Hi,
since some time GCC has BUILT_IN_IROUND{F,,L}, similar to lround() and
llround() but the result is returned as an integer. As there is no
corresponding libc function, this builtin is expanded to lround{f,l}
except when -ffast-math is used, in which case it enables slightly
shorter and faster code to be generated inline. Attached patch enables
this builtin in the gfortran frontend. For the testcase
function my_nint(x)
implicit none
real :: x
integer :: my_nint
my_nint = nint(x)
end function my_nint
compiled with "-O2 -ffast-math" on x86-64, the difference between the
assembler output of trunk and trunk+patch:
--- iround.trunk.O2fastmath.s 2012-03-15 20:12:40.045069324 +0200
+++ iround.s 2012-03-15 20:24:19.501320278 +0200
@@ -12,7 +12,7 @@ my_nint_:
movss .LC0(%rip), %xmm0
orps %xmm1, %xmm0
addss %xmm2, %xmm0
- cvttss2siq %xmm0, %rax
+ cvttss2si %xmm0, %eax
ret
.cfi_endproc
.LFE0:
Comparing the size of the object files, in the patched version the
text section is one byte shorter.
Regtested on x86_64-unknown-linux-gnu, Ok for trunk?
(As an aside, there were some recent problems with __builtin_iround(),
see PR 52592, but they seem to be fixed now so I think this should be
safe)
2012-03-15 Janne Blomqvist <j...@gcc.gnu.org>
* f95-lang.c (gfc_init_builtin_functions): Initialize
BUILT_IN_IROUND.
* mathbuiltins.def: Add IROUND.
* trans-intrinsic.c (build_round_expr): Use BUILT_IN_IROUND if
type size matches.
(gfc_build_intrinsic_lib_fndecls): Build iround functions.
--
Janne Blomqvist
diff --git a/gcc/fortran/f95-lang.c b/gcc/fortran/f95-lang.c
index 05b598f..3f28e67 100644
--- a/gcc/fortran/f95-lang.c
+++ b/gcc/fortran/f95-lang.c
@@ -773,7 +773,11 @@ gfc_init_builtin_functions (void)
gfc_define_builtin ("__builtin_fmodf", mfunc_float[1],
BUILT_IN_FMODF, "fmodf", ATTR_CONST_NOTHROW_LEAF_LIST);
- /* lround{f,,l} and llround{f,,l} */
+ /* iround{f,,l}, lround{f,,l} and llround{f,,l} */
+ ftype = build_function_type_list (integer_type_node,
+ float_type_node, NULL_TREE);
+ gfc_define_builtin("__builtin_iroundf", ftype, BUILT_IN_IROUNDF,
+ "iroundf", ATTR_CONST_NOTHROW_LEAF_LIST);
ftype = build_function_type_list (long_integer_type_node,
float_type_node, NULL_TREE);
gfc_define_builtin ("__builtin_lroundf", ftype, BUILT_IN_LROUNDF,
@@ -783,6 +787,10 @@ gfc_init_builtin_functions (void)
gfc_define_builtin ("__builtin_llroundf", ftype, BUILT_IN_LLROUNDF,
"llroundf", ATTR_CONST_NOTHROW_LEAF_LIST);
+ ftype = build_function_type_list (integer_type_node,
+ double_type_node, NULL_TREE);
+ gfc_define_builtin("__builtin_iround", ftype, BUILT_IN_IROUND,
+ "iround", ATTR_CONST_NOTHROW_LEAF_LIST);
ftype = build_function_type_list (long_integer_type_node,
double_type_node, NULL_TREE);
gfc_define_builtin ("__builtin_lround", ftype, BUILT_IN_LROUND,
@@ -792,6 +800,10 @@ gfc_init_builtin_functions (void)
gfc_define_builtin ("__builtin_llround", ftype, BUILT_IN_LLROUND,
"llround", ATTR_CONST_NOTHROW_LEAF_LIST);
+ ftype = build_function_type_list (integer_type_node,
+ long_double_type_node, NULL_TREE);
+ gfc_define_builtin("__builtin_iroundl", ftype, BUILT_IN_IROUNDL,
+ "iroundl", ATTR_CONST_NOTHROW_LEAF_LIST);
ftype = build_function_type_list (long_integer_type_node,
long_double_type_node, NULL_TREE);
gfc_define_builtin ("__builtin_lroundl", ftype, BUILT_IN_LROUNDL,
diff --git a/gcc/fortran/mathbuiltins.def b/gcc/fortran/mathbuiltins.def
index b0bcc1f..f6d9586 100644
--- a/gcc/fortran/mathbuiltins.def
+++ b/gcc/fortran/mathbuiltins.def
@@ -64,6 +64,7 @@ OTHER_BUILTIN (FMOD, "fmod", 2, true)
OTHER_BUILTIN (FREXP, "frexp", frexp, false)
OTHER_BUILTIN (LLROUND, "llround", llround, true)
OTHER_BUILTIN (LROUND, "lround", lround, true)
+OTHER_BUILTIN (IROUND, "iround", iround, true)
OTHER_BUILTIN (NEXTAFTER, "nextafter", 2, true)
OTHER_BUILTIN (POW, "pow", 1, true)
OTHER_BUILTIN (ROUND, "round", 1, true)
diff --git a/gcc/fortran/trans-intrinsic.c b/gcc/fortran/trans-intrinsic.c
index ac9f507..5e54d8e 100644
--- a/gcc/fortran/trans-intrinsic.c
+++ b/gcc/fortran/trans-intrinsic.c
@@ -376,28 +376,24 @@ build_round_expr (tree arg, tree restype)
{
tree argtype;
tree fn;
- bool longlong;
int argprec, resprec;
argtype = TREE_TYPE (arg);
argprec = TYPE_PRECISION (argtype);
resprec = TYPE_PRECISION (restype);
- /* Depending on the type of the result, choose the long int intrinsic
- (lround family) or long long intrinsic (llround). We might also
- need to convert the result afterwards. */
- if (resprec <= LONG_TYPE_SIZE)
- longlong = false;
+ /* Depending on the type of the result, choose the int intrinsic
+ (iround, available only as a builtin), long int intrinsic (lround
+ family) or long long intrinsic (llround). We might also need to
+ convert the result afterwards. */
+ if (resprec <= INT_TYPE_SIZE)
+ fn = builtin_decl_for_precision (BUILT_IN_IROUND, argprec);
+ else if (resprec <= LONG_TYPE_SIZE)
+ fn = builtin_decl_for_precision (BUILT_IN_LROUND, argprec);
else if (resprec <= LONG_LONG_TYPE_SIZE)
- longlong = true;
- else
- gcc_unreachable ();
-
- /* Now, depending on the argument type, we choose between intrinsics. */
- if (longlong)
fn = builtin_decl_for_precision (BUILT_IN_LLROUND, argprec);
else
- fn = builtin_decl_for_precision (BUILT_IN_LROUND, argprec);
+ gcc_unreachable ();
return fold_convert (restype, build_call_expr_loc (input_location,
fn, 1, arg));
@@ -623,7 +619,7 @@ gfc_build_intrinsic_lib_fndecls (void)
q-suffixed functions. */
tree type, complex_type, func_1, func_2, func_cabs, func_frexp;
- tree func_lround, func_llround, func_scalbn, func_cpow;
+ tree func_iround, func_lround, func_llround, func_scalbn, func_cpow;
memset (quad_decls, 0, sizeof(tree) * (END_BUILTINS + 1));
@@ -631,6 +627,9 @@ gfc_build_intrinsic_lib_fndecls (void)
complex_type = complex_float128_type_node;
/* type (*) (type) */
func_1 = build_function_type_list (type, type, NULL_TREE);
+ /* int (*) (type) */
+ func_iround = build_function_type_list (integer_type_node,
+ type, NULL_TREE);
/* long (*) (type) */
func_lround = build_function_type_list (long_integer_type_node,
type, NULL_TREE);
