[PATCH] Mips: Inline memcpy for R6

[Simon Dardis <Simon dot Dardis at imgtec dot com>]

Hello,

This patch enables inline memcpy for R6 which was previously 
disabled and adds support for expansion when source and 
destination are at least half-word aligned.

gcc/

  	* config/mips/mips.c (mips_expand_block_move): Enable inline memcpy
	expansion when !ISA_HAS_LWL_LWR.
	(mips_block_move_straight): Update the size of elements copied to
	account for alignment when !ISA_HAS_LWL_LWR.
	* config/mips/mips.h (MIPS_MIN_MOVE_MEM_ALIGN): New macro.

gcc/testsuite/

	* inline-memcpy-1.c: Test for inline expansion of memcpy.
	* inline-memcpy-2.c: Ditto.
	* inline-memcpy-3.c: Ditto.
	* inline-memcpy-4.c: Ditto.
	* inline-memcpy-5.c: Ditto.

Thanks,
Simon

diff --git a/gcc/config/mips/mips.c b/gcc/config/mips/mips.c
index 6f5421a..1f7c105 100644
--- a/gcc/config/mips/mips.c
+++ b/gcc/config/mips/mips.c
@@ -8187,12 +8187,22 @@ mips_block_move_straight (rtx dest, rtx src, HOST_WIDE_INT length)
      half-word alignment, it is usually better to move in half words.
      For instance, lh/lh/sh/sh is usually better than lwl/lwr/swl/swr
      and lw/lw/sw/sw is usually better than ldl/ldr/sdl/sdr.
-     Otherwise move word-sized chunks.  */
-  if (MEM_ALIGN (src) == BITS_PER_WORD / 2
-      && MEM_ALIGN (dest) == BITS_PER_WORD / 2)
-    bits = BITS_PER_WORD / 2;
+     Otherwise move word-sized chunks.
+
+     For ISA_HAS_LWL_LWR we rely on the lwl/lwr & swl/swr load. Otherwise
+     picking the minimum of alignment or BITS_PER_WORD gets us the
+     desired size for bits.  */
+
+  if (!ISA_HAS_LWL_LWR)
+    bits = MIN (BITS_PER_WORD, MIN (MEM_ALIGN (src), MEM_ALIGN (dest)));
   else
-    bits = BITS_PER_WORD;
+    {
+      if (MEM_ALIGN (src) == BITS_PER_WORD / 2
+	  && MEM_ALIGN (dest) == BITS_PER_WORD / 2)
+	bits = BITS_PER_WORD / 2;
+      else
+	bits = BITS_PER_WORD;
+    }
 
   mode = mode_for_size (bits, MODE_INT, 0);
   delta = bits / BITS_PER_UNIT;
@@ -8311,8 +8321,8 @@ bool
 mips_expand_block_move (rtx dest, rtx src, rtx length)
 {
   if (!ISA_HAS_LWL_LWR
-       && (MEM_ALIGN (src) < BITS_PER_WORD
-	   || MEM_ALIGN (dest) < BITS_PER_WORD))
+      && (MEM_ALIGN (src) < MIPS_MIN_MOVE_MEM_ALIGN
+	  || MEM_ALIGN (dest) < MIPS_MIN_MOVE_MEM_ALIGN))
     return false;
 
   if (CONST_INT_P (length))
diff --git a/gcc/config/mips/mips.h b/gcc/config/mips/mips.h
index a2380e5..6578ae5 100644
--- a/gcc/config/mips/mips.h
+++ b/gcc/config/mips/mips.h
@@ -3041,6 +3041,9 @@ while (0)
 #undef PTRDIFF_TYPE
 #define PTRDIFF_TYPE (POINTER_SIZE == 64 ? "long int" : "int")
 
+/* The minimum alignment of any expanded block move.  */
+#define MIPS_MIN_MOVE_MEM_ALIGN 16
+
 /* The maximum number of bytes that can be copied by one iteration of
    a movmemsi loop; see mips_block_move_loop.  */
 #define MIPS_MAX_MOVE_BYTES_PER_LOOP_ITER \
diff --git a/gcc/testsuite/gcc.target/mips/inline-memcpy-1.c b/gcc/testsuite/gcc.target/mips/inline-memcpy-1.c
new file mode 100644
index 0000000..5a254b1
--- /dev/null
+++ b/gcc/testsuite/gcc.target/mips/inline-memcpy-1.c
@@ -0,0 +1,16 @@
+/* { dg-options "-fno-common isa_rev>=6" } */
+/* { dg-skip-if "code quality test" { *-*-* } { "-O0" "-Os" } { "" } } */
+/* { dg-final { scan-assembler-not "\tmemcpy" } } */
+
+/* Test that memcpy is inline for target hardware
+   without swl, swr.  */
+
+#include <string.h>
+
+char c[40] __attribute__ ((aligned(8)));
+
+void
+f1 ()
+{
+  memcpy (c, "1234567890QWERTYUIOPASDFGHJKLZXCVBNM", 32);
+}
diff --git a/gcc/testsuite/gcc.target/mips/inline-memcpy-2.c b/gcc/testsuite/gcc.target/mips/inline-memcpy-2.c
new file mode 100644
index 0000000..c06be15
--- /dev/null
+++ b/gcc/testsuite/gcc.target/mips/inline-memcpy-2.c
@@ -0,0 +1,17 @@
+/* { dg-options "-fno-common isa_rev>=6" } */
+/* { dg-skip-if "code quality test" { *-*-* } { "-O0" } { "" } } */
+/* { dg-final { scan-assembler-not "\tmemcpy" } } */
+/* { dg-final { scan-assembler-times "\tsh\t" 16 } } */
+
+/* Test that inline memcpy is expanded for target hardware without
+   swl, swr when alignment is halfword and sufficent shs are produced.  */
+
+#include <string.h>
+
+char c[40] __attribute__ ((aligned(2)));
+
+void
+f1 ()
+{
+  memcpy (c, "1234567890QWERTYUIOPASDFGHJKLZXCVBNM", 32);
+}
diff --git a/gcc/testsuite/gcc.target/mips/inline-memcpy-3.c b/gcc/testsuite/gcc.target/mips/inline-memcpy-3.c
new file mode 100644
index 0000000..96a0387
--- /dev/null
+++ b/gcc/testsuite/gcc.target/mips/inline-memcpy-3.c
@@ -0,0 +1,18 @@
+/* { dg-options "-fno-common isa_rev<=5" } */
+/* { dg-skip-if "code quality test" { *-*-* } { "-O0" "-Os"} { "" } } */
+/* { dg-final { scan-assembler-not "\tmemcpy" } } */
+/* { dg-final { scan-assembler-times "swl" 8 } } */
+/* { dg-final { scan-assembler-times "swr" 8 } } */
+
+/* Test that inline memcpy for hardware with swl, swr handles subword
+   alignment and produces enough swl/swrs for mips32.  */
+
+#include <string.h>
+
+char c[40] __attribute__ ((aligned(2)));
+
+void
+f1 ()
+{
+  memcpy (c, "1234567890QWERTYUIOPASDFGHJKLZXCVBNM", 32);
+}
diff --git a/gcc/testsuite/gcc.target/mips/inline-memcpy-4.c b/gcc/testsuite/gcc.target/mips/inline-memcpy-4.c
new file mode 100644
index 0000000..0e7a22e
--- /dev/null
+++ b/gcc/testsuite/gcc.target/mips/inline-memcpy-4.c
@@ -0,0 +1,18 @@
+/* { dg-options "-fno-common isa_rev<=5 -mabi=64" } */
+/* { dg-skip-if "code quality test" { *-*-* } { "-O0" "-Os"} { "" } } */
+/* { dg-final { scan-assembler-not "\tmemcpy" } } */
+/* { dg-final { scan-assembler-times "sdl" 4 } } */
+/* { dg-final { scan-assembler-times "sdr" 4 } } */
+
+/* Test that inline memcpy for hardware with sdl, sdr handles subword
+   alignment and produces enough sdl/sdrs on n64.  */
+
+#include <string.h>
+
+char c[40] __attribute__ ((aligned(2)));
+
+void
+f1 ()
+{
+  memcpy (c, "1234567890QWERTYUIOPASDFGHJKLZXCVBNM", 32);
+}
diff --git a/gcc/testsuite/gcc.target/mips/inline-memcpy-5.c b/gcc/testsuite/gcc.target/mips/inline-memcpy-5.c
new file mode 100644
index 0000000..1b9fa16
--- /dev/null
+++ b/gcc/testsuite/gcc.target/mips/inline-memcpy-5.c
@@ -0,0 +1,18 @@
+/* { dg-options "-fno-common isa_rev<=5 -mabi=n32" } */
+/* { dg-skip-if "code quality test" { *-*-* } { "-O0" "-Os"} { "" } } */
+/* { dg-final { scan-assembler-not "\tmemcpy" } } */
+/* { dg-final { scan-assembler-times "sdl" 4 } } */
+/* { dg-final { scan-assembler-times "sdr" 4 } } */
+
+/* Test that inline memcpy for hardware with sdl, sdr handles subword
+   alignment and produces enough sdr/sdls on n32.  */
+
+#include <string.h>
+
+char c[40] __attribute__ ((aligned(2)));
+
+void
+f1 ()
+{
+  memcpy (c, "1234567890QWERTYUIOPASDFGHJKLZXCVBNM", 32);
+}
-- 
2.1.0