On 5/14/19 5:07 PM, Martin Sebor wrote: > On 5/14/19 8:55 AM, Martin Liška wrote: >> On 5/13/19 3:07 PM, Jakub Jelinek wrote: >>> On Mon, May 13, 2019 at 12:14:37PM +0200, Martin Liška wrote: >>>> 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? >>> >>> No, that would be correct only in the rare SINGLE_LIBC configurations. >>> Either you can do >>> #ifdef OPTION_GLIBC >>> return OPTION_GLIBC; >>> #else >>> return false; >>> #endif >>> or define something in config/linux.h (or .[ch]) that you can then use in >>> i386.c. >>> >>> Jakub >>> >> >> Hi. >> >> You always have nice ideas. I'm sending updated patch which addresses both >> Jakub's >> and Wilco's comments. >> > > 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) > + > > Not to be too nit-picky about the name but target.def refers to > functions rather than routines. It also defines a hook called > libc_has_function with the obvious semantics so if there's > a chance that it could be useful to query whether another libc > function is "fast" I would suggest to consider defining the hook > correspondingly, i.e., > > bool libc_has_fast_function (enum function_class) > > and naming the macro similarly. > > Martin
Hi Martin. That's a very good suggestion and I'm implementing that! Patch can bootstrap on x86_64-linux-gnu and survives regression tests. Thanks, Martin
>From f865ba74008acc651e9ef2fbfc9f2d4bb9ce8fb8 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 hook libc_has_fast_function (PR middle-end/90263). gcc/ChangeLog: 2019-05-15 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 (ix86_libc_has_fast_function): New. (TARGET_LIBC_HAS_FAST_FUNCTION): Likewise. * doc/tm.texi: Likewise. * doc/tm.texi.in: Likewise. * target.def: * 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: 2019-05-15 Martin Liska <mli...@suse.cz> PR middle-end/90263 * gcc.c-torture/compile/pr90263.c: New test. * lib/target-supports.exp: Add check_effective_target_glibc. --- gcc/builtins.c | 17 +++++++++++++++-- gcc/config/i386/i386.c | 15 +++++++++++++++ gcc/doc/tm.texi | 5 +++++ gcc/doc/tm.texi.in | 2 ++ gcc/expr.c | 13 ++++++++++++- gcc/expr.h | 4 +++- gcc/target.def | 7 +++++++ gcc/testsuite/gcc.c-torture/compile/pr90263.c | 10 ++++++++++ gcc/testsuite/lib/target-supports.exp | 11 +++++++++++ 9 files changed, 80 insertions(+), 4 deletions(-) create mode 100644 gcc/testsuite/gcc.c-torture/compile/pr90263.c diff --git a/gcc/builtins.c b/gcc/builtins.c index d37d73fc4a0..70d44927052 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.libc_has_fast_function (BUILT_IN_MEMPCPY) + && 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..6fbfe5f53e4 100644 --- a/gcc/config/i386/i386.c +++ b/gcc/config/i386/i386.c @@ -23056,6 +23056,21 @@ ix86_run_selftests (void) #define TARGET_GET_MULTILIB_ABI_NAME \ ix86_get_multilib_abi_name +static bool ix86_libc_has_fast_function (int fcode) +{ +#ifdef OPTION_GLIBC + if (OPTION_GLIBC) + return (built_in_function)fcode == BUILT_IN_MEMPCPY; + else + return false; +#else + return false; +#endif +} + +#undef TARGET_LIBC_HAS_FAST_FUNCTION +#define TARGET_LIBC_HAS_FAST_FUNCTION ix86_libc_has_fast_function + #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..0941039536b 100644 --- a/gcc/doc/tm.texi +++ b/gcc/doc/tm.texi @@ -5594,6 +5594,11 @@ This hook determines whether a function from a class of functions @var{fn_class} is present at the runtime. @end deftypefn +@deftypefn {Target Hook} bool TARGET_LIBC_HAS_FAST_FUNCTION (int @var{fcode}) +This hook determines whether a function from a class of functions +@var{fn_class} has a fast implementation. +@end deftypefn + @defmac NEXT_OBJC_RUNTIME Set this macro to 1 to use the "NeXT" Objective-C message sending conventions by default. This calling convention involves passing the object, the selector diff --git a/gcc/doc/tm.texi.in b/gcc/doc/tm.texi.in index fe1194ef91a..17560fce6b7 100644 --- a/gcc/doc/tm.texi.in +++ b/gcc/doc/tm.texi.in @@ -3991,6 +3991,8 @@ macro, a reasonable default is used. @hook TARGET_LIBC_HAS_FUNCTION +@hook TARGET_LIBC_HAS_FAST_FUNCTION + @defmac NEXT_OBJC_RUNTIME Set this macro to 1 to use the "NeXT" Objective-C message sending conventions by default. This calling convention involves passing the object, the selector 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..23e260cb535 100644 --- a/gcc/target.def +++ b/gcc/target.def @@ -2547,6 +2547,13 @@ DEFHOOK bool, (enum function_class fn_class), default_libc_has_function) +DEFHOOK +(libc_has_fast_function, + "This hook determines whether a function from a class of functions\n\ +@var{fn_class} has a fast implementation.", + bool, (int fcode), + default_libc_has_fast_function) + /* 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/testsuite/gcc.c-torture/compile/pr90263.c b/gcc/testsuite/gcc.c-torture/compile/pr90263.c new file mode 100644 index 00000000000..df3ab0fc1cd --- /dev/null +++ b/gcc/testsuite/gcc.c-torture/compile/pr90263.c @@ -0,0 +1,10 @@ +/* PR middle-end/90263 */ +/* { dg-require-effective-target glibc } */ + +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-*-* } } } } } */ diff --git a/gcc/testsuite/lib/target-supports.exp b/gcc/testsuite/lib/target-supports.exp index da132369a4b..2f6f990d285 100644 --- a/gcc/testsuite/lib/target-supports.exp +++ b/gcc/testsuite/lib/target-supports.exp @@ -9300,6 +9300,17 @@ proc check_effective_target_arm_v8_3a_complex_neon_hw { } { } [add_options_for_arm_v8_3a_complex_neon ""]] } +# Returns 1 if the target is using glibc, 0 otherwise. + +proc check_effective_target_glibc { } { + return [check_no_compiler_messages glibc_object assembly { + #include <stdlib.h> + #if !defined(__GLIBC__) + #error undefined + #endif + }] +} + # Return 1 if the target plus current options supports a vector # complex addition with rotate of half and single float modes, 0 otherwise. # -- 2.21.0