diff mbox

[rs6000] Fix analyze_swaps to handle vperm for large and small code models

Message ID 1448980341.3133.4.camel@gnopaine
State New
Headers show

Commit Message

Bill Schmidt Dec. 1, 2015, 2:32 p.m. UTC 
Hi,

Uli Weigand discovered that the gcc.target/powerpc/swaps-p8-21.c test
case fails when large and small code models are used, rather than the
default medium code model.  This is because analyze_swaps is determining
whether the mask used for a vperm insn is loaded from the constant pool,
and there is an extra indirection for such loads when the large or small
code model is used.  This patch changes analyze_swaps to handle the
extra indirection correctly.  A new test case variant is added to check
for it.

Bootstrapped and tested on powerpc64le-unknown-linux-gnu with no
regressions.  Ok for trunk?

Thanks,
Bill


[gcc]

2015-12-01  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>

	* config/rs6000/rs6000.c (const_load_sequence_p): Handle extra
	indirection for large and small code models.
	(adjust_vperm): Likewise.

[gcc/testsuite]

2015-12-01  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>

	* gcc.target/powerpc/swaps-p8-22.c: New.
diff mbox

Patch

Index: gcc/config/rs6000/rs6000.c
===================================================================
--- gcc/config/rs6000/rs6000.c	(revision 231083)
+++ gcc/config/rs6000/rs6000.c	(working copy)
@@ -36613,7 +36613,12 @@  const_load_sequence_p (swap_web_entry *insn_entry,
 	  rtx base, offset;
 	  if (GET_CODE (tocrel_body) != SET)
 	    return false;
-	  if (!toc_relative_expr_p (SET_SRC (tocrel_body), false))
+	  /* There is an extra level of indirection for small/large
+	     code models.  */
+	  rtx tocrel_expr = SET_SRC (tocrel_body);
+	  if (GET_CODE (tocrel_expr) == MEM)
+	    tocrel_expr = XEXP (tocrel_expr, 0);
+	  if (!toc_relative_expr_p (tocrel_expr, false))
 	    return false;
 	  split_const (XVECEXP (tocrel_base, 0, 0), &base, &offset);
 	  if (GET_CODE (base) != SYMBOL_REF || !CONSTANT_POOL_ADDRESS_P (base))
@@ -37294,10 +37299,19 @@  adjust_vperm (rtx_insn *insn)
      to set tocrel_base; otherwise it would be unnecessary as we've
      already established it will return true.  */
   rtx base, offset;
-  if (!toc_relative_expr_p (SET_SRC (PATTERN (tocrel_insn)), false))
+  rtx tocrel_expr = SET_SRC (PATTERN (tocrel_insn));
+  /* There is an extra level of indirection for small/large code models.  */
+  if (GET_CODE (tocrel_expr) == MEM)
+    tocrel_expr = XEXP (tocrel_expr, 0);
+  if (!toc_relative_expr_p (tocrel_expr, false))
     gcc_unreachable ();
   split_const (XVECEXP (tocrel_base, 0, 0), &base, &offset);
   rtx const_vector = get_pool_constant (base);
+  /* With the extra indirection, get_pool_constant will produce the
+     real constant from the reg_equal expression, so get the real
+     constant.  */
+  if (GET_CODE (const_vector) == SYMBOL_REF)
+    const_vector = get_pool_constant (const_vector);
   gcc_assert (GET_CODE (const_vector) == CONST_VECTOR);
 
   /* Create an adjusted mask from the initial mask.  */
Index: gcc/testsuite/gcc.target/powerpc/swaps-p8-22.c
===================================================================
--- gcc/testsuite/gcc.target/powerpc/swaps-p8-22.c	(revision 0)
+++ gcc/testsuite/gcc.target/powerpc/swaps-p8-22.c	(working copy)
@@ -0,0 +1,29 @@ 
+/* { dg-do compile { target { powerpc64le-*-* } } } */
+/* { dg-skip-if "do not override -mcpu" { powerpc*-*-* } { "-mcpu=*" } { "-mcpu=power8" } }
+/* { dg-options "-O2 -mcpu=power8 -maltivec -mcmodel=large" } */
+
+/* The expansion for vector character multiply introduces a vperm operation.
+   This tests that changing the vperm mask allows us to remove all swaps
+   from the generated code.  It is a duplicate of swaps-p8-21.c, except
+   that it applies the large code model, which requires an extra indirection
+   in the load of the constant mask.  */
+
+#include <altivec.h>
+
+void abort ();
+
+vector unsigned char r;
+vector unsigned char v =
+  { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
+vector unsigned char i =
+  { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16 };
+
+int main ()
+{
+  int j;
+  r = v * i;
+  return 0;
+}
+
+/* { dg-final { scan-assembler-times "vperm" 1 } } */
+/* { dg-final { scan-assembler-not "xxpermdi" } } */