This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Prefer mempcpy to memcpy on x86_64 target (PR middle-end/81657).
On 03/13/2018 04:23 PM, Jakub Jelinek wrote:
> On Tue, Mar 13, 2018 at 04:19:21PM +0100, Martin Liška wrote:
>>> Yes, see e.g. TARGET_LIBC_HAS_FUNCTION target hook,
>>> where in particular linux_libc_has_function deals with various C libraries.
>>> Of course, in this case you need another target hook, that is dependent both
>>> on the target backend and C library.
>>>
>>> It would be nice to make the target hook a little bit more generic as well,
>>> e.g. pass it enum builtin_function and query if it is fast, slow or
>>> unknown, or even some kind of cost, where the caller could ask for cost of
>>> BUILT_IN_MEMCPY and BUILT_IN_MEMPCPY and decide based on the relative costs.
>>
>> Let me start with simple return enum value of FAST,SLOW,UNKNOWN. I've added new hook
>> definition to gcc/config/gnu-user.h that will point to gnu_libc_function_implementation.
>> I would like to implement the function in gcc/targhooks.c, but I don't know how to
>> make ifdef according to target?
>
> Put there just the default implementation (everything is UNKNOWN?).
>
>> One another issue is that built_in_function is enum defined in tree.h. Thus I'll replace the
>> callback argument with int, that will be casted. One last issue: am I right that I'll have to define
>> TARGET_LIBC_FUNCTION_IMPLEMENTATION in each config file (similar to no_c99_libc_has_function)?
>
> And define the i386/x86_64 glibc one in config/i386/*.h, check there
> OPTION_GLIBC and only in that case return something other than UNKNOWN.
>
> And redefine TARGET_LIBC_FUNCTION_IMPLEMENTATION only in that case.
>
> Jakub
>
Hi.
I'm sending V2 that can survive bootstrap and regression tests on both x86_64 and ppc64le.
Martin
>From 222c7c205a7afc144dc123d2b378a057dcf8816f Mon Sep 17 00:00:00 2001
From: marxin <mliska@suse.cz>
Date: Wed, 14 Mar 2018 09:44:18 +0100
Subject: [PATCH] Introduce new libc_func_speed target hook (PR
middle-end/81657).
gcc/ChangeLog:
2018-03-14 Martin Liska <mliska@suse.cz>
PR middle-end/81657
* builtins.c (expand_builtin_memory_copy_args): Handle situation
when libc library provides a fast mempcpy implementation/
* config/i386/i386-protos.h (gnu_libc_func_speed): New.
* config/i386/i386.c (enum libc_speed): Likewise.
(ix86_libc_func_speed): Likewise.
(TARGET_LIBC_FUNC_SPEED): Likewise.
* coretypes.h (enum libc_speed): Likewise.
* doc/tm.texi: Document new target hook.
* doc/tm.texi.in: Likewise.
* expr.c (emit_block_move_hints): Handle libc bail out argument.
* expr.h (emit_block_move_hints): Add new parameters.
* target.def: Add new hook.
* targhooks.c (enum libc_speed): New enum.
(default_libc_func_speed): Provide a default hook
implementation.
* targhooks.h (default_libc_func_speed): Likewise.
gcc/testsuite/ChangeLog:
2018-03-14 Martin Liska <mliska@suse.cz>
* gcc.c-torture/execute/builtins/mempcpy.c (main_test): Adjust
to not use mempcpy.
* gcc.dg/string-opt-1.c: Adjust for i386 target.
---
gcc/builtins.c | 13 ++++++++++++-
gcc/config/i386/i386-protos.h | 2 ++
gcc/config/i386/i386.c | 20 ++++++++++++++++++++
gcc/coretypes.h | 7 +++++++
gcc/doc/tm.texi | 4 ++++
gcc/doc/tm.texi.in | 1 +
gcc/expr.c | 16 +++++++++++++++-
gcc/expr.h | 4 +++-
gcc/target.def | 7 +++++++
gcc/targhooks.c | 6 ++++++
gcc/targhooks.h | 1 +
.../gcc.c-torture/execute/builtins/mempcpy.c | 2 +-
gcc/testsuite/gcc.dg/string-opt-1.c | 4 ++--
13 files changed, 81 insertions(+), 6 deletions(-)
diff --git a/gcc/builtins.c b/gcc/builtins.c
index 85affa74510..eb038dd45b3 100644
--- a/gcc/builtins.c
+++ b/gcc/builtins.c
@@ -3651,13 +3651,24 @@ expand_builtin_memory_copy_args (tree dest, tree src, tree len,
src_mem = get_memory_rtx (src, len);
set_mem_align (src_mem, src_align);
+ bool is_move_done;
+
/* Copy word part most expediently. */
+ bool bail_out_libcall = endp == 1
+ && targetm.libc_func_speed ((int)BUILT_IN_MEMPCPY) == FAST_SPEED;
dest_addr = emit_block_move_hints (dest_mem, src_mem, len_rtx,
CALL_EXPR_TAILCALL (exp)
&& (endp == 0 || target == const0_rtx)
? BLOCK_OP_TAILCALL : BLOCK_OP_NORMAL,
expected_align, expected_size,
- min_size, max_size, probable_max_size);
+ min_size, max_size, probable_max_size,
+ bail_out_libcall, &is_move_done);
+
+ /* Bail out when a mempcpy call would be expanded as libcall and when
+ we have a target that provides a fast implementation
+ of mempcpy routine. */
+ if (!is_move_done)
+ return NULL_RTX;
if (dest_addr == 0)
{
diff --git a/gcc/config/i386/i386-protos.h b/gcc/config/i386/i386-protos.h
index ef7c818986f..d3fc515845b 100644
--- a/gcc/config/i386/i386-protos.h
+++ b/gcc/config/i386/i386-protos.h
@@ -47,6 +47,8 @@ extern void ix86_reset_previous_fndecl (void);
extern bool ix86_using_red_zone (void);
+extern enum libc_speed gnu_libc_func_speed (int fn);
+
#ifdef RTX_CODE
extern int standard_80387_constant_p (rtx);
extern const char *standard_80387_constant_opcode (rtx);
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index af24c6ec5ba..b3eb50a42f3 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -2733,6 +2733,23 @@ ix86_using_red_zone (void)
&& (!cfun->machine->has_local_indirect_jump
|| cfun->machine->indirect_branch_type == indirect_branch_keep));
}
+
+enum libc_speed
+ix86_libc_func_speed (int fn)
+{
+ enum built_in_function f = (built_in_function)fn;
+
+ if (!OPTION_GLIBC)
+ return UNKNOWN_SPEED;
+
+ switch (f)
+ {
+ case BUILT_IN_MEMPCPY:
+ return FAST_SPEED;
+ default:
+ return UNKNOWN_SPEED;
+ }
+}
/* Return a string that documents the current -m options. The caller is
responsible for freeing the string. */
@@ -52061,6 +52078,9 @@ ix86_run_selftests (void)
#undef TARGET_WARN_PARAMETER_PASSING_ABI
#define TARGET_WARN_PARAMETER_PASSING_ABI ix86_warn_parameter_passing_abi
+#undef TARGET_LIBC_FUNC_SPEED
+#define TARGET_LIBC_FUNC_SPEED ix86_libc_func_speed
+
#if CHECKING_P
#undef TARGET_RUN_TARGET_SELFTESTS
#define TARGET_RUN_TARGET_SELFTESTS selftest::ix86_run_selftests
diff --git a/gcc/coretypes.h b/gcc/coretypes.h
index 283b4eb33fe..8123df7ccc5 100644
--- a/gcc/coretypes.h
+++ b/gcc/coretypes.h
@@ -384,6 +384,13 @@ enum excess_precision_type
EXCESS_PRECISION_TYPE_FAST
};
+enum libc_speed
+{
+ FAST_SPEED,
+ SLOW_SPEED,
+ UNKNOWN_SPEED
+};
+
/* Support for user-provided GGC and PCH markers. The first parameter
is a pointer to a pointer, the second a cookie. */
typedef void (*gt_pointer_operator) (void *, void *);
diff --git a/gcc/doc/tm.texi b/gcc/doc/tm.texi
index bd8b917ba82..0f7c91a22c4 100644
--- a/gcc/doc/tm.texi
+++ b/gcc/doc/tm.texi
@@ -5501,6 +5501,10 @@ macro, a reasonable default is used.
This hook determines whether a function from a class of functions
@var{fn_class} is present at the runtime.
@end deftypefn
+@deftypefn {Target Hook} libc_speed TARGET_LIBC_FUNC_SPEED (int @var{fn})
+This hook determines whether a function from libc has a fast implementation
+@var{fn} is present at the runtime.
+@end deftypefn
@defmac NEXT_OBJC_RUNTIME
Set this macro to 1 to use the "NeXT" Objective-C message sending conventions
diff --git a/gcc/doc/tm.texi.in b/gcc/doc/tm.texi.in
index b0207146e8c..4bb2998a8a1 100644
--- a/gcc/doc/tm.texi.in
+++ b/gcc/doc/tm.texi.in
@@ -3933,6 +3933,7 @@ macro, a reasonable default is used.
@end defmac
@hook TARGET_LIBC_HAS_FUNCTION
+@hook TARGET_LIBC_FUNC_SPEED
@defmac NEXT_OBJC_RUNTIME
Set this macro to 1 to use the "NeXT" Objective-C message sending conventions
diff --git a/gcc/expr.c b/gcc/expr.c
index 00660293f72..b6c13652d79 100644
--- a/gcc/expr.c
+++ b/gcc/expr.c
@@ -1554,6 +1554,8 @@ compare_by_pieces (rtx arg0, rtx arg1, unsigned HOST_WIDE_INT len,
MIN_SIZE is the minimal size of block to move
MAX_SIZE is the maximal size of block to move, if it can not be represented
in unsigned HOST_WIDE_INT, than it is mask of all ones.
+ If BAIL_OUT_LIBCALL is set true, do not emit library call and set
+ *IS_MOVE_DONE to false.
Return the address of the new block, if memcpy is called and returns it,
0 otherwise. */
@@ -1563,12 +1565,17 @@ emit_block_move_hints (rtx x, rtx y, rtx size, enum block_op_methods method,
unsigned int expected_align, HOST_WIDE_INT expected_size,
unsigned HOST_WIDE_INT min_size,
unsigned HOST_WIDE_INT max_size,
- unsigned HOST_WIDE_INT probable_max_size)
+ unsigned HOST_WIDE_INT probable_max_size,
+ bool bail_out_libcall, bool *is_move_done)
{
bool may_use_call;
rtx retval = 0;
unsigned int align;
+ /* When not doing a bail out, we always emit a memory move. */
+ if (is_move_done)
+ *is_move_done = true;
+
gcc_assert (size);
if (CONST_INT_P (size) && INTVAL (size) == 0)
return 0;
@@ -1625,6 +1632,13 @@ emit_block_move_hints (rtx x, rtx y, rtx size, enum block_op_methods method,
&& ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (x))
&& ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (y)))
{
+ if (bail_out_libcall)
+ {
+ if (is_move_done)
+ *is_move_done = false;
+ return retval;
+ }
+
/* Since x and y are passed to a libcall, mark the corresponding
tree EXPR as addressable. */
tree y_expr = MEM_EXPR (y);
diff --git a/gcc/expr.h b/gcc/expr.h
index b3d523bcb24..023bc5aec47 100644
--- a/gcc/expr.h
+++ b/gcc/expr.h
@@ -110,7 +110,9 @@ extern rtx emit_block_move_hints (rtx, rtx, rtx, enum block_op_methods,
unsigned int, HOST_WIDE_INT,
unsigned HOST_WIDE_INT,
unsigned HOST_WIDE_INT,
- unsigned HOST_WIDE_INT);
+ unsigned HOST_WIDE_INT,
+ bool bail_out_libcall = false,
+ bool *is_move_done = NULL);
extern rtx emit_block_cmp_hints (rtx, rtx, rtx, tree, rtx, bool,
by_pieces_constfn, void *);
extern bool emit_storent_insn (rtx to, rtx from);
diff --git a/gcc/target.def b/gcc/target.def
index c5b2a1e7e71..3bbddc82776 100644
--- a/gcc/target.def
+++ b/gcc/target.def
@@ -2639,6 +2639,13 @@ DEFHOOK
bool, (enum function_class fn_class),
default_libc_has_function)
+DEFHOOK
+(libc_func_speed,
+ "This hook determines whether a function from libc has a fast implementation\n\
+@var{fn} is present at the runtime.",
+ libc_speed, (int fn),
+ default_libc_func_speed)
+
/* True if new jumps cannot be created, to replace existing ones or
not, at the current point in the compilation. */
DEFHOOK
diff --git a/gcc/targhooks.c b/gcc/targhooks.c
index fafcc6c5196..af0c87ff9fe 100644
--- a/gcc/targhooks.c
+++ b/gcc/targhooks.c
@@ -1642,6 +1642,12 @@ no_c99_libc_has_function (enum function_class fn_class ATTRIBUTE_UNUSED)
return false;
}
+enum libc_speed
+default_libc_func_speed (int)
+{
+ return UNKNOWN_SPEED;
+}
+
tree
default_builtin_tm_load_store (tree ARG_UNUSED (type))
{
diff --git a/gcc/targhooks.h b/gcc/targhooks.h
index 8a4393f2ba4..7508673ad0a 100644
--- a/gcc/targhooks.h
+++ b/gcc/targhooks.h
@@ -205,6 +205,7 @@ extern bool default_have_conditional_execution (void);
extern bool default_libc_has_function (enum function_class);
extern bool no_c99_libc_has_function (enum function_class);
extern bool gnu_libc_has_function (enum function_class);
+extern enum libc_speed default_libc_func_speed (int);
extern tree default_builtin_tm_load_store (tree);
diff --git a/gcc/testsuite/gcc.c-torture/execute/builtins/mempcpy.c b/gcc/testsuite/gcc.c-torture/execute/builtins/mempcpy.c
index d82e2232d7b..91e1c87f83f 100644
--- a/gcc/testsuite/gcc.c-torture/execute/builtins/mempcpy.c
+++ b/gcc/testsuite/gcc.c-torture/execute/builtins/mempcpy.c
@@ -62,7 +62,7 @@ main_test (void)
mempcpy (p + 5, s3, 1);
if (memcmp (p, "ABCDEFg", 8))
abort ();
- mempcpy (p + 6, s1 + 1, l1);
+ memcpy (p + 6, s1 + 1, l1);
if (memcmp (p, "ABCDEF2", 8))
abort ();
}
diff --git a/gcc/testsuite/gcc.dg/string-opt-1.c b/gcc/testsuite/gcc.dg/string-opt-1.c
index 2f060732bf0..851c8b04a33 100644
--- a/gcc/testsuite/gcc.dg/string-opt-1.c
+++ b/gcc/testsuite/gcc.dg/string-opt-1.c
@@ -48,5 +48,5 @@ main (void)
return 0;
}
-/* { dg-final { scan-assembler-not "\<mempcpy\>" } } */
-/* { dg-final { scan-assembler "memcpy" } } */
+/* { dg-final { scan-assembler-not "\<mempcpy\>" { target { i?86-*-* x86_64-*-* } } } } */
+/* { dg-final { scan-assembler "memcpy" { target { ! { i?86-*-* x86_64-*-* } } } } } */
--
2.16.2