diff mbox

patch to fix PR64617

Message ID 54CBC3E7.9030907@redhat.com
State New
Headers show

Commit Message

Vladimir Makarov Jan. 30, 2015, 5:48 p.m. UTC
The following patch fixes

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64617

The patch was bootstrapped and tested on x86-64 and ppc64.

Committed as rev.220294.

2015-01-30  Vladimir Makarov  <vmakarov@redhat.com>

         PR target/64617
         * lra-constraints.c (prohibited_class_reg_set_mode_p): New
     function.
         (process_alt_operands): Use it.
         (curr_insn_transform): Check the optional reload pseudo class is
         ok for the mode.

2015-01-30  Vladimir Makarov  <vmakarov@redhat.com>

         PR target/64617
         * gcc.target/i386/pr64617.c: New test.
diff mbox

Patch

Index: lra-constraints.c
===================================================================
--- lra-constraints.c	(revision 220126)
+++ lra-constraints.c	(working copy)
@@ -1637,6 +1637,22 @@  reg_in_class_p (rtx reg, enum reg_class
   return in_class_p (reg, cl, NULL);
 }
 
+/* Return true if SET of RCLASS contains no hard regs which can be
+   used in MODE.  */
+static bool
+prohibited_class_reg_set_mode_p (enum reg_class rclass,
+				 HARD_REG_SET &set,
+				 enum machine_mode mode)
+{
+  HARD_REG_SET temp;
+  
+  lra_assert (hard_reg_set_subset_p (set, reg_class_contents[rclass]));
+  COPY_HARD_REG_SET (temp, set);
+  AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
+  return (hard_reg_set_subset_p
+	  (temp, ira_prohibited_class_mode_regs[rclass][mode]));
+}
+
 /* Major function to choose the current insn alternative and what
    operands should be reloaded and how.	 If ONLY_ALTERNATIVE is not
    negative we should consider only this alternative.  Return false if
@@ -2311,28 +2327,20 @@  process_alt_operands (int only_alternati
 		     not hold the mode value.  */
 		  && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
 					   [this_alternative][0],
-					   GET_MODE (*curr_id->operand_loc[nop])))
-		{
-		  HARD_REG_SET temp;
-		  
-		  COPY_HARD_REG_SET (temp, this_alternative_set);
-		  AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
+					   GET_MODE (*curr_id->operand_loc[nop]))
 		  /* The above condition is not enough as the first
 		     reg in ira_class_hard_regs can be not aligned for
 		     multi-words mode values.  */
-		  if (hard_reg_set_subset_p (temp,
-					     ira_prohibited_class_mode_regs
-					     [this_alternative]
-					     [GET_MODE (*curr_id->operand_loc[nop])]))
-		    {
-		      if (lra_dump_file != NULL)
-			fprintf
-			  (lra_dump_file,
-			   "            alt=%d: reload pseudo for op %d "
-			   " can not hold the mode value -- refuse\n",
-			   nalt, nop);
-		      goto fail;
-		    }
+		  && (prohibited_class_reg_set_mode_p
+		      (this_alternative, this_alternative_set,
+		       GET_MODE (*curr_id->operand_loc[nop]))))
+		{
+		  if (lra_dump_file != NULL)
+		    fprintf (lra_dump_file,
+			     "            alt=%d: reload pseudo for op %d "
+			     " can not hold the mode value -- refuse\n",
+			     nalt, nop);
+		  goto fail;
 		}
 
 	      /* Check strong discouragement of reload of non-constant
@@ -3732,6 +3740,11 @@  curr_insn_transform (bool check_only_p)
 	      && regno < new_regno_start
 	      && ! lra_former_scratch_p (regno)
 	      && reg_renumber[regno] < 0
+	      /* Check that the optional reload pseudo will be able to
+		 hold given mode value.  */
+	      && ! (prohibited_class_reg_set_mode_p
+		    (goal_alt[i], reg_class_contents[goal_alt[i]],
+		     PSEUDO_REGNO_MODE (regno)))
 	      && (curr_insn_set == NULL_RTX
 		  || !((REG_P (SET_SRC (curr_insn_set))
 			|| MEM_P (SET_SRC (curr_insn_set))
Index: testsuite/gcc.target/i386/pr64617.c
===================================================================
--- testsuite/gcc.target/i386/pr64617.c	(revision 0)
+++ testsuite/gcc.target/i386/pr64617.c	(working copy)
@@ -0,0 +1,37 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O -ftree-vectorize -mavx512bw -march=slm" } */
+
+unsigned short out2[128 * 8], b0, b1, b2, b3, b4, b5, b6, b7, b8;
+
+void
+foo (unsigned short a0, unsigned short a1, unsigned short a2,
+     unsigned short a3, unsigned short a4, unsigned short a5,
+     unsigned short a6, unsigned short a7, unsigned short a8)
+{
+  int i;
+  for (i = 0; i < 128; i++)
+    {
+      out2[i * 4] = a0 + 8;
+    }
+  for (i = 0; i < 128; i++)
+    {
+      b0 = a0 + 8;
+      b1 = a1 + 7;
+      b2 = a2 + 6;
+      b3 = a3 + 5;
+      b4 = a4 + 4;
+      b5 = a5 + 3;
+      b6 = a6 + 2;
+      b7 = a7 + 1;
+      b8 = a8 + 9;
+
+      out2[i * 8] = b0;
+      out2[i * 8 + 1] = b1;
+      out2[i * 8 + 2] = b4;
+      out2[i * 8 + 3] = b5;
+      out2[i * 8 + 4] = b6;
+      out2[i * 8 + 5] = b2;
+      out2[i * 8 + 6] = b7;
+      out2[i * 8 + 7] = b8;
+    }
+}