This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[RFC] Enable SSE2 math intrinsics for fpmath=sse and fast-math
- From: Richard Guenther <rguenth at tat dot physik dot uni-tuebingen dot de>
- To: gcc-patches at gcc dot gnu dot org
- Date: Wed, 08 Jun 2005 15:14:33 +0200
- Subject: [RFC] Enable SSE2 math intrinsics for fpmath=sse and fast-math
On top of the sseregparm patch this adds support for emitting calls
to Intel libimf.a SSE2 intrinsics. They are enabled if the target
supports SSE2 and unsafe math optimizations are on.
Of course this helps you only, if you have Intel libimf.a around.
And it raises the question of where to put our own SSE2 intrinsics.
There are four options:
1. in libgcc
2. in glibc libm
3. in both
4. in to-be-created gcc libm
Where for option 3 we could play linker tricks to have simple
SSE ABI wrappers to libm in libgcc and have them overridden by
true implementations in libm, if available (does this work?).
I'm open for suggestions, though in the long term I'd like 4 most.
I'm also curious for benchmarks compared to Intel of say, PovRay.
Thanks,
Richard.
2005-06-08 Richard Guenther <rguenth@gcc.gnu.org>
* c-common.c (builtin_function_2): Add *idecl argument
to return 2nd created decl.
(c_common_nodes_and_builtins): Store both, builtin and
implicit decl.
* config/i386/i386.c (ix86_init_builtins): With SSE2
support and fpmath=sse change (some) builtin math
functions to be sseregparm and redirect their assembler_name
to an appropriate library routine.
Index: c-common.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/c-common.c,v
retrieving revision 1.634
diff -c -3 -p -r1.634 c-common.c
*** c-common.c 6 Jun 2005 19:31:24 -0000 1.634
--- c-common.c 8 Jun 2005 12:34:55 -0000
*************** static tree builtin_function_2 (const ch
*** 2499,2505 ****
enum built_in_function function_code,
enum built_in_class cl, int library_name_p,
bool nonansi_p,
! tree attrs);
/* Make a variant type in the proper way for C/C++, propagating qualifiers
down to the element type of an array. */
--- 2502,2508 ----
enum built_in_function function_code,
enum built_in_class cl, int library_name_p,
bool nonansi_p,
! tree attrs, tree *idecl);
/* Make a variant type in the proper way for C/C++, propagating qualifiers
down to the element type of an array. */
*************** c_common_nodes_and_builtins (void)
*** 3236,3249 ****
NONANSI_P, ATTRS, IMPLICIT, COND) \
if (NAME && COND) \
{ \
! tree decl; \
\
gcc_assert ((!BOTH_P && !FALLBACK_P) \
|| !strncmp (NAME, "__builtin_", \
strlen ("__builtin_"))); \
\
if (!BOTH_P) \
! decl = lang_hooks.builtin_function (NAME, builtin_types[TYPE], \
ENUM, \
CLASS, \
(FALLBACK_P \
--- 3239,3253 ----
NONANSI_P, ATTRS, IMPLICIT, COND) \
if (NAME && COND) \
{ \
! tree decl, idecl; \
\
gcc_assert ((!BOTH_P && !FALLBACK_P) \
|| !strncmp (NAME, "__builtin_", \
strlen ("__builtin_"))); \
\
if (!BOTH_P) \
! idecl = decl \
! = lang_hooks.builtin_function (NAME, builtin_types[TYPE], \
ENUM, \
CLASS, \
(FALLBACK_P \
*************** c_common_nodes_and_builtins (void)
*** 3259,3269 ****
CLASS, \
FALLBACK_P, \
NONANSI_P, \
! built_in_attributes[(int) ATTRS]); \
\
built_in_decls[(int) ENUM] = decl; \
if (IMPLICIT) \
! implicit_built_in_decls[(int) ENUM] = decl; \
}
#include "builtins.def"
#undef DEF_BUILTIN
--- 3263,3274 ----
CLASS, \
FALLBACK_P, \
NONANSI_P, \
! built_in_attributes[(int) ATTRS], \
! &idecl); \
\
built_in_decls[(int) ENUM] = decl; \
if (IMPLICIT) \
! implicit_built_in_decls[(int) ENUM] = idecl; \
}
#include "builtins.def"
#undef DEF_BUILTIN
*************** builtin_function_2 (const char *builtin_
*** 3374,3380 ****
tree builtin_type, tree type,
enum built_in_function function_code,
enum built_in_class cl, int library_name_p,
! bool nonansi_p, tree attrs)
{
tree bdecl = NULL_TREE;
tree decl = NULL_TREE;
--- 3379,3385 ----
tree builtin_type, tree type,
enum built_in_function function_code,
enum built_in_class cl, int library_name_p,
! bool nonansi_p, tree attrs, tree *idecl)
{
tree bdecl = NULL_TREE;
tree decl = NULL_TREE;
*************** builtin_function_2 (const char *builtin_
*** 3389,3394 ****
--- 3394,3400 ----
decl = lang_hooks.builtin_function (name, type, function_code, cl,
NULL, attrs);
+ *idecl = decl;
return (bdecl != 0 ? bdecl : decl);
}
Index: config/i386/i386.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/i386/i386.c,v
retrieving revision 1.826
diff -c -3 -p -r1.826 i386.c
*** config/i386/i386.c 8 Jun 2005 05:05:18 -0000 1.826
--- config/i386/i386.c 8 Jun 2005 13:08:39 -0000
*************** ix86_init_builtins (void)
*** 13620,13625 ****
--- 13681,13735 ----
{
if (TARGET_MMX)
ix86_init_mmx_sse_builtins ();
+
+ if (flag_unsafe_math_optimizations
+ && ix86_fpmath == FPMATH_SSE)
+ {
+ /* Patch some builtins with sseregparm attribute and
+ redirect their assembler_name. Note that we need
+ to change assembler_name for both __builtin_NAME
+ and NAME, while the (shared) type needs to be fixed
+ only once. */
+ #define PATCH_SSE_BUILTIN(fncode, symbol) \
+ { \
+ tree sym = get_identifier (symbol); \
+ tree fn = built_in_decls[fncode]; \
+ TYPE_ATTRIBUTES (TREE_TYPE (fn)) \
+ = tree_cons (get_identifier ("sseregparm"), \
+ NULL_TREE, \
+ TYPE_ATTRIBUTES (TREE_TYPE (fn))); \
+ SET_DECL_ASSEMBLER_NAME (fn, sym); \
+ fn = implicit_built_in_decls[fncode]; \
+ if (fn) \
+ SET_DECL_ASSEMBLER_NAME (fn, sym); \
+ }
+
+ PATCH_SSE_BUILTIN (BUILT_IN_ACOS, "__libm_sse2_acos")
+ PATCH_SSE_BUILTIN (BUILT_IN_ACOSF, "__libm_sse2_acosf")
+ PATCH_SSE_BUILTIN (BUILT_IN_ASIN, "__libm_sse2_asin")
+ PATCH_SSE_BUILTIN (BUILT_IN_ASINF, "__libm_sse2_asinf")
+ PATCH_SSE_BUILTIN (BUILT_IN_ATAN2, "__libm_sse2_atan2")
+ PATCH_SSE_BUILTIN (BUILT_IN_ATAN, "__libm_sse2_atan")
+ PATCH_SSE_BUILTIN (BUILT_IN_ATANF, "__libm_sse2_atanf")
+ PATCH_SSE_BUILTIN (BUILT_IN_COS, "__libm_sse2_cos")
+ PATCH_SSE_BUILTIN (BUILT_IN_COSF, "__libm_sse2_cosf")
+ PATCH_SSE_BUILTIN (BUILT_IN_EXP, "__libm_sse2_exp")
+ PATCH_SSE_BUILTIN (BUILT_IN_EXPF, "__libm_sse2_expf")
+ PATCH_SSE_BUILTIN (BUILT_IN_LOG10, "__libm_sse2_log10")
+ PATCH_SSE_BUILTIN (BUILT_IN_LOG10F, "__libm_sse2_log10f")
+ PATCH_SSE_BUILTIN (BUILT_IN_LOG, "__libm_sse2_log")
+ PATCH_SSE_BUILTIN (BUILT_IN_LOGF, "__libm_sse2_logf")
+ PATCH_SSE_BUILTIN (BUILT_IN_POW, "__libm_sse2_pow")
+ PATCH_SSE_BUILTIN (BUILT_IN_POWF, "__libm_sse2_powf")
+ PATCH_SSE_BUILTIN (BUILT_IN_SIN, "__libm_sse2_sin")
+ PATCH_SSE_BUILTIN (BUILT_IN_SINCOS, "__libm_sse2_sincos")
+ PATCH_SSE_BUILTIN (BUILT_IN_SINCOSF, "__libm_sse2_sincosf")
+ PATCH_SSE_BUILTIN (BUILT_IN_SINF, "__libm_sse2_sinf")
+ PATCH_SSE_BUILTIN (BUILT_IN_TAN, "__libm_sse2_tan")
+ PATCH_SSE_BUILTIN (BUILT_IN_TANF, "__libm_sse2_tanf")
+
+ #undef PATCH_SSE_BUILTIN
+ }
}
/* Set up all the MMX/SSE builtins. This is not called if TARGET_MMX