Patchwork [rs6000] Fix insertion of nop[s] to force dependent load into new dispatch group.

login
register
mail settings
Submitter Pat Haugen
Date May 23, 2012, 3:48 p.m.
Message ID <4FBD06CE.4060709@linux.vnet.ibm.com>
Download mbox | patch
Permalink /patch/160956/
State New
Headers show

Comments

Pat Haugen - May 23, 2012, 3:48 p.m.
The following patch fixes existing code that tried to prevent load-hit-store (LHS) from being in the same dispatch group. The main problem was use of the wrong dependency list in is_costly_group(), but I also added code to verify the memory refs overlap and to emit group ending nops for those ISAs that have them.

Bootstrap/regtest on powerpc64-linux with no new regressions. Ok for trunk?

-Pat


2012-05-23  Pat Haugen <pthaugen@us.ibm.com>

         * config/rs6000/rs6000.c (rs6000_option_override_internal): Change
         rs6000_sched_costly_dep default to true_store_to_load_dep_costly.
         (adjacent_mem_locations): Move some code to and call...
         (get_memref_parts): ...new function.
         (mem_locations_overlap): New function.
         (rs6000_adjust_priority): Adjust calls to is_load_insn/is_store_insn.
         (is_mem_ref): Rename to...
         (find_mem_ref): ...this. Return MEM rtx.
         (get_store_dest): Remove function.
         (is_load_insn1, is_load_insn, is_store_insn1, is_store_insn): Add
         new parameter and adjust calls.
         (rs6000_is_costly_dependence): Update calls for extra arg. Make sure
         mem refs overlap for true_store_to_load_dep_costly.
         (rs6000_sched_reorder2): Update calls for extra arg. Adjust args
         passed to adjacent_mem_locations.
         (is_costly_group): Walk resolved dependency list.
         (force_new_group): Emit group ending nop for Power6/Power7.
         * config/rs6000/rs6000.md (UNSPEC_GRP_END_NOP): New enum value.
         (group_ending_nop): New define_insn.

gcc/testsuite:
         * gcc.target/powerpc/lhs-1.c: New.
         * gcc.target/powerpc/lhs-2.c: New.
         * gcc.target/powerpc/lhs-3.c: New.
David Edelsohn - May 24, 2012, 2:20 p.m.
On Wed, May 23, 2012 at 11:48 AM, Pat Haugen
<pthaugen@linux.vnet.ibm.com> wrote:
> The following patch fixes existing code that tried to prevent load-hit-store
> (LHS) from being in the same dispatch group. The main problem was use of the
> wrong dependency list in is_costly_group(), but I also added code to verify
> the memory refs overlap and to emit group ending nops for those ISAs that
> have them.
>
> Bootstrap/regtest on powerpc64-linux with no new regressions. Ok for trunk?
>
> -Pat
>
>
> 2012-05-23  Pat Haugen <pthaugen@us.ibm.com>
>
>        * config/rs6000/rs6000.c (rs6000_option_override_internal): Change
>        rs6000_sched_costly_dep default to true_store_to_load_dep_costly.
>        (adjacent_mem_locations): Move some code to and call...
>        (get_memref_parts): ...new function.
>        (mem_locations_overlap): New function.
>        (rs6000_adjust_priority): Adjust calls to is_load_insn/is_store_insn.
>        (is_mem_ref): Rename to...
>        (find_mem_ref): ...this. Return MEM rtx.
>        (get_store_dest): Remove function.
>        (is_load_insn1, is_load_insn, is_store_insn1, is_store_insn): Add
>        new parameter and adjust calls.
>        (rs6000_is_costly_dependence): Update calls for extra arg. Make sure
>        mem refs overlap for true_store_to_load_dep_costly.
>        (rs6000_sched_reorder2): Update calls for extra arg. Adjust args
>        passed to adjacent_mem_locations.
>        (is_costly_group): Walk resolved dependency list.
>        (force_new_group): Emit group ending nop for Power6/Power7.
>        * config/rs6000/rs6000.md (UNSPEC_GRP_END_NOP): New enum value.
>        (group_ending_nop): New define_insn.
>
> gcc/testsuite:
>        * gcc.target/powerpc/lhs-1.c: New.
>        * gcc.target/powerpc/lhs-2.c: New.
>        * gcc.target/powerpc/lhs-3.c: New.

I thought there was common infrastructure for recursively walking an
insn for use in the memref comparison code, but, as you mentioned
privately, other parts of GCC implement it manually.


+/* Determine is PAT refers to memory. If so, set MEM_REF to the MEM rtx
+   and return true.  */

I assume you mean "... if PAT" not "is PAT".

Okay with that typo fix.

Thanks, David
Pat Haugen - July 5, 2012, 4:16 p.m.
On 05/23/2012 10:48 AM, Pat Haugen wrote:
> The following patch fixes existing code that tried to prevent load-hit-store (LHS) from being in the same dispatch group. The main problem was use of the wrong dependency list in is_costly_group(), but I also added code to verify the memory refs overlap and to emit group ending nops for those ISAs that have them.
>
> Bootstrap/regtest on powerpc64-linux with no new regressions. Ok for trunk?
>
> -Pat
>
>
> 2012-05-23  Pat Haugen <pthaugen@us.ibm.com>
>
>          * config/rs6000/rs6000.c (rs6000_option_override_internal): Change
>          rs6000_sched_costly_dep default to true_store_to_load_dep_costly.
>          (adjacent_mem_locations): Move some code to and call...
>          (get_memref_parts): ...new function.
>          (mem_locations_overlap): New function.
>          (rs6000_adjust_priority): Adjust calls to is_load_insn/is_store_insn.
>          (is_mem_ref): Rename to...
>          (find_mem_ref): ...this. Return MEM rtx.
>          (get_store_dest): Remove function.
>          (is_load_insn1, is_load_insn, is_store_insn1, is_store_insn): Add
>          new parameter and adjust calls.
>          (rs6000_is_costly_dependence): Update calls for extra arg. Make sure
>          mem refs overlap for true_store_to_load_dep_costly.
>          (rs6000_sched_reorder2): Update calls for extra arg. Adjust args
>          passed to adjacent_mem_locations.
>          (is_costly_group): Walk resolved dependency list.
>          (force_new_group): Emit group ending nop for Power6/Power7.
>          * config/rs6000/rs6000.md (UNSPEC_GRP_END_NOP): New enum value.
>          (group_ending_nop): New define_insn.
>
> gcc/testsuite:
>          * gcc.target/powerpc/lhs-1.c: New.
>          * gcc.target/powerpc/lhs-2.c: New.
>          * gcc.target/powerpc/lhs-3.c: New.
>
>

Ok for backport to 4.7 branch now that that is open again?

-Pat
David Edelsohn - July 5, 2012, 4:51 p.m.
On Thu, Jul 5, 2012 at 12:16 PM, Pat Haugen <pthaugen@linux.vnet.ibm.com> wrote:

> Ok for backport to 4.7 branch now that that is open again?

Yes.

Thanks, David

Patch

Index: gcc/config/rs6000/rs6000.c
===================================================================
--- gcc/config/rs6000/rs6000.c	(revision 187663)
+++ gcc/config/rs6000/rs6000.c	(working copy)
@@ -1005,9 +1005,8 @@  static bool is_microcoded_insn (rtx);
 static bool is_nonpipeline_insn (rtx);
 static bool is_cracked_insn (rtx);
 static bool is_branch_slot_insn (rtx);
-static bool is_load_insn (rtx);
-static rtx get_store_dest (rtx pat);
-static bool is_store_insn (rtx);
+static bool is_load_insn (rtx, rtx *);
+static bool is_store_insn (rtx, rtx *);
 static bool set_to_load_agen (rtx,rtx);
 static bool adjacent_mem_locations (rtx,rtx);
 static int rs6000_adjust_priority (rtx, int);
@@ -3046,7 +3045,7 @@  rs6000_option_override_internal (bool gl
 
   /* Handle -msched-costly-dep option.  */
   rs6000_sched_costly_dep
-    = (rs6000_sched_groups ? store_to_load_dep_costly : no_dep_costly);
+    = (rs6000_sched_groups ? true_store_to_load_dep_costly : no_dep_costly);
 
   if (rs6000_sched_costly_dep_str)
     {
@@ -22965,49 +22964,78 @@  set_to_load_agen (rtx out_insn, rtx in_i
   return false;
 }
 
-/* The function returns true if the target storage location of
-   out_insn is adjacent to the target storage location of in_insn */
-/* Return 1 if memory locations are adjacent.  */
+/* Try to determine base/offset/size parts of the given MEM.
+   Return true if successful, false if all the values couldn't
+   be determined.
+
+   This function only looks for REG or REG+CONST address forms.
+   REG+REG address form will return false. */
 
 static bool
-adjacent_mem_locations (rtx insn1, rtx insn2)
+get_memref_parts (rtx mem, rtx *base, HOST_WIDE_INT *offset,
+		  HOST_WIDE_INT *size)
 {
+  rtx addr_rtx;
+  if MEM_SIZE_KNOWN_P (mem)
+    *size = MEM_SIZE (mem);
+  else
+    return false;
 
-  rtx a = get_store_dest (PATTERN (insn1));
-  rtx b = get_store_dest (PATTERN (insn2));
+  if (GET_CODE (XEXP (mem, 0)) == PRE_MODIFY)
+    addr_rtx = XEXP (XEXP (mem, 0), 1);
+  else
+    addr_rtx = (XEXP (mem, 0));
 
-  if ((GET_CODE (XEXP (a, 0)) == REG
-       || (GET_CODE (XEXP (a, 0)) == PLUS
-	   && GET_CODE (XEXP (XEXP (a, 0), 1)) == CONST_INT))
-      && (GET_CODE (XEXP (b, 0)) == REG
-	  || (GET_CODE (XEXP (b, 0)) == PLUS
-	      && GET_CODE (XEXP (XEXP (b, 0), 1)) == CONST_INT)))
+  if (GET_CODE (addr_rtx) == REG)
+    {
+      *base = addr_rtx;
+      *offset = 0;
+    }
+  else if (GET_CODE (addr_rtx) == PLUS
+	   && CONST_INT_P (XEXP (addr_rtx, 1)))
     {
-      HOST_WIDE_INT val0 = 0, val1 = 0, val_diff;
-      rtx reg0, reg1;
+      *base = XEXP (addr_rtx, 0);
+      *offset = INTVAL (XEXP (addr_rtx, 1));
+    }
+  else
+    return false;
 
-      if (GET_CODE (XEXP (a, 0)) == PLUS)
-        {
-	  reg0 = XEXP (XEXP (a, 0), 0);
-	  val0 = INTVAL (XEXP (XEXP (a, 0), 1));
-        }
-      else
-	reg0 = XEXP (a, 0);
+  return true;
+}
 
-      if (GET_CODE (XEXP (b, 0)) == PLUS)
-        {
-	  reg1 = XEXP (XEXP (b, 0), 0);
-	  val1 = INTVAL (XEXP (XEXP (b, 0), 1));
-        }
-      else
-	reg1 = XEXP (b, 0);
+/* The function returns true if the target storage location of
+   mem1 is adjacent to the target storage location of mem2 */
+/* Return 1 if memory locations are adjacent.  */
 
-      val_diff = val1 - val0;
+static bool
+adjacent_mem_locations (rtx mem1, rtx mem2)
+{
+  rtx reg1, reg2;
+  HOST_WIDE_INT off1, size1, off2, size2;
 
-      return ((REGNO (reg0) == REGNO (reg1))
-	      && ((MEM_SIZE_KNOWN_P (a) && val_diff == MEM_SIZE (a))
-		  || (MEM_SIZE_KNOWN_P (b) && val_diff == -MEM_SIZE (b))));
-    }
+  if (get_memref_parts (mem1, &reg1, &off1, &size1)
+      && get_memref_parts (mem2, &reg2, &off2, &size2))
+    return ((REGNO (reg1) == REGNO (reg2))
+	    && ((off1 + size1 == off2)
+		|| (off2 + size2 == off1)));
+
+  return false;
+}
+
+/* This function returns true if it can be determined that the two MEM
+   locations overlap by at least 1 byte based on base reg/offset/size. */
+
+static bool
+mem_locations_overlap (rtx mem1, rtx mem2)
+{
+  rtx reg1, reg2;
+  HOST_WIDE_INT off1, size1, off2, size2;
+
+  if (get_memref_parts (mem1, &reg1, &off1, &size1)
+      && get_memref_parts (mem2, &reg2, &off2, &size2))
+    return ((REGNO (reg1) == REGNO (reg2))
+	    && (((off1 <= off2) && (off1 + size1 > off2))
+		|| ((off2 <= off1) && (off2 + size2 > off1))));
 
   return false;
 }
@@ -23021,6 +23049,7 @@  adjacent_mem_locations (rtx insn1, rtx i
 static int
 rs6000_adjust_priority (rtx insn ATTRIBUTE_UNUSED, int priority)
 {
+  rtx load_mem, str_mem;
   /* On machines (like the 750) which have asymmetric integer units,
      where one integer unit can do multiply and divides and the other
      can't, reduce the priority of multiply/divide so it is scheduled
@@ -23072,8 +23101,8 @@  rs6000_adjust_priority (rtx insn ATTRIBU
     }
 
   if (rs6000_cpu == PROCESSOR_POWER6
-      && ((load_store_pendulum == -2 && is_load_insn (insn))
-          || (load_store_pendulum == 2 && is_store_insn (insn))))
+      && ((load_store_pendulum == -2 && is_load_insn (insn, &load_mem))
+          || (load_store_pendulum == 2 && is_store_insn (insn, &str_mem))))
     /* Attach highest priority to insn if the scheduler has just issued two
        stores and this instruction is a load, or two loads and this instruction
        is a store. Power6 wants loads and stores scheduled alternately
@@ -23198,54 +23227,63 @@  rs6000_use_sched_lookahead_guard (rtx in
   return 1;
 }
 
-/* Determine is PAT refers to memory.  */
+/* Determine is PAT refers to memory. If so, set MEM_REF to the MEM rtx
+   and return true.  */
 
 static bool
-is_mem_ref (rtx pat)
+find_mem_ref (rtx pat, rtx *mem_ref)
 {
   const char * fmt;
   int i, j;
-  bool ret = false;
 
   /* stack_tie does not produce any real memory traffic.  */
   if (tie_operand (pat, VOIDmode))
     return false;
 
   if (GET_CODE (pat) == MEM)
-    return true;
+    {
+      *mem_ref = pat;
+      return true;
+    }
 
   /* Recursively process the pattern.  */
   fmt = GET_RTX_FORMAT (GET_CODE (pat));
 
-  for (i = GET_RTX_LENGTH (GET_CODE (pat)) - 1; i >= 0 && !ret; i--)
+  for (i = GET_RTX_LENGTH (GET_CODE (pat)) - 1; i >= 0; i--)
     {
       if (fmt[i] == 'e')
-	ret |= is_mem_ref (XEXP (pat, i));
+	{
+	  if (find_mem_ref (XEXP (pat, i), mem_ref))
+	    return true;
+	}
       else if (fmt[i] == 'E')
 	for (j = XVECLEN (pat, i) - 1; j >= 0; j--)
-	  ret |= is_mem_ref (XVECEXP (pat, i, j));
+	  {
+	    if (find_mem_ref (XVECEXP (pat, i, j), mem_ref))
+	      return true;
+	  }
     }
 
-  return ret;
+  return false;
 }
 
 /* Determine if PAT is a PATTERN of a load insn.  */
 
 static bool
-is_load_insn1 (rtx pat)
+is_load_insn1 (rtx pat, rtx *load_mem)
 {
   if (!pat || pat == NULL_RTX)
     return false;
 
   if (GET_CODE (pat) == SET)
-    return is_mem_ref (SET_SRC (pat));
+    return find_mem_ref (SET_SRC (pat), load_mem);
 
   if (GET_CODE (pat) == PARALLEL)
     {
       int i;
 
       for (i = 0; i < XVECLEN (pat, 0); i++)
-	if (is_load_insn1 (XVECEXP (pat, 0, i)))
+	if (is_load_insn1 (XVECEXP (pat, 0, i), load_mem))
 	  return true;
     }
 
@@ -23255,7 +23293,7 @@  is_load_insn1 (rtx pat)
 /* Determine if INSN loads from memory.  */
 
 static bool
-is_load_insn (rtx insn)
+is_load_insn (rtx insn, rtx *load_mem)
 {
   if (!insn || !INSN_P (insn))
     return false;
@@ -23263,26 +23301,26 @@  is_load_insn (rtx insn)
   if (GET_CODE (insn) == CALL_INSN)
     return false;
 
-  return is_load_insn1 (PATTERN (insn));
+  return is_load_insn1 (PATTERN (insn), load_mem);
 }
 
 /* Determine if PAT is a PATTERN of a store insn.  */
 
 static bool
-is_store_insn1 (rtx pat)
+is_store_insn1 (rtx pat, rtx *str_mem)
 {
   if (!pat || pat == NULL_RTX)
     return false;
 
   if (GET_CODE (pat) == SET)
-    return is_mem_ref (SET_DEST (pat));
+    return find_mem_ref (SET_DEST (pat), str_mem);
 
   if (GET_CODE (pat) == PARALLEL)
     {
       int i;
 
       for (i = 0; i < XVECLEN (pat, 0); i++)
-	if (is_store_insn1 (XVECEXP (pat, 0, i)))
+	if (is_store_insn1 (XVECEXP (pat, 0, i), str_mem))
 	  return true;
     }
 
@@ -23292,38 +23330,12 @@  is_store_insn1 (rtx pat)
 /* Determine if INSN stores to memory.  */
 
 static bool
-is_store_insn (rtx insn)
+is_store_insn (rtx insn, rtx *str_mem)
 {
   if (!insn || !INSN_P (insn))
     return false;
 
-  return is_store_insn1 (PATTERN (insn));
-}
-
-/* Return the dest of a store insn.  */
-
-static rtx
-get_store_dest (rtx pat)
-{
-  gcc_assert (is_store_insn1 (pat));
-
-  if (GET_CODE (pat) == SET)
-    return SET_DEST (pat);
-  else if (GET_CODE (pat) == PARALLEL)
-    {
-      int i;
-
-      for (i = 0; i < XVECLEN (pat, 0); i++)
-	{
-	  rtx inner_pat = XVECEXP (pat, 0, i);
-	  if (GET_CODE (inner_pat) == SET
-	      && is_mem_ref (SET_DEST (inner_pat)))
-	    return inner_pat;
-	}
-    }
-  /* We shouldn't get here, because we should have either a simple
-     store insn or a store with update which are covered above.  */
-  gcc_unreachable();
+  return is_store_insn1 (PATTERN (insn), str_mem);
 }
 
 /* Returns whether the dependence between INSN and NEXT is considered
@@ -23334,6 +23346,7 @@  rs6000_is_costly_dependence (dep_t dep,
 {
   rtx insn;
   rtx next;
+  rtx load_mem, str_mem;
 
   /* If the flag is not enabled - no dependence is considered costly;
      allow all dependent insns in the same group.
@@ -23351,15 +23364,16 @@  rs6000_is_costly_dependence (dep_t dep,
   next = DEP_CON (dep);
 
   if (rs6000_sched_costly_dep == store_to_load_dep_costly
-      && is_load_insn (next)
-      && is_store_insn (insn))
+      && is_load_insn (next, &load_mem)
+      && is_store_insn (insn, &str_mem))
     /* Prevent load after store in the same group.  */
     return true;
 
   if (rs6000_sched_costly_dep == true_store_to_load_dep_costly
-      && is_load_insn (next)
-      && is_store_insn (insn)
-      && DEP_TYPE (dep) == REG_DEP_TRUE)
+      && is_load_insn (next, &load_mem)
+      && is_store_insn (insn, &str_mem)
+      && DEP_TYPE (dep) == REG_DEP_TRUE
+      && mem_locations_overlap(str_mem, load_mem))
      /* Prevent load after store in the same group if it is a true
 	dependence.  */
      return true;
@@ -23486,12 +23500,12 @@  rs6000_sched_reorder2 (FILE *dump, int s
     {
       int pos;
       int i;
-      rtx tmp;
+      rtx tmp, load_mem, str_mem;
 
-      if (is_store_insn (last_scheduled_insn))
+      if (is_store_insn (last_scheduled_insn, &str_mem))
         /* Issuing a store, swing the load_store_pendulum to the left */
         load_store_pendulum--;
-      else if (is_load_insn (last_scheduled_insn))
+      else if (is_load_insn (last_scheduled_insn, &load_mem))
         /* Issuing a load, swing the load_store_pendulum to the right */
         load_store_pendulum++;
       else
@@ -23510,7 +23524,7 @@  rs6000_sched_reorder2 (FILE *dump, int s
 
           while (pos >= 0)
             {
-              if (is_load_insn (ready[pos]))
+              if (is_load_insn (ready[pos], &load_mem))
                 {
                   /* Found a load.  Move it to the head of the ready list,
                      and adjust it's priority so that it is more likely to
@@ -23536,7 +23550,7 @@  rs6000_sched_reorder2 (FILE *dump, int s
 
           while (pos >= 0)
             {
-              if (is_load_insn (ready[pos])
+              if (is_load_insn (ready[pos], &load_mem)
                   && !sel_sched_p ()
 		  && INSN_PRIORITY_KNOWN (ready[pos]))
                 {
@@ -23563,15 +23577,16 @@  rs6000_sched_reorder2 (FILE *dump, int s
 
           while (pos >= 0)
             {
-              if (is_store_insn (ready[pos]))
+              if (is_store_insn (ready[pos], &str_mem))
                 {
+		  rtx str_mem2;
                   /* Maintain the index of the first store found on the
                      list */
                   if (first_store_pos == -1)
                     first_store_pos = pos;
 
-                  if (is_store_insn (last_scheduled_insn)
-                      && adjacent_mem_locations (last_scheduled_insn,ready[pos]))
+                  if (is_store_insn (last_scheduled_insn, &str_mem2)
+                      && adjacent_mem_locations (str_mem, str_mem2))
                     {
                       /* Found an adjacent store.  Move it to the head of the
                          ready list, and adjust it's priority so that it is
@@ -23615,7 +23630,7 @@  rs6000_sched_reorder2 (FILE *dump, int s
 
           while (pos >= 0)
             {
-              if (is_store_insn (ready[pos])
+              if (is_store_insn (ready[pos], &str_mem)
                   && !sel_sched_p ()
 		  && INSN_PRIORITY_KNOWN (ready[pos]))
                 {
@@ -23899,7 +23914,7 @@  is_costly_group (rtx *group_insns, rtx n
       if (!insn)
 	continue;
 
-      FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep)
+      FOR_EACH_DEP (insn, SD_LIST_RES_FORW, sd_it, dep)
 	{
 	  rtx next = DEP_CON (dep);
 
@@ -23963,12 +23978,20 @@  force_new_group (int sched_verbose, FILE
       if (can_issue_more && !is_branch_slot_insn (next_insn))
 	can_issue_more--;
 
-      while (can_issue_more > 0)
+      /* Power6 and Power7 have special group ending nop. */
+      if (rs6000_cpu_attr == CPU_POWER6 || rs6000_cpu_attr == CPU_POWER7)
 	{
-	  nop = gen_nop ();
+	  nop = gen_group_ending_nop ();
 	  emit_insn_before (nop, next_insn);
-	  can_issue_more--;
+	  can_issue_more = 0;
 	}
+      else
+	while (can_issue_more > 0)
+	  {
+	    nop = gen_nop ();
+	    emit_insn_before (nop, next_insn);
+	    can_issue_more--;
+	  }
 
       *group_end = true;
       return 0;
Index: gcc/config/rs6000/rs6000.md
===================================================================
--- gcc/config/rs6000/rs6000.md	(revision 187663)
+++ gcc/config/rs6000/rs6000.md	(working copy)
@@ -125,6 +125,7 @@  (define_c_enum "unspec"
    UNSPEC_LFIWAX
    UNSPEC_LFIWZX
    UNSPEC_FCTIWUZ
+   UNSPEC_GRP_END_NOP
   ])
 
 ;;
@@ -15586,6 +15587,16 @@  (define_insn "nop"
   [(const_int 0)]
   ""
   "{cror 0,0,0|nop}")
+
+(define_insn "group_ending_nop"
+  [(unspec [(const_int 0)] UNSPEC_GRP_END_NOP)]
+  ""
+  "*
+{
+  if (rs6000_cpu_attr == CPU_POWER6)
+    return \"ori 1,1,0\";
+  return \"ori 2,2,0\";
+}")
 
 ;; Define the subtract-one-and-jump insns, starting with the template
 ;; so loop.c knows what to generate.
Index: gcc/testsuite/gcc.target/powerpc/lhs-1.c
===================================================================
--- gcc/testsuite/gcc.target/powerpc/lhs-1.c	(revision 0)
+++ gcc/testsuite/gcc.target/powerpc/lhs-1.c	(revision 0)
@@ -0,0 +1,22 @@ 
+/* { dg-do compile { target { powerpc*-*-* } } } */
+/* { dg-skip-if "" { powerpc*-*-darwin* } { "*" } { "" } } */
+/* { dg-options "-O2 -mcpu=power5" } */
+/* { dg-final { scan-assembler-times "nop" 3 } } */
+
+/* Test generation of nops in load hit store situation.  */
+
+typedef union {
+  double val;
+  struct {
+    unsigned int w1;
+    unsigned int w2;
+  };
+} words;
+
+unsigned int f (double d)
+{
+  words u;
+  u.val = d;
+  return u.w2;
+}
+
Index: gcc/testsuite/gcc.target/powerpc/lhs-2.c
===================================================================
--- gcc/testsuite/gcc.target/powerpc/lhs-2.c	(revision 0)
+++ gcc/testsuite/gcc.target/powerpc/lhs-2.c	(revision 0)
@@ -0,0 +1,21 @@ 
+/* { dg-do compile { target { powerpc*-*-* } } } */
+/* { dg-skip-if "" { powerpc*-*-darwin* } { "*" } { "" } } */
+/* { dg-options "-O2 -mcpu=power6 -msched-groups" } */
+/* { dg-final { scan-assembler "ori 1,1,0" } } */
+
+/* Test generation of group ending nop in load hit store situation.  */
+typedef union {
+  double val;
+  struct {
+    unsigned int w1;
+    unsigned int w2;
+  };
+} words;
+
+unsigned int f (double d)
+{
+  words u;
+  u.val = d;
+  return u.w2;
+}
+
Index: gcc/testsuite/gcc.target/powerpc/lhs-3.c
===================================================================
--- gcc/testsuite/gcc.target/powerpc/lhs-3.c	(revision 0)
+++ gcc/testsuite/gcc.target/powerpc/lhs-3.c	(revision 0)
@@ -0,0 +1,21 @@ 
+/* { dg-do compile { target { powerpc*-*-* } } } */
+/* { dg-skip-if "" { powerpc*-*-darwin* } { "*" } { "" } } */
+/* { dg-options "-O2 -mcpu=power7" } */
+/* { dg-final { scan-assembler "ori 2,2,0" } } */
+
+/* Test generation of group ending nop in load hit store situation.  */
+typedef union {
+  double val;
+  struct {
+    unsigned int w1;
+    unsigned int w2;
+  };
+} words;
+
+unsigned int f (double d)
+{
+  words u;
+  u.val = d;
+  return u.w2;
+}
+