On 5/10/19 11:21 AM, Jakub Jelinek wrote:
> On Fri, May 10, 2019 at 11:04:12AM +0200, Martin Liška wrote:
>> --- a/gcc/config/i386/i386.h
>> +++ b/gcc/config/i386/i386.h
>> @@ -1906,6 +1906,9 @@ typedef struct ix86_args {
>>
>> #define CLEAR_RATIO(speed) ((speed) ? MIN (6, ix86_cost->move_ratio) : 2)
>>
>> +/* C library provides fast implementation of mempcpy function. */
>> +#define TARGET_HAS_FAST_MEMPCPY_ROUTINE 1
>> +
>
> 1) we shouldn't be adding further target macros, but target hooks
Done.
> 2) I don't think this is a property of the x86 target, but of x86 glibc,
> so you should set it on x86 glibc only (i.e. i?86/x86_64 linux and hurd
> when using glibc, not newlib, nor bionic/android, nor uclibc, nor musl)
I've implemented the in i386.c with DEFAULT_LIBC == LIBC_GLIBC. Hope it's
correct?
>
>> --- a/gcc/testsuite/gcc.c-torture/execute/builtins/mempcpy.c
>> +++ b/gcc/testsuite/gcc.c-torture/execute/builtins/mempcpy.c
>> @@ -56,9 +56,8 @@ main_test (void)
>> if (__builtin_mempcpy (p, "ABCDE", 6) != p + 6 || memcmp (p, "ABCDE", 6))
>> abort ();
>>
>> - /* If the result of mempcpy is ignored, gcc should use memcpy.
>> - This should be optimized always, so set inside_main again. */
>> - inside_main = 1;
>> + /* Set inside main in order to not abort because of usafe of mempcpy. */
>> + inside_main = 0;
>> mempcpy (p + 5, s3, 1);
>> if (memcmp (p, "ABCDEFg", 8))
>> abort ();
>
> Why this? Do you mean mempcpy is called here, even when the lhs is unused?
> We should be calling memcpy in that case.
Yes, that's a stupid mistake. I'm attaching updated version of the patch
that works fine in x86_64-linux-gnu.
Martin
>
> Jakub
>
>From 05b36e6a04e061b46638ec6f23ad2027f033caf9 Mon Sep 17 00:00:00 2001
From: marxin <mli...@suse.cz>
Date: Mon, 29 Apr 2019 13:46:25 +0200
Subject: [PATCH] Come up with has_fast_mempcpy_routine hook (PR
middle-end/90263).
gcc/ChangeLog:
2019-05-13 Martin Liska <mli...@suse.cz>
PR middle-end/90263
* builtins.c (expand_builtin_memory_copy_args): When having a
target with fast mempcpy implementation do now use memcpy.
* config/i386/i386.c (glibc_has_fast_mempcpy_routine): New.
(TARGET_HAS_FAST_MEMPCPY_ROUTINE): Likewise.
* doc/tm.texi: Likewise.
* doc/tm.texi.in: Likewise.
* target.def: Likewise.
* expr.c (emit_block_move_hints): Add 2 new arguments.
* expr.h (emit_block_move_hints): Bail out when libcall
to memcpy would be used.
gcc/testsuite/ChangeLog:
2019-05-13 Martin Liska <mli...@suse.cz>
PR middle-end/90263
* gcc.c-torture/pr90263.c: New test.
---
gcc/builtins.c | 17 +++++++++++++++--
gcc/config/i386/i386.c | 10 ++++++++++
gcc/doc/tm.texi | 4 ++++
gcc/doc/tm.texi.in | 2 ++
gcc/expr.c | 13 ++++++++++++-
gcc/expr.h | 4 +++-
gcc/target.def | 6 ++++++
gcc/testsuite/gcc.c-torture/pr90263.c | 9 +++++++++
8 files changed, 61 insertions(+), 4 deletions(-)
create mode 100644 gcc/testsuite/gcc.c-torture/pr90263.c
diff --git a/gcc/builtins.c b/gcc/builtins.c
index d37d73fc4a0..06b19b62b43 100644
--- a/gcc/builtins.c
+++ b/gcc/builtins.c
@@ -3839,6 +3839,8 @@ expand_builtin_memory_copy_args (tree dest, tree src, tree len,
unsigned HOST_WIDE_INT max_size;
unsigned HOST_WIDE_INT probable_max_size;
+ bool is_move_done;
+
/* If DEST is not a pointer type, call the normal function. */
if (dest_align == 0)
return NULL_RTX;
@@ -3888,11 +3890,22 @@ expand_builtin_memory_copy_args (tree dest, tree src, tree len,
if (CALL_EXPR_TAILCALL (exp)
&& (retmode == RETURN_BEGIN || target == const0_rtx))
method = BLOCK_OP_TAILCALL;
- if (retmode == RETURN_END && target != const0_rtx)
+ bool use_mempcpy_call = (targetm.has_fast_mempcpy_routine ()
+ && retmode == RETURN_END
+ && target != const0_rtx);
+ if (use_mempcpy_call)
method = BLOCK_OP_NO_LIBCALL_RET;
dest_addr = emit_block_move_hints (dest_mem, src_mem, len_rtx, method,
expected_align, expected_size,
- min_size, max_size, probable_max_size);
+ min_size, max_size, probable_max_size,
+ use_mempcpy_call, &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 == pc_rtx)
return NULL_RTX;
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index cc0ae3fcfd3..a0fc9786568 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -23056,6 +23056,16 @@ ix86_run_selftests (void)
#define TARGET_GET_MULTILIB_ABI_NAME \
ix86_get_multilib_abi_name
+#if DEFAULT_LIBC == LIBC_GLIBC
+static bool glibc_has_fast_mempcpy_routine (void)
+{
+ return true;
+}
+
+#undef TARGET_HAS_FAST_MEMPCPY_ROUTINE
+#define TARGET_HAS_FAST_MEMPCPY_ROUTINE glibc_has_fast_mempcpy_routine
+#endif
+
#if CHECKING_P
#undef TARGET_RUN_TARGET_SELFTESTS
#define TARGET_RUN_TARGET_SELFTESTS selftest::ix86_run_selftests
diff --git a/gcc/doc/tm.texi b/gcc/doc/tm.texi
index 8c8978bb13a..dbe1ef7f4a1 100644
--- a/gcc/doc/tm.texi
+++ b/gcc/doc/tm.texi
@@ -1935,6 +1935,10 @@ call_2 may be NULL or a call insn.
This hook returns name of multilib ABI name.
@end deftypefn
+@deftypefn {Target Hook} bool TARGET_HAS_FAST_MEMPCPY_ROUTINE (void)
+Return true if a target has a fast mempcpy routine.
+@end deftypefn
+
@findex fixed_regs
@findex call_used_regs
@findex global_regs
diff --git a/gcc/doc/tm.texi.in b/gcc/doc/tm.texi.in
index fe1194ef91a..a64205c221b 100644
--- a/gcc/doc/tm.texi.in
+++ b/gcc/doc/tm.texi.in
@@ -1713,6 +1713,8 @@ of @code{CALL_USED_REGISTERS}.
@hook TARGET_GET_MULTILIB_ABI_NAME
+@hook TARGET_HAS_FAST_MEMPCPY_ROUTINE
+
@findex fixed_regs
@findex call_used_regs
@findex global_regs
diff --git a/gcc/expr.c b/gcc/expr.c
index fa15b7eceae..c78bc74c0d9 100644
--- a/gcc/expr.c
+++ b/gcc/expr.c
@@ -1561,12 +1561,16 @@ 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)
{
int may_use_call;
rtx retval = 0;
unsigned int align;
+ if (is_move_done)
+ *is_move_done = true;
+
gcc_assert (size);
if (CONST_INT_P (size) && INTVAL (size) == 0)
return 0;
@@ -1628,6 +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;
+ }
+
if (may_use_call < 0)
return pc_rtx;
diff --git a/gcc/expr.h b/gcc/expr.h
index 17c3962436a..6eb70bf12f1 100644
--- a/gcc/expr.h
+++ b/gcc/expr.h
@@ -114,7 +114,9 @@ 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 66cee075018..7bff5cbd313 100644
--- a/gcc/target.def
+++ b/gcc/target.def
@@ -5797,6 +5797,12 @@ DEFHOOK
const char *, (void),
hook_constcharptr_void_null)
+DEFHOOK
+(has_fast_mempcpy_routine,
+ "Return true if a target has a fast mempcpy routine.",
+ bool, (void),
+ hook_bool_void_false)
+
DEFHOOK
(remove_extra_call_preserved_regs,
"This hook removes registers from the set of call-clobbered registers\n\
diff --git a/gcc/testsuite/gcc.c-torture/pr90263.c b/gcc/testsuite/gcc.c-torture/pr90263.c
new file mode 100644
index 00000000000..a348f6265d3
--- /dev/null
+++ b/gcc/testsuite/gcc.c-torture/pr90263.c
@@ -0,0 +1,9 @@
+/* PR middle-end/90263 */
+
+int *f (int *p, int *q, long n)
+{
+ return __builtin_mempcpy (p, q, n);
+}
+
+/* { dg-final { scan-assembler "mempcpy" { target { i?86-*-* x86_64-*-* } } } } */
+/* { dg-final { scan-assembler "memcpy" { target { ! { i?86-*-* x86_64-*-* } } } } } */
--
2.21.0
