| Message ID | 1473153792-30946-1-git-send-email-james.greenhalgh@arm.com |
|---|---|
| State | New |
| Headers | show |
On Tue, 6 Sep 2016, James Greenhalgh wrote: > In c-family/c-cppbuiltin.c I've updated cpp_iec_559_value such that also > allow setting __GEC_IEC_559 if FLT_EVAL_METHOD=16, and I've updated > c_cpp_builtins to handle the new value, and use the new enum names. I think the special cases in this patch show that the abstractions are wrong. How about instead having more than one target macro / hook. One would indicate that excess precision is used by insn patterns (and be set only for i386 and m68k). Another would indicate the API-level excess precision requested by the target and might take an argument to indicate whether the "fast" or "standard" case is in use (in the x86 case the results would differ depending on the argument, in the ARM case they wouldn't). * If the first hook returns true, excess precision is unpredictable in the "fast" case. Otherwise excess precision is predictable. * For short-circuiting excess_precision_type, maybe an internal setting EXCESS_PRECISION_NONE would make sense, and other settings would be turned into it if permitted by the results of the second hook. (For s390, talk to the maintainers, but I think they really need to eliminate the bogus float_t definition in glibc and the FLT_EVAL_METHOD setting that goes along with it. This is the sort of theoretical ABI change that should be safe in practice.) > + machine_mode float16_type_mode = (FLOATN_TYPE_NODE (0) > + ? TYPE_MODE (FLOATN_TYPE_NODE (0)) > + : VOIDmode); This is obviously wrong. Use float16_type_node. > + ||mode == TYPE_MODE (float_type_node)); Missing space after ||.
On Wed, 7 Sep 2016, Joseph Myers wrote: > How about instead having more than one target macro / hook. One would > indicate that excess precision is used by insn patterns (and be set only > for i386 and m68k). Another would indicate the API-level excess precision Or, maybe there would be a single hook taking a tristate argument. Target hooks need to provide the following information: (a) What excess precision is implicit in the insn patterns (that is, when a value described in middle-end IR as having a particular mode/type may in fact implicitly have extra range and precision because insn patterns produce results with such extra range and precision, not just with the range and precision of the output mode). (b) What excess precision should be added explicitly to the IR by the front end in "fast" mode. (c) What excess precision should be added explicitly to the IR by the front end in "standard" mode. All of these may be represented by FLT_EVAL_METHOD values. In what follows, "none" means no excess precision (whether the value is 0 or 16); "unpredictable" means inherently unpredictable even in the absence of register spills (like -mfpmath=sse+387), so FLT_EVAL_METHOD == -1. In principle there could be cases of predictable excess precision that have to map to -1 because there is no other value they could map to, but in practice I don't expect that to be an issue (given the TS 18661-3 values of FLT_EVAL_METHOD being available). (a) should always be "none" except for the existing x86 and m68k cases. (b) is "none" in all existing cases, but we have the issue of ARM cases without direct binary16 arithmetic where it would be desirable for it to apply excess precision to _Float16 values. (c) is not "none" at present in exactly those cases where TARGET_FLT_EVAL_METHOD is nonzero (but in the case of s390 this is really a target bug that should be fixed). It might also be not "none" in future for ARM cases like in (b). (a) can be "unpredictable". (b) and (c) never can. Rather than init_excess_precision setting flag_excess_precision, possibly turning "standard" into "fast", I think it should set some variable that describes the result of whichever of (b) and (c) is applicable - and in the cases where "standard" turns into "fast", it would simply happen that both (b) and (c) produce the same result. The effective excess precision seen by the user is the result of applying first one of (b) and (c), then (a). If (a) is not "none", this is not entirely predictable. It's a broken compiler configuration if applying (c) yields a type on which (a) is not a no-op, except in the case where (a) is "unpredictable" and (c) is "none". I'm not aware of any likely cases where a type would actually get promoted twice by applying those two operations. Right now TARGET_FLT_EVAL_METHOD_NON_DEFAULT is used to give errors for -fexcess-precision=standard for languages not supporting it. With a conversion to hooks, that needs to be rethought. The point is to give an error if predictability was requested but cannot be achieved, so I suppose ideally the error should be about (a) being not "none", together with -fexcess-precision=standard being used. But if the relevant back-end options aren't available at this point to use the hook for (a), the error could just be given for all targets (for those languages when that option is given). Effective predictability, for __GCC_IEC_559 in flag_iso mode, means that (a) does nothing to any type resulting from whichever of (b) and (c) is in effect. The way __LIBGCC_*_EXCESS_PRECISION__ is used is about eliminating excess precision from results assigned to variables - meaning it should be about (a) only. That leaves the question of setting FLT_EVAL_METHOD. It should relate to the effective excess precision seen by the user, the combination of whichever of (b) and (c) is in effect with (a). The only problem is the case where that combination is most precisely described by "16", which as discussed is not a C11 value and may affect existing code not expecting such a value. The value -1 is compatible with C11 and TS 18661-3 but suboptimal, while the value 0 is compatible with C11 only, not with TS 18661-3 even when no feature test macros are defined. We already have the option -fno-fp-int-builtin-inexact to ensure certain built-in functions follow TS 18661-1 semantics. It might be reasonable to have a new option to enable FLT_EVAL_METHOD using new values. However, I'd be inclined to think that such an option should be on by default for -std=gnu*, only off for strict conformance modes. (There would be both __FLT_EVAL_METHOD__ and __FLT_EVAL_METHOD_C99__, say, predefined macros, so that <float.h> could also always use the new value if __STDC_WANT_IEC_60559_TYPES_EXT__ is defined.)
diff --git a/gcc/c-family/c-cppbuiltin.c b/gcc/c-family/c-cppbuiltin.c index ee4d233..f278aff 100644 --- a/gcc/c-family/c-cppbuiltin.c +++ b/gcc/c-family/c-cppbuiltin.c @@ -735,10 +735,13 @@ cpp_iec_559_value (void) excess precision to mean lack of IEEE 754 support. The same applies to unpredictable contraction. For C++, and outside strict conformance mode, do not consider these options to mean - lack of IEEE 754 support. */ + lack of IEEE 754 support. FLT_EVAL_METHOD_PROMOTE_TO_FLOAT and + FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16 both give predictable excess + precision. */ if (flag_iso && !c_dialect_cxx () - && TARGET_FLT_EVAL_METHOD != 0 + && TARGET_FLT_EVAL_METHOD != FLT_EVAL_METHOD_PROMOTE_TO_FLOAT + && TARGET_FLT_EVAL_METHOD != FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16 && flag_excess_precision_cmdline != EXCESS_PRECISION_STANDARD) ret = 0; if (flag_iso @@ -1118,24 +1121,28 @@ c_cpp_builtins (cpp_reader *pfile) } builtin_define_with_value (macro_name, suffix, 0); bool excess_precision = false; - if (TARGET_FLT_EVAL_METHOD != 0 - && mode != TYPE_MODE (long_double_type_node) - && (mode == TYPE_MODE (float_type_node) - || mode == TYPE_MODE (double_type_node))) - switch (TARGET_FLT_EVAL_METHOD) - { - case -1: - case 2: - excess_precision = true; - break; - - case 1: - excess_precision = mode == TYPE_MODE (float_type_node); - break; - - default: - gcc_unreachable (); - } + machine_mode float16_type_mode = (FLOATN_TYPE_NODE (0) + ? TYPE_MODE (FLOATN_TYPE_NODE (0)) + : VOIDmode); + switch (TARGET_FLT_EVAL_METHOD) + { + case FLT_EVAL_METHOD_UNPREDICTABLE: + case FLT_EVAL_METHOD_PROMOTE_TO_LONG_DOUBLE: + excess_precision = (mode == float16_type_mode + || mode == TYPE_MODE (float_type_node) + || mode == TYPE_MODE (double_type_node)); + break; + + case FLT_EVAL_METHOD_PROMOTE_TO_DOUBLE: + excess_precision = (mode == float16_type_mode + ||mode == TYPE_MODE (float_type_node)); + break; + case FLT_EVAL_METHOD_PROMOTE_TO_FLOAT: + excess_precision = mode == float16_type_mode; + break; + default: + gcc_unreachable (); + } macro_name = (char *) alloca (strlen (name) + sizeof ("__LIBGCC__EXCESS_" "PRECISION__")); diff --git a/gcc/flag-types.h b/gcc/flag-types.h index dd57e16..adfe074 100644 --- a/gcc/flag-types.h +++ b/gcc/flag-types.h @@ -158,6 +158,16 @@ enum excess_precision EXCESS_PRECISION_STANDARD }; +/* The possible values for FLT_EVAL_METHOD. */ +enum flt_eval_method +{ + FLT_EVAL_METHOD_UNPREDICTABLE = -1, + FLT_EVAL_METHOD_PROMOTE_TO_FLOAT = 0, + FLT_EVAL_METHOD_PROMOTE_TO_DOUBLE = 1, + FLT_EVAL_METHOD_PROMOTE_TO_LONG_DOUBLE = 2, + FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16 = 16 +}; + /* Type of stack check. */ enum stack_check_type { diff --git a/gcc/toplev.c b/gcc/toplev.c index 66099ec..052f414 100644 --- a/gcc/toplev.c +++ b/gcc/toplev.c @@ -1700,15 +1700,23 @@ init_excess_precision (void) int flt_eval_method = TARGET_FLT_EVAL_METHOD; switch (flt_eval_method) { - case -1: - case 0: - /* Either the target acts unpredictably (-1) or has all the - operations required not to have excess precision (0). */ + case FLT_EVAL_METHOD_UNPREDICTABLE: + case FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16: + /* Either the target acts unpredictably (-1) or has the required + operations for any type we support. */ flag_excess_precision = EXCESS_PRECISION_FAST; break; - case 1: - case 2: - /* In these cases, predictable excess precision makes + case FLT_EVAL_METHOD_PROMOTE_TO_FLOAT: + /* Most targets will be in this case (FLT_EVAL_METHOD = 0). + If the target supports _Float16, we need to predictably + calculate it in the precision of float, otherwise, we can + use EXCESS_PRECISION_FAST. */ + if (targetm.floatn_mode (16, false) == VOIDmode) + flag_excess_precision = EXCESS_PRECISION_FAST; + break; + case FLT_EVAL_METHOD_PROMOTE_TO_DOUBLE: + case FLT_EVAL_METHOD_PROMOTE_TO_LONG_DOUBLE: + /* In these cases, predictable excess precision always makes sense. */ break; default: diff --git a/gcc/tree.c b/gcc/tree.c index 33e6f97..cbab851 100644 --- a/gcc/tree.c +++ b/gcc/tree.c @@ -8836,50 +8836,93 @@ build_complex_type (tree component_type) tree excess_precision_type (tree type) { + int flt_eval_method = TARGET_FLT_EVAL_METHOD; + machine_mode float16_type_mode = (float16_type_node + ? TYPE_MODE (float16_type_node) + : VOIDmode); + machine_mode float_type_mode = TYPE_MODE (float_type_node); + machine_mode double_type_mode = TYPE_MODE (float_type_node); if (flag_excess_precision != EXCESS_PRECISION_FAST) { - int flt_eval_method = TARGET_FLT_EVAL_METHOD; switch (TREE_CODE (type)) { case REAL_TYPE: - switch (flt_eval_method) { - case 1: - if (TYPE_MODE (type) == TYPE_MODE (float_type_node)) - return double_type_node; - break; - case 2: - if (TYPE_MODE (type) == TYPE_MODE (float_type_node) - || TYPE_MODE (type) == TYPE_MODE (double_type_node)) - return long_double_type_node; + machine_mode type_mode = TYPE_MODE (type); + switch (flt_eval_method) + { + case FLT_EVAL_METHOD_PROMOTE_TO_FLOAT: + if (type_mode == float16_type_mode) + return float_type_node; + break; + case FLT_EVAL_METHOD_PROMOTE_TO_DOUBLE: + if (type_mode == float16_type_mode + || type_mode == float_type_mode) + return double_type_node; + break; + case FLT_EVAL_METHOD_PROMOTE_TO_LONG_DOUBLE: + if (type_mode == float16_type_mode + || type_mode == float_type_mode + || type_mode == double_type_mode) + return long_double_type_node; + break; + default: + gcc_unreachable (); + } break; - default: - gcc_unreachable (); } - break; case COMPLEX_TYPE: - if (TREE_CODE (TREE_TYPE (type)) != REAL_TYPE) - return NULL_TREE; - switch (flt_eval_method) { - case 1: - if (TYPE_MODE (TREE_TYPE (type)) == TYPE_MODE (float_type_node)) - return complex_double_type_node; - break; - case 2: - if (TYPE_MODE (TREE_TYPE (type)) == TYPE_MODE (float_type_node) - || (TYPE_MODE (TREE_TYPE (type)) - == TYPE_MODE (double_type_node))) - return complex_long_double_type_node; + if (TREE_CODE (TREE_TYPE (type)) != REAL_TYPE) + return NULL_TREE; + machine_mode type_mode = TYPE_MODE (TREE_TYPE (type)); + switch (flt_eval_method) + { + case FLT_EVAL_METHOD_PROMOTE_TO_FLOAT: + if (type_mode == float16_type_mode) + return complex_float_type_node; + break; + case FLT_EVAL_METHOD_PROMOTE_TO_DOUBLE: + if (type_mode == float16_type_mode + || type_mode == float_type_mode) + return complex_double_type_node; + break; + case FLT_EVAL_METHOD_PROMOTE_TO_LONG_DOUBLE: + if (type_mode == float16_type_mode + || type_mode == float_type_mode + || type_mode == double_type_mode) + return complex_long_double_type_node; + break; + default: + gcc_unreachable (); + } break; - default: - gcc_unreachable (); } - break; default: break; } } + else + { + /* A special case for EXCESS_PRECISION_FAST if we need to calculate in + at-least-float precision. If this is a _Float16 + type, and we would calculate it in float precision + (FLT_EVAL_METHOD == FLT_EVAL_METHOD_AT_LEAST_FLOAT), then the target + probably doesn't have HFmode operations. We are therefore + likely to generate better code by promoting once here. If we + don't, we'll promote before arithmetic operations so we can use + their SFmode counterparts. */ + if (flt_eval_method == FLT_EVAL_METHOD_PROMOTE_TO_FLOAT + && float16_type_node) + { + if (TREE_CODE (type) == REAL_TYPE + && TYPE_MODE (type) == float16_type_mode) + return float_type_node; + if (TREE_CODE (type) == COMPLEX_TYPE + && TYPE_MODE (TREE_TYPE (type)) == float16_type_mode) + return complex_float_type_node; + } + } return NULL_TREE; }
Hi, ISO/IEC TS 18661-3 redefines FLT_EVAL_METHOD. In this patch I'm interested in updating the code in GCC to handle two of the changes. The redefinition of the meanings of {0, 1, 2}, and the meaning of "N" for the specific case of N=16. In ISO/IEC TS 18661-3, these are defined as so: 0 Evaluate all operations and constants, whose semantic type has at most the range and precision of float, to the range and precision of float; evaluate all other operations and constants to the range and precision of the semantic type. 1 Evaluate all operations and constants, whose semantic type has at most the range and precision of double, to the range and precision of double; evaluate all other operations and constants to the range and precision of the semantic type. 2 Evaluate all operations and constants, whose semantic type has at most the range and precision of long double, to the range and precision of long double; evaluate all other operations and constants to the range and precision of the semantic type. N, where _FloatN is a supported interchange floating type Evaluate operations and constants whose semantic type has at most the range and precision of the _FloatN type, to the range and precision of the _FloatN type; evaluate all other operations and constants to the range and precision of the semantic type; When we enable support for _Float16 in AArch64 we would like, where we have the ARMv8.2-A 16-bit floating point extensions available, to evaluate operations to the range and precision of that type. Where we do not have the extensions available, we'd like to evaluate operations to the range and precision of float. This will require adding support for FLT_EVAL_METHOD=16. In most cases this is simply duplicating the exceptions that exist for FLT_EVAL_METHOD==0 and teaching some of the excess-precision logic that it may need to handle _Float16 types if FLT_EVAL_METHOD=0. In this patch, I first add a new enum for the 5 float methods, just to save the repeated magic numbers now we are up to 5 of them. I've added this to flag-types.h, but there may be a better home for it. In c-family/c-cppbuiltin.c I've updated cpp_iec_559_value such that also allow setting __GEC_IEC_559 if FLT_EVAL_METHOD=16, and I've updated c_cpp_builtins to handle the new value, and use the new enum names. Then I've updated init_excess_precision in toplev.c with the new enum names, and with logic to understand that for targets that provide _Float16, we can't set flag_excess_precision to "fast" when FLT_EVAL_METHOD=0. This is because FLT_EVAL_METHOD=0 requires predictably using the precision of float for _Float16 types. In tree.c, I've modified excess_precision_type with the logic discussed above, promoting _Float16 to the appropriate type if we are required to. I've also added a special case that allows promoting to "float" from an _Float16 type in the case that -fexcess-precision=fast. If we don't do this, then the "fast" case will spend more time promoting and demoting between HFmode and SFmode and the consequence will be slower code. Bootstrapped on AArch64 and x86_64. OK? Thanks, James --- gcc/ 2016-09-06 James Greenhalgh <james.greenhalgh@arm.com> * flag-types.h (flt_eval_method): New. * toplev.c (init_excess_precision): Handle FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16. * tree.c (excess_precision_type): Likewise. gcc/c-family/ 2016-09-06 James Greenhalgh <james.greenhalgh@arm.com> * c-cppbuiltin.c (cpp_iec_559_value): Support FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16. (c_cpp_builtins): Likewise.