expr.c (emit_block_move_via_libcall): Export.

	* expr.c (emit_block_move_via_libcall): Export.
	(clear_storage_via_libcall): Rename to ...
	(set_storage_via_libcall): ... this one; handle arbitrary value to set.
	(clear_storage): Update to set_storage_via_libcall.
	* expr.h (emit_block_move_via_libcall): Declare
	* config/i386/i386.h (stringop_alg): New enum.
	(MAX_STRINGOP_ALGS): New constant.
	(stringop_algs): New struct.
	(processor_costs): Add memcpy/memset descriptors.
	(TARGET_REP_MOVEL_OPTIMAL): Remove.
	* config/i386/i386.md (movsi_insv_1_rex64): New pattern.
	(strmovsi, strmovdi, strsetsi, strsetdi): Update; accept all operands
	for memset.
	* config/i386/i386.opt (minline-stringops-dynamically,
	mstringop-strategy): New
	parameters.
	* config/i386/i386-prostos.h (ix86_expand_movmem): Update prototype.
	(ix86_expand-clrmem): Rename to ...
	(ix86_expand_setmem): ... this one; update prototype.
	* config/i386/i386.c (DUMMY_STRINGOP_ALGS): New constant.
	(size_cost, i386_cost, i486_cost, pentium_cost, pentiumpro_cost,
	geode_cost, k6_cost, athlon_cost, k8_cost, pentium4_cost,
	nocona_cost, core2_cost, generic_cost, generic64_cost,
	generic32_cost): Add memcpy/memset descriptors.
	(x86_rep_movl_optimal): Remove.
	(stringop_alg): New static variable.
	(ix86_expand_aligntest): Handle predictions.
	(override_options): Add strgop_alg handling.
	(predict_jump): New function.
	(scale_counter): New function.
	(expand_set_or_movmem_via_loop): New function.
	(expand_movmem_via_rep_mov): New function.
	(expand_setmem_via_rep_stots): New function.
	(emit_strmov): New function.
	(expand_movmem_epilogue): New function.
	(expand_setmem_epilogue_via_loop): New function.
	(expand_setmem_epilogue): New function.
	(expand_movmem_prologue): New function.
	(expand_setmem_prologue): New function.
	(decide_alg): New function.
	(decide_alignment): New function.
	(ix86_exand_movmem): Rewrite.
	(promote_duplicated_reg): New function.
	(ix86_expand_clrmem): Rename to ...
	(ix86_expand_setmem): ... this one. Rewrite.

	* invoke.texi (minline-stringops-dynamically): New command line option.
	(mstringop-strategy): Likewise.

From-SVN: r119252

gcc/ChangeLog

 2006-11-27  Jan Hubicka  <jh@suse.cz>
 
+	* expr.c (emit_block_move_via_libcall): Export.
+	(clear_storage_via_libcall): Rename to ...
+	(set_storage_via_libcall): ... this one; handle arbitrary value to set.
+	(clear_storage): Update to set_storage_via_libcall.
+	* expr.h (emit_block_move_via_libcall): Declare
+	* config/i386/i386.h (stringop_alg): New enum.
+	(MAX_STRINGOP_ALGS): New constant.
+	(stringop_algs): New struct.
+	(processor_costs): Add memcpy/memset descriptors.
+	(TARGET_REP_MOVEL_OPTIMAL): Remove.
+	* config/i386/i386.md (movsi_insv_1_rex64): New pattern.
+	(strmovsi, strmovdi, strsetsi, strsetdi): Update; accept all operands
+	for memset.
+	* config/i386/i386.opt (minline-stringops-dynamically,
+	mstringop-strategy): New
+	parameters.
+	* config/i386/i386-prostos.h (ix86_expand_movmem): Update prototype.
+	(ix86_expand-clrmem): Rename to ...
+	(ix86_expand_setmem): ... this one; update prototype.
+	* config/i386/i386.c (DUMMY_STRINGOP_ALGS): New constant.
+	(size_cost, i386_cost, i486_cost, pentium_cost, pentiumpro_cost,
+	geode_cost, k6_cost, athlon_cost, k8_cost, pentium4_cost,
+	nocona_cost, core2_cost, generic_cost, generic64_cost,
+	generic32_cost): Add memcpy/memset descriptors.
+	(x86_rep_movl_optimal): Remove.
+	(stringop_alg): New static variable.
+	(ix86_expand_aligntest): Handle predictions.
+	(override_options): Add strgop_alg handling.
+	(predict_jump): New function.
+	(scale_counter): New function.
+	(expand_set_or_movmem_via_loop): New function.
+	(expand_movmem_via_rep_mov): New function.
+	(expand_setmem_via_rep_stots): New function.
+	(emit_strmov): New function.
+	(expand_movmem_epilogue): New function.
+	(expand_setmem_epilogue_via_loop): New function.
+	(expand_setmem_epilogue): New function.
+	(expand_movmem_prologue): New function.
+	(expand_setmem_prologue): New function.
+	(decide_alg): New function.
+	(decide_alignment): New function.
+	(ix86_exand_movmem): Rewrite.
+	(promote_duplicated_reg): New function.
+	(ix86_expand_clrmem): Rename to ...
+	(ix86_expand_setmem): ... this one. Rewrite.
+
+	* invoke.texi (minline-stringops-dynamically): New command line option.
+	(mstringop-strategy): Likewise.
+
+2006-11-27  Jan Hubicka  <jh@suse.cz>
+
 	* cfgexpand.c (construct_exit_block): Don't disturb end of last BB.
 	* rtl.texi (NOTE_INSN_FUNCTION_END): Remove.
 	* final.c (final_scan_insn): Don't scan NOTE_INSN_FUNCTION_END.

gcc/config/i386/i386-protos.h

@@ -50,8 +50,8 @@ extern bool x86_extended_QIreg_mentioned_p (rtx);
 extern bool x86_extended_reg_mentioned_p (rtx);
 extern enum machine_mode ix86_cc_mode (enum rtx_code, rtx, rtx);
 
-extern int ix86_expand_movmem (rtx, rtx, rtx, rtx);
-extern int ix86_expand_clrmem (rtx, rtx, rtx);
+extern int ix86_expand_movmem (rtx, rtx, rtx, rtx, rtx, rtx);
+extern int ix86_expand_setmem (rtx, rtx, rtx, rtx, rtx, rtx);
 extern int ix86_expand_strlen (rtx, rtx, rtx, rtx);
 
 extern bool legitimate_constant_p (rtx);

gcc/config/i386/i386.h

@@ -19,6 +19,38 @@ along with GCC; see the file COPYING.  If not, write to
 the Free Software Foundation, 51 Franklin Street, Fifth Floor,
 Boston, MA 02110-1301, USA.  */
 
+/* Algorithm to expand string function with.  */
+enum stringop_alg
+{
+   no_stringop,
+   libcall,
+   rep_prefix_1_byte,
+   rep_prefix_4_byte,
+   rep_prefix_8_byte,
+   loop_1_byte,
+   loop,
+   unrolled_loop
+};
+#define NAX_STRINGOP_ALGS 4
+/* Specify what algorithm to use for stringops on known size.
+   When size is unknown, the UNKNOWN_SIZE alg is used.  When size is
+   known at compile time or estimated via feedback, the SIZE array
+   is walked in order until MAX is greater then the estimate (or -1
+   means infinity).  Corresponding ALG is used then.  
+   For example initializer:
+    {{256, loop}, {-1, rep_prefix_4_byte}}		
+   will use loop for blocks smaller or equal to 256 bytes, rep prefix will
+   be used otherwise.
+*/
+struct stringop_algs
+{
+  const enum stringop_alg unknown_size;
+  const struct stringop_strategy {
+    const int max;
+    const enum stringop_alg alg;
+  } size [NAX_STRINGOP_ALGS];
+};
+
 /* The purpose of this file is to define the characteristics of the i386,
    independent of assembler syntax or operating system.
 
@@ -84,6 +116,9 @@ struct processor_costs {
   const int fabs;		/* cost of FABS instruction.  */
   const int fchs;		/* cost of FCHS instruction.  */
   const int fsqrt;		/* cost of FSQRT instruction.  */
+				/* Specify what algorithm
+				   to use for stringops on unknown size.  */
+  struct stringop_algs memcpy[2], memset[2];
 };
 
 extern const struct processor_costs *ix86_cost;
@@ -217,7 +252,6 @@ extern int x86_prefetch_sse;
 #define TARGET_PREFETCH_SSE (x86_prefetch_sse)
 #define TARGET_SHIFT1 (x86_shift1 & TUNEMASK)
 #define TARGET_USE_FFREEP (x86_use_ffreep & TUNEMASK)
-#define TARGET_REP_MOVL_OPTIMAL (x86_rep_movl_optimal & TUNEMASK)
 #define TARGET_INTER_UNIT_MOVES (x86_inter_unit_moves & TUNEMASK)
 #define TARGET_FOUR_JUMP_LIMIT (x86_four_jump_limit & TUNEMASK)
 #define TARGET_SCHEDULE (x86_schedule & TUNEMASK)

gcc/config/i386/i386.md

@@ -1805,6 +1805,16 @@
   [(set_attr "type" "imov")
    (set_attr "mode" "QI")])
 
+(define_insn "*movsi_insv_1_rex64"
+  [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "+Q")
+			 (const_int 8)
+			 (const_int 8))
+	(match_operand:SI 1 "nonmemory_operand" "Qn"))]
+  "TARGET_64BIT"
+  "mov{b}\t{%b1, %h0|%h0, %b1}"
+  [(set_attr "type" "imov")
+   (set_attr "mode" "QI")])
+
 (define_insn "movdi_insv_1_rex64"
   [(set (zero_extract:DI (match_operand 0 "ext_register_operand" "+Q")
 			 (const_int 8)
@@ -18160,9 +18170,10 @@
    (use (match_operand:BLK 1 "memory_operand" ""))
    (use (match_operand:SI 2 "nonmemory_operand" ""))
    (use (match_operand:SI 3 "const_int_operand" ""))]
-  "! optimize_size || TARGET_INLINE_ALL_STRINGOPS"
+  ""
 {
- if (ix86_expand_movmem (operands[0], operands[1], operands[2], operands[3]))
+ if (ix86_expand_movmem (operands[0], operands[1], operands[2], operands[3],
+			 operands[3], constm1_rtx))
    DONE;
  else
    FAIL;
@@ -18175,7 +18186,8 @@
    (use (match_operand:DI 3 "const_int_operand" ""))]
   "TARGET_64BIT"
 {
- if (ix86_expand_movmem (operands[0], operands[1], operands[2], operands[3]))
+ if (ix86_expand_movmem (operands[0], operands[1], operands[2], operands[3],
+			 operands[3], constm1_rtx))
    DONE;
  else
    FAIL;
@@ -18450,11 +18462,9 @@
     (use (match_operand 3 "const_int_operand" ""))]
   ""
 {
- /* If value to set is not zero, use the library routine.  */
- if (operands[2] != const0_rtx)
-   FAIL;
-
- if (ix86_expand_clrmem (operands[0], operands[1], operands[3]))
+ if (ix86_expand_setmem (operands[0], operands[1],
+			 operands[2], operands[3],
+			 operands[3], constm1_rtx))
    DONE;
  else
    FAIL;
@@ -18464,14 +18474,14 @@
    [(use (match_operand:BLK 0 "memory_operand" ""))
     (use (match_operand:DI 1 "nonmemory_operand" ""))
     (use (match_operand 2 "const_int_operand" ""))
-    (use (match_operand 3 "const_int_operand" ""))]
+    (use (match_operand 3 "const_int_operand" ""))
+    (use (match_operand 4 "const_int_operand" ""))
+    (use (match_operand 5 "const_int_operand" ""))]
   "TARGET_64BIT"
 {
- /* If value to set is not zero, use the library routine.  */
- if (operands[2] != const0_rtx)
-   FAIL;
-
- if (ix86_expand_clrmem (operands[0], operands[1], operands[3]))
+ if (ix86_expand_setmem (operands[0], operands[1],
+			 operands[2], operands[3],
+			 operands[3], constm1_rtx))
    DONE;
  else
    FAIL;

gcc/config/i386/i386.opt

@@ -125,6 +125,10 @@ minline-all-stringops
 Target Report Mask(INLINE_ALL_STRINGOPS)
 Inline all known string operations
 
+minline-stringops-dynamically
+Target Report Mask(INLINE_STRINGOPS_DYNAMICALLY)
+Inline memset/memcpy string operations, but perform inline version only for small blocks
+
 mintel-syntax
 Target Undocumented
 ;; Deprecated
@@ -221,6 +225,10 @@ mstack-arg-probe
 Target Report Mask(STACK_PROBE)
 Enable stack probing
 
+mstringop-strategy=
+Target RejectNegative Joined Var(ix86_stringop_string)
+Chose strategy to generate stringop using
+
 mtls-dialect=
 Target RejectNegative Joined Var(ix86_tls_dialect_string)
 Use given thread-local storage dialect

gcc/doc/invoke.texi

@@ -9721,6 +9721,19 @@ aligned at least to 4 byte boundary.  This enables more inlining, increase code
 size, but may improve performance of code that depends on fast memcpy, strlen
 and memset for short lengths.
 
+@item -minline-stringops-dynamically
+@opindex minline-stringops-dynamically
+For string operation of unknown size, inline runtime checks so for small
+blocks inline code is used, while for large blocks librarly call is used.
+
+@item -mstringop-strategy=@var{alg}
+@optindex mstringop-strategy=@var{alg}
+Overwrite internal decision heuristic about particular algorithm to inline
+string opteration with.  The allowed values are @code{rep_byte},
+@code{rep_4byte}, @code{rep_8byte} for expanding using i386 @code{rep} prefix
+of specified size, @code{loop}, @code{unrolled_loop} for expanding inline loop,
+@code{libcall} for always expanding library call.
+
 @item -momit-leaf-frame-pointer
 @opindex momit-leaf-frame-pointer
 Don't keep the frame pointer in a register for leaf functions.  This

gcc/expr.c

@@ -127,7 +127,6 @@ static void move_by_pieces_1 (rtx (*) (rtx, ...), enum machine_mode,
 			      struct move_by_pieces *);
 static bool block_move_libcall_safe_for_call_parm (void);
 static bool emit_block_move_via_movmem (rtx, rtx, rtx, unsigned);
-static rtx emit_block_move_via_libcall (rtx, rtx, rtx, bool);
 static tree emit_block_move_libcall_fn (int);
 static void emit_block_move_via_loop (rtx, rtx, rtx, unsigned);
 static rtx clear_by_pieces_1 (void *, HOST_WIDE_INT, enum machine_mode);
@@ -135,7 +134,6 @@ static void clear_by_pieces (rtx, unsigned HOST_WIDE_INT, unsigned int);
 static void store_by_pieces_1 (struct store_by_pieces *, unsigned int);
 static void store_by_pieces_2 (rtx (*) (rtx, ...), enum machine_mode,
 			       struct store_by_pieces *);
-static rtx clear_storage_via_libcall (rtx, rtx, bool);
 static tree clear_storage_libcall_fn (int);
 static rtx compress_float_constant (rtx, rtx);
 static rtx get_subtarget (rtx);
@@ -1336,7 +1334,7 @@ emit_block_move_via_movmem (rtx x, rtx y, rtx size, unsigned int align)
 /* A subroutine of emit_block_move.  Expand a call to memcpy.
    Return the return value from memcpy, 0 otherwise.  */
 
-static rtx
+rtx
 emit_block_move_via_libcall (rtx dst, rtx src, rtx size, bool tailcall)
 {
   rtx dst_addr, src_addr;
@@ -2540,8 +2538,8 @@ clear_storage (rtx object, rtx size, enum block_op_methods method)
   else if (set_storage_via_setmem (object, size, const0_rtx, align))
     ;
   else
-    return clear_storage_via_libcall (object, size,
-				      method == BLOCK_OP_TAILCALL);
+    return set_storage_via_libcall (object, size, const0_rtx,
+				    method == BLOCK_OP_TAILCALL);
 
   return NULL;
 }
@@ -2549,10 +2547,10 @@ clear_storage (rtx object, rtx size, enum block_op_methods method)
 /* A subroutine of clear_storage.  Expand a call to memset.
    Return the return value of memset, 0 otherwise.  */
 
-static rtx
-clear_storage_via_libcall (rtx object, rtx size, bool tailcall)
+rtx
+set_storage_via_libcall (rtx object, rtx size, rtx val, bool tailcall)
 {
-  tree call_expr, arg_list, fn, object_tree, size_tree;
+  tree call_expr, arg_list, fn, object_tree, size_tree, val_tree;
   enum machine_mode size_mode;
   rtx retval;
 
@@ -2572,11 +2570,14 @@ clear_storage_via_libcall (rtx object, rtx size, bool tailcall)
      for returning pointers, we could end up generating incorrect code.  */
 
   object_tree = make_tree (ptr_type_node, object);
+  if (GET_CODE (val) != CONST_INT)
+    val = convert_to_mode (TYPE_MODE (integer_type_node), val, 1);
   size_tree = make_tree (sizetype, size);
+  val_tree = make_tree (integer_type_node, val);
 
   fn = clear_storage_libcall_fn (true);
   arg_list = tree_cons (NULL_TREE, size_tree, NULL_TREE);
-  arg_list = tree_cons (NULL_TREE, integer_zero_node, arg_list);
+  arg_list = tree_cons (NULL_TREE, val_tree, arg_list);
   arg_list = tree_cons (NULL_TREE, object_tree, arg_list);
 
   /* Now we have to build up the CALL_EXPR itself.  */
@@ -2590,7 +2591,7 @@ clear_storage_via_libcall (rtx object, rtx size, bool tailcall)
   return retval;
 }
 
-/* A subroutine of clear_storage_via_libcall.  Create the tree node
+/* A subroutine of set_storage_via_libcall.  Create the tree node
    for the function we use for block clears.  The first time FOR_CALL
    is true, we call assemble_external.  */
 

gcc/expr.h

@@ -377,6 +377,7 @@ extern void init_block_move_fn (const char *);
 extern void init_block_clear_fn (const char *);
 
 extern rtx emit_block_move (rtx, rtx, rtx, enum block_op_methods);
+extern rtx emit_block_move_via_libcall (rtx, rtx, rtx, bool);
 
 /* Copy all or part of a value X into registers starting at REGNO.
    The number of registers to be filled is NREGS.  */
@@ -423,6 +424,8 @@ extern void use_group_regs (rtx *, rtx);
 /* Write zeros through the storage of OBJECT.
    If OBJECT has BLKmode, SIZE is its length in bytes.  */
 extern rtx clear_storage (rtx, rtx, enum block_op_methods);
+/* The same, but always output an library call.  */
+rtx set_storage_via_libcall (rtx, rtx, rtx, bool);
 
 /* Expand a setmem pattern; return true if successful.  */
 extern bool set_storage_via_setmem (rtx, rtx, rtx, unsigned int);

gcc/testsuite/ChangeLog

+2006-11-27  Jan Hubicka  <jh@suse.cz>
+
+	* gcc.target/i386/memcpy-1.c: Adjust size.
+	* testsuite/gcc.dg/visibility-11.c: Likewise.
+
 2006-11-27  Richard Guenther  <rguenther@suse.de>
 
 	PR middle-end/25620

gcc/testsuite/gcc.dg/visibility-11.c

@@ -13,7 +13,7 @@
 extern void* memcpy (void *, const void *, __SIZE_TYPE__);
 #pragma GCC visibility pop
 
-struct a { int a[1024]; };
+struct a { int a[4096]; };
 
 extern void *bar (struct a *, struct a *, int);
 

gcc/testsuite/gcc.target/i386/memcpy-1.c

@@ -9,9 +9,9 @@
 /* A and B are aligned, but we used to lose track of it.
    Ensure that memcpy is inlined and alignment prologue is missing.  */
 
-char a[900];
-char b[900];
+char a[2048];
+char b[2048];
 t()
 {
-  __builtin_memcpy (a,b,900);
+  __builtin_memcpy (a,b,2048);
 }