gcc/

	PR rtl-optimization/28982
	* reload.c (find_reloads_address_1): Use RELOAD_OTHER for the
	index of a PRE_MODIFY or POST_MODIFY address.
	* reload1.c (inc_for_reload): Use find_replacement on the original
	base and index registers.

gcc/testsuite/
	PR rtl-optimization/28982
	* gcc.c-torture/execute/pr28982a.c: New test.
	* gcc.c-torture/execute/pr28982b.c: Likewise.


git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@116919 138bc75d-0d04-0410-961f-82ee72b054a4

gcc/ChangeLog

@@ -1,3 +1,11 @@
+2006-09-13  Richard Sandiford  <[email protected]>
+
+	PR rtl-optimization/28982
+	* reload.c (find_reloads_address_1): Use RELOAD_OTHER for the
+	index of a PRE_MODIFY or POST_MODIFY address.
+	* reload1.c (inc_for_reload): Use find_replacement on the original
+	base and index registers.
+
 2006-09-12  H.J. Lu  <[email protected]>
 
 	* doc/invoke.texi (mpreferred-stack-boundary): Remove exception

gcc/reload.c

@@ -5541,12 +5541,18 @@ find_reloads_address_1 (enum machine_mode mode, rtx x, int context,
 	/* Require index register (or constant).  Let's just handle the
 	   register case in the meantime... If the target allows
 	   auto-modify by a constant then we could try replacing a pseudo
-	   register with its equivalent constant where applicable.  */
+	   register with its equivalent constant where applicable.
+
+	   If we later decide to reload the whole PRE_MODIFY or
+	   POST_MODIFY, inc_for_reload might clobber the reload register
+	   before reading the index.  The index register might therefore
+	   need to live longer than a TYPE reload normally would, so be
+	   conservative and class it as RELOAD_OTHER.  */
 	if (REG_P (XEXP (op1, 1)))
 	  if (!REGNO_OK_FOR_INDEX_P (REGNO (XEXP (op1, 1))))
 	    find_reloads_address_1 (mode, XEXP (op1, 1), 1, code, SCRATCH,
-				    &XEXP (op1, 1), opnum, type, ind_levels,
-				    insn);
+				    &XEXP (op1, 1), opnum, RELOAD_OTHER,
+				    ind_levels, insn);
 
 	gcc_assert (REG_P (XEXP (op1, 0)));
 

gcc/reload1.c

@@ -8177,7 +8177,7 @@ static rtx
 inc_for_reload (rtx reloadreg, rtx in, rtx value, int inc_amount)
 {
   /* REG or MEM to be copied and incremented.  */
-  rtx incloc = XEXP (value, 0);
+  rtx incloc = find_replacement (&XEXP (value, 0));
   /* Nonzero if increment after copying.  */
   int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
 	      || GET_CODE (value) == POST_MODIFY);
@@ -8186,7 +8186,7 @@ inc_for_reload (rtx reloadreg, rtx in, rtx value, int inc_amount)
   rtx add_insn;
   int code;
   rtx store;
-  rtx real_in = in == value ? XEXP (in, 0) : in;
+  rtx real_in = in == value ? incloc : in;
 
   /* No hard register is equivalent to this register after
      inc/dec operation.  If REG_LAST_RELOAD_REG were nonzero,
@@ -8198,7 +8198,7 @@ inc_for_reload (rtx reloadreg, rtx in, rtx value, int inc_amount)
   if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
     {
       gcc_assert (GET_CODE (XEXP (value, 1)) == PLUS);
-      inc = XEXP (XEXP (value, 1), 1);
+      inc = find_replacement (&XEXP (XEXP (value, 1), 1));
     }
   else
     {

gcc/testsuite/ChangeLog

@@ -1,3 +1,9 @@
+2006-09-13  Richard Sandiford  <[email protected]>
+
+	PR rtl-optimization/28982
+	* gcc.c-torture/execute/pr28982a.c: New test.
+	* gcc.c-torture/execute/pr28982b.c: Likewise.
+
 2006-09-12  Eric Christopher  <[email protected]>
 
 	* gcc.target/x86_64/abi/asm-support-darwin.s: New.

gcc/testsuite/gcc.c-torture/execute/pr28982a.c

@@ -0,0 +1,65 @@
+/* PR rtl-optimization/28982.  Function foo() does the equivalent of:
+
+     float tmp_results[NVARS];
+     for (int i = 0; i < NVARS; i++)
+       {
+	 int inc = incs[i];
+	 float *ptr = ptrs[i], result = 0;
+	 for (int j = 0; j < n; j++)
+	   result += *ptr, ptr += inc;
+	 tmp_results[i] = result;
+       }
+     memcpy (results, tmp_results, sizeof (results));
+
+   but without the outermost loop.  The idea is to create high register
+   pressure and ensure that some INC and PTR variables are spilled.
+
+   On ARM targets, sequences like "result += *ptr, ptr += inc" can
+   usually be implemented using (mem (post_modify ...)), and we do
+   indeed create such MEMs before reload for this testcase.  However,
+   (post_modify ...) is not a valid address for coprocessor loads, so
+   for -mfloat-abi=softfp, reload reloads the POST_MODIFY into a base
+   register.  GCC did not deal correctly with cases where the base and
+   index of the POST_MODIFY are themselves reloaded.  */
+#define NITER 4
+#define NVARS 20
+#define MULTI(X) \
+  X( 0), X( 1), X( 2), X( 3), X( 4), X( 5), X( 6), X( 7), X( 8), X( 9), \
+  X(10), X(11), X(12), X(13), X(14), X(15), X(16), X(17), X(18), X(19)
+
+#define DECLAREI(INDEX) inc##INDEX = incs[INDEX]
+#define DECLAREF(INDEX) *ptr##INDEX = ptrs[INDEX], result##INDEX = 0
+#define LOOP(INDEX) result##INDEX += *ptr##INDEX, ptr##INDEX += inc##INDEX
+#define COPYOUT(INDEX) results[INDEX] = result##INDEX
+
+float *ptrs[NVARS];
+float results[NVARS];
+int incs[NVARS];
+
+void __attribute__((noinline))
+foo (int n)
+{
+  int MULTI (DECLAREI);
+  float MULTI (DECLAREF);
+  while (n--)
+    MULTI (LOOP);
+  MULTI (COPYOUT);
+}
+
+float input[NITER * NVARS];
+
+int
+main (void)
+{
+  int i;
+
+  for (i = 0; i < NVARS; i++)
+    ptrs[i] = input + i, incs[i] = i;
+  for (i = 0; i < NITER * NVARS; i++)
+    input[i] = i;
+  foo (NITER);
+  for (i = 0; i < NVARS; i++)
+    if (results[i] != i * NITER * (NITER + 1) / 2)
+      return 1;
+  return 0;
+}

gcc/testsuite/gcc.c-torture/execute/pr28982b.c

@@ -0,0 +1,58 @@
+/* Like pr28982a.c, but with the spill slots outside the range of
+   a single sp-based load on ARM.  This test tests for cases where
+   the addresses in the base and index reloads require further reloads.  */
+#if defined(STACK_SIZE) && STACK_SIZE <= 0x80100
+int main (void) { return 0; }
+#else
+#define NITER 4
+#define NVARS 20
+#define MULTI(X) \
+  X( 0), X( 1), X( 2), X( 3), X( 4), X( 5), X( 6), X( 7), X( 8), X( 9), \
+  X(10), X(11), X(12), X(13), X(14), X(15), X(16), X(17), X(18), X(19)
+
+#define DECLAREI(INDEX) inc##INDEX = incs[INDEX]
+#define DECLAREF(INDEX) *ptr##INDEX = ptrs[INDEX], result##INDEX = 0
+#define LOOP(INDEX) result##INDEX += *ptr##INDEX, ptr##INDEX += inc##INDEX
+#define COPYOUT(INDEX) results[INDEX] = result##INDEX
+
+float *ptrs[NVARS];
+float results[NVARS];
+int incs[NVARS];
+
+struct big { int i[0x10000]; };
+void __attribute__((noinline))
+bar (struct big b)
+{
+  incs[0] += b.i[0];
+}
+
+void __attribute__((noinline))
+foo (int n)
+{
+  struct big b = {};
+  int MULTI (DECLAREI);
+  float MULTI (DECLAREF);
+  while (n--)
+    MULTI (LOOP);
+  MULTI (COPYOUT);
+  bar (b);
+}
+
+float input[NITER * NVARS];
+
+int
+main (void)
+{
+  int i;
+
+  for (i = 0; i < NVARS; i++)
+    ptrs[i] = input + i, incs[i] = i;
+  for (i = 0; i < NITER * NVARS; i++)
+    input[i] = i;
+  foo (NITER);
+  for (i = 0; i < NVARS; i++)
+    if (results[i] != i * NITER * (NITER + 1) / 2)
+      return 1;
+  return 0;
+}
+#endif