_cmove_arith_check_and_fail;
 
-      /* If the insn uses a reg set in BB_A return false.  */
+      /* If the insn uses a reg set in BB_A return false
+	 or try to collect register list for renaming.  */
       FOR_EACH_INSN_USE (use, b_insn)
 	{
-	  if (bitmap_bit_p (bba_sets, DF_REF_REGNO (use)))
+	  if (bitmap_bit_p (intersections, DF_REF_REGNO (use)))
 	    {
-	      BITMAP_FREE (bba_sets);
-	      return false;
+	      if (param_ifcvt_allow_register_renaming < 1)
+		  goto end_cmove_arith_check_and_fail;
+
+	      /* Those regs should be renamed.  We can't rename CC reg, but
+		 possibly we can provide combined comparison in the future.  */
+	      if (GET_MODE_CLASS (GET_MODE (DF_REF_REG (use))) == MODE_CC)
+		goto end_cmove_arith_check_and_fail;
+	      bitmap_set_bit (conflict_regs, DF_REF_REGNO (use));
 	    }
 	}
-
     }
 
   BITMAP_FREE (bba_sets);
+  BITMAP_FREE (intersections);
   return true;
+
+end_cmove_arith_check_and_fail:
+  BITMAP_FREE (bba_sets);
+  BITMAP_FREE (intersections);
+  return false;
 }
 
 /* Emit copies of all the active instructions in BB except the last.
@@ -2103,6 +2113,142 @@ noce_emit_bb (rtx last_insn, basic_block bb, bool simple)
   return true;
 }
 
+/* This function tries to rename regs that intersect with considered bb
+   inside condition expression.  Condition expression will be moved down
+   if the optimization will be applied, so it is essential to be sure that
+   all intersected registers will be renamed otherwise transformation
+   can't be applied.  Function returns true if renaming was successful
+   and optimization can proceed futher.  */
+
+static bool
+noce_rename_regs_in_cond (struct noce_if_info *if_info, bitmap cond_rename_regs)
+{
+  bool success = true;
+  if (bitmap_empty_p (cond_rename_regs))
+    return true;
+  if (param_ifcvt_allow_register_renaming < 2)
+    return false;
+  df_ref use;
+  rtx_insn *cmp_insn = if_info->cond_earliest;
+  /*  Jump instruction as a condion currently unsupported.  */
+  if (JUMP_P (cmp_insn))
+    return false;
+  rtx_insn *before_cmp = PREV_INSN (cmp_insn);
+  start_sequence ();
+  rtx_insn *copy_of_cmp = as_a <rtx_insn *> (copy_rtx (cmp_insn));
+  basic_block cmp_block = BLOCK_FOR_INSN (cmp_insn);
+  FOR_EACH_INSN_USE (use, cmp_insn)
+    {
+      if (bitmap_bit_p (cond_rename_regs, DF_REF_REGNO (use)))
+	{
+	  rtx use_reg = DF_REF_REG (use);
+	  rtx tmp = gen_reg_rtx (GET_MODE (use_reg));
+	  if (!validate_replace_rtx (use_reg, tmp, copy_of_cmp))
+	    {
+	      end_sequence ();
+	      return false;
+	    }
+	  noce_emit_move_insn (tmp, use_reg);
+	}
+    }
+
+  emit_insn (PATTERN (copy_of_cmp));
+  rtx_insn *seq = get_insns ();
+  unshare_all_rtl_in_chain (seq);
+  end_sequence ();
+
+  emit_insn_after_setloc (seq, before_cmp, INSN_LOCATION (cmp_insn));
+  delete_insn_and_edges (cmp_insn);
+  rtx_insn *insn;
+  FOR_BB_INSNS (cmp_block, insn)
+    df_insn_rescan (insn);
+
+  if_info->cond = noce_get_condition (if_info->jump,
+				      &copy_of_cmp,
+				      if_info->then_else_reversed);
+  if_info->cond_earliest = copy_of_cmp;
+  if_info->rev_cond = NULL_RTX;
+
+  return success;
+}
+
+/* This function tries to rename regs that intersect with considered bb.
+   return true if the renaming was successful and optimization can
+   proceed futher, false otherwise.  */
+static bool
+noce_rename_regs_in_bb (basic_block test_bb, bitmap rename_regs)
+{
+  if (bitmap_empty_p (rename_regs))
+    return true;
+  rtx_insn *insn;
+  rtx_insn *last_insn = last_active_insn (test_bb, FALSE);
+  bool res = true;
+  start_sequence ();
+  FOR_BB_INSNS (test_bb, insn)
+    {
+      if (!active_insn_p (insn))
+	continue;
+      /* Only ssets are supported for now.  */
+      rtx sset = single_set (insn);
+      gcc_assert (sset);
+      rtx x = SET_DEST (sset);
+      if (!REG_P (x) || !bitmap_bit_p (rename_regs, REGNO (x)))
+	continue;
+      /* Do not need to rename dest in the last instruction
+	 it will be renamed anyway.  */
+      if (insn == last_insn)
+	continue;
+      machine_mode mode = GET_MODE (x);
+      rtx tmp = gen_reg_rtx (mode);
+      if (!validate_replace_rtx_part (x, tmp, &SET_DEST (sset), insn))
+	{
+	  gcc_assert (insn != last_insn);
+	  /* We can generate additional move for such case,
+	     but it will increase register preasure.
+	     For now just stop transformation.  */
+	  rtx result_rtx = SET_DEST (single_set (last_insn));
+	  if (REG_P (result_rtx) && (x != result_rtx))
+	    {
+	      res = false;
+	      break;
+	    }
+	  if (!validate_replace_rtx (x, tmp, insn))
+	    gcc_unreachable ();
+	  noce_emit_move_insn (tmp,x);
+	}
+      set_used_flags (insn);
+      rtx_insn *rename_candidate;
+      for (rename_candidate = NEXT_INSN (insn);
+	   rename_candidate && rename_candidate!= NEXT_INSN (BB_END (test_bb));
+	   rename_candidate = NEXT_INSN (rename_candidate))
+	{
+	  if (!reg_overlap_mentioned_p (x, rename_candidate))
+	    continue;
+
+	  int replace_res = TRUE;
+	  if (rename_candidate == last_insn)
+	    {
+	      validate_replace_src_group (x, tmp, rename_candidate);
+	      replace_res = apply_change_group ();
+	    }
+	  else
+	    replace_res = validate_replace_rtx (x, tmp, rename_candidate);
+	  gcc_assert (replace_res);
+	  set_used_flags (rename_candidate);
+	}
+      set_used_flags (x);
+      set_used_flags (tmp);
+    }
+    rtx_insn *seq = get_insns ();
+    unshare_all_rtl_in_chain (seq);
+    end_sequence ();
+    emit_insn_before_setloc (seq, first_active_insn (test_bb),
+			     INSN_LOCATION (first_active_insn (test_bb)));
+    FOR_BB_INSNS (test_bb, insn)
+      df_insn_rescan (insn);
+  return res;
+}
+
 /* Try more complex cases involving conditional_move.  */
 
 static int
@@ -2185,11 +2331,30 @@ noce_try_cmove_arith (struct noce_if_info *if_info)
 	  std::swap (then_bb, else_bb);
 	}
     }
-
+  bitmap else_bb_rename_regs = BITMAP_ALLOC (&reg_obstack);
+  bitmap then_bb_rename_regs = BITMAP_ALLOC (&reg_obstack);
   if (then_bb && else_bb
-      && (!bbs_ok_for_cmove_arith (then_bb, else_bb,  if_info->orig_x)
-	  || !bbs_ok_for_cmove_arith (else_bb, then_bb,  if_info->orig_x)))
-    return FALSE;
+      && (!bbs_ok_for_cmove_arith (then_bb, else_bb,
+				   if_info->orig_x,
+				   then_bb_rename_regs)
+	  || !bbs_ok_for_cmove_arith (else_bb, then_bb,
+				      if_info->orig_x,
+				      else_bb_rename_regs)))
+    {
+      BITMAP_FREE (then_bb_rename_regs);
+      BITMAP_FREE (else_bb_rename_regs);
+      return FALSE;
+    }
+  bool prepass_renaming = noce_rename_regs_in_bb (then_bb,
+						  then_bb_rename_regs)
+			  && noce_rename_regs_in_bb (else_bb,
+						     else_bb_rename_regs);
+
+  BITMAP_FREE (then_bb_rename_regs);
+  BITMAP_FREE (else_bb_rename_regs);
+
+  if (!prepass_renaming)
+   return FALSE;
 
   start_sequence ();
 
@@ -3072,7 +3237,8 @@ noce_operand_ok (const_rtx op)
 
 static bool
 bb_valid_for_noce_process_p (basic_block test_bb, rtx cond,
-			      unsigned int *cost, bool *simple_p)
+			     unsigned int *cost, bool *simple_p,
+			     bitmap cond_rename_regs)
 {
   if (!test_bb)
     return false;
@@ -3112,8 +3278,9 @@ bb_valid_for_noce_process_p (basic_block test_bb, rtx cond,
   rtx_insn *prev_last_insn = PREV_INSN (last_insn);
   gcc_assert (prev_last_insn);
 
-  /* For now, disallow setting x multiple times in test_bb.  */
-  if (REG_P (x) && reg_set_between_p (x, first_insn, prev_last_insn))
+  if (REG_P (x)
+      && reg_set_between_p (x, first_insn, prev_last_insn)
+      && param_ifcvt_allow_register_renaming < 1)
     return false;
 
   bitmap test_bb_temps = BITMAP_ALLOC (&reg_obstack);
@@ -3125,25 +3292,35 @@ bb_valid_for_noce_process_p (basic_block test_bb, rtx cond,
   rtx_insn *insn;
   FOR_BB_INSNS (test_bb, insn)
     {
-      if (insn != last_insn)
-	{
-	  if (!active_insn_p (insn))
-	    continue;
+      if (insn == last_insn)
+	continue;
+      if (!active_insn_p (insn))
+	continue;
 
-	  if (!insn_valid_noce_process_p (insn, cc))
-	    goto free_bitmap_and_fail;
+      if (!insn_valid_noce_process_p (insn, cc))
+	goto free_bitmap_and_fail;
 
-	  rtx sset = single_set (insn);
-	  gcc_assert (sset);
+      rtx sset = single_set (insn);
+      gcc_assert (sset);
 
-	  if (contains_mem_rtx_p (SET_SRC (sset))
-	      || !REG_P (SET_DEST (sset))
-	      || reg_overlap_mentioned_p (SET_DEST (sset), cond))
-	    goto free_bitmap_and_fail;
+      if (contains_mem_rtx_p (SET_SRC (sset))
+	  || !REG_P (SET_DEST (sset)))
+	goto free_bitmap_and_fail;
 
-	  potential_cost += pattern_cost (sset, speed_p);
-	  bitmap_set_bit (test_bb_temps, REGNO (SET_DEST (sset)));
+      if (reg_overlap_mentioned_p (SET_DEST (sset), cond))
+	{
+	  if (param_ifcvt_allow_register_renaming < 1)
+	    goto free_bitmap_and_fail;
+	  rtx sset_dest = SET_DEST (sset);
+	  if (REG_P (sset_dest)
+	      && (GET_MODE_CLASS (GET_MODE (sset_dest)) != MODE_CC))
+	    bitmap_set_bit (cond_rename_regs, REGNO (sset_dest));
+	  else
+	    goto free_bitmap_and_fail;
 	}
+	potential_cost += pattern_cost (sset, speed_p);
+	if (SET_DEST (sset) != SET_DEST (last_set))
+	  bitmap_set_bit (test_bb_temps, REGNO (SET_DEST (sset)));
     }
 
   /* If any of the intermediate results in test_bb are live after test_bb
@@ -3777,15 +3954,29 @@ noce_process_if_block (struct noce_if_info *if_info)
 
   bool speed_p = optimize_bb_for_speed_p (test_bb);
   unsigned int then_cost = 0, else_cost = 0;
+  bitmap cond_rename_regs = BITMAP_ALLOC (&reg_obstack);
   if (!bb_valid_for_noce_process_p (then_bb, cond, &then_cost,
-				    &if_info->then_simple))
-    return false;
+				    &if_info->then_simple, cond_rename_regs))
+    {
+      BITMAP_FREE (cond_rename_regs);
+      return false;
+    }
 
   if (else_bb
       && !bb_valid_for_noce_process_p (else_bb, cond, &else_cost,
-				       &if_info->else_simple))
-    return false;
+				       &if_info->else_simple, cond_rename_regs))
+    {
+      BITMAP_FREE (cond_rename_regs);
+      return false;
+    }
 
+  if (!noce_rename_regs_in_cond (if_info, cond_rename_regs))
+    {
+      BITMAP_FREE (cond_rename_regs);
+      return false;
+    }
+  BITMAP_FREE (cond_rename_regs);
+  cond = if_info->cond;
   if (speed_p)
     if_info->original_cost += average_cost (then_cost, else_cost,
 					    find_edge (test_bb, then_bb));
@@ -5823,12 +6014,13 @@ if_convert (bool after_combine)
 {
   basic_block bb;
   int pass;
-
   if (optimize == 1)
     {
       df_live_add_problem ();
       df_live_set_all_dirty ();
     }
+  free_dominance_info (CDI_DOMINATORS);
+  cleanup_cfg (CLEANUP_EXPENSIVE);
 
   /* Record whether we are after combine pass.  */
   ifcvt_after_combine = after_combine;
@@ -5933,7 +6125,6 @@ rest_of_handle_if_conversion (void)
 	  dump_reg_info (dump_file);
 	  dump_flow_info (dump_file, dump_flags);
 	}
-      cleanup_cfg (CLEANUP_EXPENSIVE);
       if_convert (false);
       if (num_updated_if_blocks)
 	/* Get rid of any dead CC-related instructions.  */
diff --git a/gcc/params.opt b/gcc/params.opt
index d2196dc68..ba87f820b 100644
--- a/gcc/params.opt
+++ b/gcc/params.opt
@@ -669,6 +669,10 @@ Maximum permissible cost for the sequence that would be generated by the RTL if-
 Common Joined UInteger Var(param_max_rtl_if_conversion_unpredictable_cost) Init(40) IntegerRange(0, 200) Param Optimization
 Maximum permissible cost for the sequence that would be generated by the RTL if-conversion pass for a branch that is considered unpredictable.
 
+-param=ifcvt-allow-register-renaming=
+Common Joined UInteger Var(param_ifcvt_allow_register_renaming) IntegerRange(0, 2) Param Optimization
+Allow RTL if-conversion pass to aggressively rename registers in basic blocks.  Sometimes additional moves will be created.
+
 -param=max-sched-extend-regions-iters=
 Common Joined UInteger Var(param_max_sched_extend_regions_iters) Param Optimization
 The maximum number of iterations through CFG to extend regions.
diff --git a/gcc/testsuite/gcc.c-torture/execute/ifcvt-renaming-1.c b/gcc/testsuite/gcc.c-torture/execute/ifcvt-renaming-1.c
new file mode 100644
index 000000000..65c4d4140
--- /dev/null
+++ b/gcc/testsuite/gcc.c-torture/execute/ifcvt-renaming-1.c
@@ -0,0 +1,35 @@
+
+extern void abort(void);
+
+__attribute__ ((noinline))
+int foo (int x, int y, int z, int a, int b)
+{
+  if (a < 2) {
+      if (a == 0) {
+	  if (x - y < 0)
+	    x = x - y + z;
+	  else
+	    x = x - y;
+	}
+      else {
+	  if (x + y >= z)
+	    x = x + y - z;
+	  else
+	    x = x + y;
+	}
+    }
+  return x;
+}
+
+int main(void) {
+  if (foo (5,10,7,0,1) != 2) // x - y + z = -5 + 7 = 2
+    abort ();
+  if (foo (50,10,7,0,1) != 40) // x - y = 40
+    abort ();
+  if (foo (5,10,7,1,1) != 8) // x + y - z = 5 + 10 - 7 = 8
+    abort ();
+  if (foo (5,10,70,1,1) != 15) // x + y = 15
+    abort ();
+  return 0;
+}
+
diff --git a/gcc/testsuite/gcc.dg/ifcvt-6.c b/gcc/testsuite/gcc.dg/ifcvt-6.c
new file mode 100644
index 000000000..be9a67b3f
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/ifcvt-6.c
@@ -0,0 +1,27 @@
+/* { dg-do compile { target { aarch64*-*-* } } } */
+/* { dg-options "-fdump-rtl-ce1 -O2 --param max-rtl-if-conversion-unpredictable-cost=100 --param max-rtl-if-conversion-predictable-cost=100 --param=ifcvt-allow-register-renaming=2 -fifcvt-allow-complicated-cmps" } */
+
+typedef unsigned int uint16_t;
+
+uint16_t
+foo (uint16_t x, uint16_t y, uint16_t z, uint16_t a,
+     uint16_t b, uint16_t c, uint16_t d) {
+  int i = 1;
+  int j = 1;
+  if (a > b) {
+      j = x;
+      if (b > c)
+	i = y;
+      else
+	i = z;
+    }
+  else {
+      j = y;
+      if (c > d)
+	i = z;
+    }
+  return i * j;
+}
+
+/* { dg-final { scan-rtl-dump "7 true changes made" "ce1" } } */
+
-- 
2.33.0