From f19a327077ecc34a51487761378b9edb43c82997 Mon Sep 17 00:00:00 2001 From: liuhongt Date: Mon, 2 Aug 2021 10:56:45 +0800 Subject: [PATCH] Support -fexcess-precision=16 which will enable FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16 when backend supports _Float16. gcc/ada/ChangeLog: * gcc-interface/misc.c (gnat_post_options): Issue an error for -fexcess-precision=16. gcc/c-family/ChangeLog: * c-common.c (excess_precision_mode_join): Update below comments. (c_ts18661_flt_eval_method): Set excess_precision_type to EXCESS_PRECISION_TYPE_FLOAT16 when -fexcess-precision=16. * c-cppbuiltin.c (cpp_atomic_builtins): Update below comments. (c_cpp_flt_eval_method_iec_559): Set excess_precision_type to EXCESS_PRECISION_TYPE_FLOAT16 when -fexcess-precision=16. gcc/ChangeLog: * common.opt: Support -fexcess-precision=16. * config/aarch64/aarch64.c (aarch64_excess_precision): Return FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16 when EXCESS_PRECISION_TYPE_FLOAT16. * config/arm/arm.c (arm_excess_precision): Ditto. * config/i386/i386.c (ix86_get_excess_precision): Ditto. * config/m68k/m68k.c (m68k_excess_precision): Issue an error when EXCESS_PRECISION_TYPE_FLOAT16. * config/s390/s390.c (s390_excess_precision): Ditto. * coretypes.h (enum excess_precision_type): Add EXCESS_PRECISION_TYPE_FLOAT16. * doc/tm.texi (TARGET_C_EXCESS_PRECISION): Update documents. * doc/tm.texi.in (TARGET_C_EXCESS_PRECISION): Ditto. * doc/extend.texi (Half-Precision): Document -fexcess-precision=16. * flag-types.h (enum excess_precision): Add EXCESS_PRECISION_FLOAT16. * target.def (excess_precision): Update document. * tree.c (excess_precision_type): Set excess_precision_type to EXCESS_PRECISION_FLOAT16 when -fexcess-precision=16. gcc/fortran/ChangeLog: * options.c (gfc_post_options): Issue an error for -fexcess-precision=16. gcc/testsuite/ChangeLog: * gcc.target/i386/float16-6.c: New test. * gcc.target/i386/float16-7.c: New test. --- gcc/ada/gcc-interface/misc.c | 3 +++ gcc/c-family/c-common.c | 6 ++++-- gcc/c-family/c-cppbuiltin.c | 6 ++++-- gcc/common.opt | 5 ++++- gcc/config/aarch64/aarch64.c | 1 + gcc/config/arm/arm.c | 1 + gcc/config/i386/i386.c | 5 +++++ gcc/config/m68k/m68k.c | 3 +++ gcc/config/s390/s390.c | 3 +++ gcc/coretypes.h | 3 ++- gcc/doc/extend.texi | 3 ++- gcc/doc/tm.texi | 14 ++++++++++---- gcc/doc/tm.texi.in | 3 +++ gcc/flag-types.h | 3 ++- gcc/fortran/options.c | 3 +++ gcc/target.def | 11 +++++++---- gcc/testsuite/gcc.target/i386/float16-6.c | 8 ++++++++ gcc/testsuite/gcc.target/i386/float16-7.c | 9 +++++++++ gcc/tree.c | 3 ++- 19 files changed, 76 insertions(+), 17 deletions(-) create mode 100644 gcc/testsuite/gcc.target/i386/float16-6.c create mode 100644 gcc/testsuite/gcc.target/i386/float16-7.c diff --git a/gcc/ada/gcc-interface/misc.c b/gcc/ada/gcc-interface/misc.c index 186367ac6d1..96199bd4b63 100644 --- a/gcc/ada/gcc-interface/misc.c +++ b/gcc/ada/gcc-interface/misc.c @@ -256,6 +256,9 @@ gnat_post_options (const char **pfilename ATTRIBUTE_UNUSED) /* Excess precision other than "fast" requires front-end support. */ if (flag_excess_precision == EXCESS_PRECISION_STANDARD) sorry ("%<-fexcess-precision=standard%> for Ada"); + else if (flag_excess_precision == EXCESS_PRECISION_FLOAT16) + sorry ("%<-fexcess-precision=16%> for Ada"); + flag_excess_precision = EXCESS_PRECISION_FAST; /* No psABI change warnings for Ada. */ diff --git a/gcc/c-family/c-common.c b/gcc/c-family/c-common.c index 017e41537ac..c6757f093ac 100644 --- a/gcc/c-family/c-common.c +++ b/gcc/c-family/c-common.c @@ -8778,7 +8778,7 @@ excess_precision_mode_join (enum flt_eval_method x, This relates to the effective excess precision seen by the user, which is the join point of the precision the target requests for - -fexcess-precision={standard,fast} and the implicit excess precision + -fexcess-precision={standard,fast,16} and the implicit excess precision the target uses. */ static enum flt_eval_method @@ -8790,7 +8790,9 @@ c_ts18661_flt_eval_method (void) enum excess_precision_type flag_type = (flag_excess_precision == EXCESS_PRECISION_STANDARD ? EXCESS_PRECISION_TYPE_STANDARD - : EXCESS_PRECISION_TYPE_FAST); + : (flag_excess_precision == EXCESS_PRECISION_FLOAT16 + ? EXCESS_PRECISION_TYPE_FLOAT16 + : EXCESS_PRECISION_TYPE_FAST)); enum flt_eval_method requested = targetm.c.excess_precision (flag_type); diff --git a/gcc/c-family/c-cppbuiltin.c b/gcc/c-family/c-cppbuiltin.c index 3fa62bc4fe7..48cbefd8bf8 100644 --- a/gcc/c-family/c-cppbuiltin.c +++ b/gcc/c-family/c-cppbuiltin.c @@ -753,7 +753,7 @@ cpp_atomic_builtins (cpp_reader *pfile) /* Return TRUE if the implicit excess precision in which the back-end will compute floating-point calculations is not more than the explicit excess precision that the front-end will apply under - -fexcess-precision=[standard|fast]. + -fexcess-precision=[standard|fast|16]. More intuitively, return TRUE if the excess precision proposed by the front-end is the excess precision that will actually be used. */ @@ -764,7 +764,9 @@ c_cpp_flt_eval_method_iec_559 (void) enum excess_precision_type front_end_ept = (flag_excess_precision == EXCESS_PRECISION_STANDARD ? EXCESS_PRECISION_TYPE_STANDARD - : EXCESS_PRECISION_TYPE_FAST); + : (flag_excess_precision == EXCESS_PRECISION_FLOAT16 + ? EXCESS_PRECISION_TYPE_FLOAT16 + : EXCESS_PRECISION_TYPE_FAST)); enum flt_eval_method back_end = targetm.c.excess_precision (EXCESS_PRECISION_TYPE_IMPLICIT); diff --git a/gcc/common.opt b/gcc/common.opt index 7d69ab5ef7c..f103a7de004 100644 --- a/gcc/common.opt +++ b/gcc/common.opt @@ -1518,7 +1518,7 @@ Perform a number of minor, expensive optimizations. fexcess-precision= Common Joined RejectNegative Enum(excess_precision) Var(flag_excess_precision) Init(EXCESS_PRECISION_DEFAULT) Optimization SetByCombined --fexcess-precision=[fast|standard] Specify handling of excess floating-point precision. +-fexcess-precision=[fast|standard|16] Specify handling of excess floating-point precision. Enum Name(excess_precision) Type(enum excess_precision) UnknownError(unknown excess precision style %qs) @@ -1529,6 +1529,9 @@ Enum(excess_precision) String(fast) Value(EXCESS_PRECISION_FAST) EnumValue Enum(excess_precision) String(standard) Value(EXCESS_PRECISION_STANDARD) +EnumValue +Enum(excess_precision) String(16) Value(EXCESS_PRECISION_FLOAT16) + ; Whether we permit the extended set of values for FLT_EVAL_METHOD ; introduced in ISO/IEC TS 18661-3, or limit ourselves to those in C99/C11. fpermitted-flt-eval-methods= diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c index 26d59ba1e13..1fbe9e0daa0 100644 --- a/gcc/config/aarch64/aarch64.c +++ b/gcc/config/aarch64/aarch64.c @@ -25045,6 +25045,7 @@ aarch64_excess_precision (enum excess_precision_type type) ? FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16 : FLT_EVAL_METHOD_PROMOTE_TO_FLOAT); case EXCESS_PRECISION_TYPE_IMPLICIT: + case EXCESS_PRECISION_TYPE_FLOAT16: return FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16; default: gcc_unreachable (); diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c index 5c929417f93..f1e628253d0 100644 --- a/gcc/config/arm/arm.c +++ b/gcc/config/arm/arm.c @@ -25612,6 +25612,7 @@ arm_excess_precision (enum excess_precision_type type) ? FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16 : FLT_EVAL_METHOD_PROMOTE_TO_FLOAT); case EXCESS_PRECISION_TYPE_IMPLICIT: + case EXCESS_PRECISION_TYPE_FLOAT16: return FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16; default: gcc_unreachable (); diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c index bfefbd718bb..210fc4234a3 100644 --- a/gcc/config/i386/i386.c +++ b/gcc/config/i386/i386.c @@ -23599,6 +23599,11 @@ ix86_get_excess_precision (enum excess_precision_type type) return (type == EXCESS_PRECISION_TYPE_STANDARD ? FLT_EVAL_METHOD_PROMOTE_TO_FLOAT : FLT_EVAL_METHOD_UNPREDICTABLE); + case EXCESS_PRECISION_TYPE_FLOAT16: + if (TARGET_80387 + && !(TARGET_SSE_MATH && TARGET_SSE)) + error ("%<-fexcess-precision=16%> is not compatible with %<-mfpmath=387%>"); + return FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16; default: gcc_unreachable (); } diff --git a/gcc/config/m68k/m68k.c b/gcc/config/m68k/m68k.c index 3f63c60fa92..0248eb719a8 100644 --- a/gcc/config/m68k/m68k.c +++ b/gcc/config/m68k/m68k.c @@ -7115,6 +7115,9 @@ m68k_excess_precision (enum excess_precision_type type) return FLT_EVAL_METHOD_PROMOTE_TO_FLOAT; return FLT_EVAL_METHOD_PROMOTE_TO_LONG_DOUBLE; + case EXCESS_PRECISION_TYPE_FLOAT16: + error ("%<-fexcess-precision=16%> is not supported on this target"); + break; default: gcc_unreachable (); } diff --git a/gcc/config/s390/s390.c b/gcc/config/s390/s390.c index 673a1340285..54dd6332c3a 100644 --- a/gcc/config/s390/s390.c +++ b/gcc/config/s390/s390.c @@ -16549,6 +16549,9 @@ s390_excess_precision (enum excess_precision_type type) ensure consistency with the implementation in glibc, report that float is evaluated to the range and precision of double. */ return FLT_EVAL_METHOD_PROMOTE_TO_DOUBLE; + case EXCESS_PRECISION_TYPE_FLOAT16: + error ("%<-fexcess-precision=16%> is not supported on this target"); + break; default: gcc_unreachable (); } diff --git a/gcc/coretypes.h b/gcc/coretypes.h index 726fcaddda2..b4f530d57ac 100644 --- a/gcc/coretypes.h +++ b/gcc/coretypes.h @@ -425,7 +425,8 @@ enum excess_precision_type { EXCESS_PRECISION_TYPE_IMPLICIT, EXCESS_PRECISION_TYPE_STANDARD, - EXCESS_PRECISION_TYPE_FAST + EXCESS_PRECISION_TYPE_FAST, + EXCESS_PRECISION_TYPE_FLOAT16 }; /* Level of size optimization. */ diff --git a/gcc/doc/extend.texi b/gcc/doc/extend.texi index 777dc0f529f..f08238b6ca9 100644 --- a/gcc/doc/extend.texi +++ b/gcc/doc/extend.texi @@ -1160,7 +1160,8 @@ On x86 targets with SSE2 enabled, without @option{-mavx512fp16}, software emulation and the @code{float} instructions. The default behavior for @code{FLT_EVAL_METHOD} is to keep the intermediate result of the operation as 32-bit precision. This may lead to inconsistent behavior between software -emulation and AVX512-FP16 instructions. +emulation and AVX512-FP16 instructions. Using @option{-fexcess-precision=16} +will force round back after each operation. @node Decimal Float @section Decimal Floating Types diff --git a/gcc/doc/tm.texi b/gcc/doc/tm.texi index f68f42638a1..be8148583d8 100644 --- a/gcc/doc/tm.texi +++ b/gcc/doc/tm.texi @@ -982,20 +982,26 @@ Do not define this macro if it would never modify @var{m}. Return a value, with the same meaning as the C99 macro @code{FLT_EVAL_METHOD} that describes which excess precision should be applied. @var{type} is either @code{EXCESS_PRECISION_TYPE_IMPLICIT}, -@code{EXCESS_PRECISION_TYPE_FAST}, or -@code{EXCESS_PRECISION_TYPE_STANDARD}. For +@code{EXCESS_PRECISION_TYPE_FAST}, +@code{EXCESS_PRECISION_TYPE_STANDARD}, or +@code{EXCESS_PRECISION_TYPE_FLOAT16}. For @code{EXCESS_PRECISION_TYPE_IMPLICIT}, the target should return which precision and range operations will be implictly evaluated in regardless of the excess precision explicitly added. For -@code{EXCESS_PRECISION_TYPE_STANDARD} and +@code{EXCESS_PRECISION_TYPE_STANDARD}, +@code{EXCESS_PRECISION_TYPE_FLOAT16}, and @code{EXCESS_PRECISION_TYPE_FAST}, the target should return the explicit excess precision that should be added depending on the value set for @option{-fexcess-precision=@r{[}standard@r{|}fast@r{]}}. Note that unpredictable explicit excess precision does not make sense, so a target should never return @code{FLT_EVAL_METHOD_UNPREDICTABLE} -when @var{type} is @code{EXCESS_PRECISION_TYPE_STANDARD} or +when @var{type} is @code{EXCESS_PRECISION_TYPE_STANDARD}, +@code{EXCESS_PRECISION_TYPE_FLOAT16} or @code{EXCESS_PRECISION_TYPE_FAST}. @end deftypefn +Return a value, with the same meaning as the C99 macro +@code{FLT_EVAL_METHOD} that describes which excess precision should be +applied. @deftypefn {Target Hook} machine_mode TARGET_PROMOTE_FUNCTION_MODE (const_tree @var{type}, machine_mode @var{mode}, int *@var{punsignedp}, const_tree @var{funtype}, int @var{for_return}) Like @code{PROMOTE_MODE}, but it is applied to outgoing function arguments or diff --git a/gcc/doc/tm.texi.in b/gcc/doc/tm.texi.in index fdf16b901c5..d088eee4afe 100644 --- a/gcc/doc/tm.texi.in +++ b/gcc/doc/tm.texi.in @@ -929,6 +929,9 @@ Do not define this macro if it would never modify @var{m}. @end defmac @hook TARGET_C_EXCESS_PRECISION +Return a value, with the same meaning as the C99 macro +@code{FLT_EVAL_METHOD} that describes which excess precision should be +applied. @hook TARGET_PROMOTE_FUNCTION_MODE diff --git a/gcc/flag-types.h b/gcc/flag-types.h index cc41b2ae103..45a2338d5f6 100644 --- a/gcc/flag-types.h +++ b/gcc/flag-types.h @@ -202,7 +202,8 @@ enum excess_precision { EXCESS_PRECISION_DEFAULT, EXCESS_PRECISION_FAST, - EXCESS_PRECISION_STANDARD + EXCESS_PRECISION_STANDARD, + EXCESS_PRECISION_FLOAT16 }; /* The options for which values of FLT_EVAL_METHOD are permissible. */ diff --git a/gcc/fortran/options.c b/gcc/fortran/options.c index 1723f689a57..847e20e8829 100644 --- a/gcc/fortran/options.c +++ b/gcc/fortran/options.c @@ -267,6 +267,9 @@ gfc_post_options (const char **pfilename) support. */ if (flag_excess_precision == EXCESS_PRECISION_STANDARD) sorry ("%<-fexcess-precision=standard%> for Fortran"); + else if (flag_excess_precision == EXCESS_PRECISION_FLOAT16) + sorry ("%<-fexcess-precision=16%> for Fortran"); + flag_excess_precision = EXCESS_PRECISION_FAST; /* Fortran allows associative math - but we cannot reassociate if diff --git a/gcc/target.def b/gcc/target.def index 28a34f1d51b..bfa819609c2 100644 --- a/gcc/target.def +++ b/gcc/target.def @@ -6225,18 +6225,21 @@ DEFHOOK "Return a value, with the same meaning as the C99 macro\n\ @code{FLT_EVAL_METHOD} that describes which excess precision should be\n\ applied. @var{type} is either @code{EXCESS_PRECISION_TYPE_IMPLICIT},\n\ -@code{EXCESS_PRECISION_TYPE_FAST}, or\n\ -@code{EXCESS_PRECISION_TYPE_STANDARD}. For\n\ +@code{EXCESS_PRECISION_TYPE_FAST},\n\ +@code{EXCESS_PRECISION_TYPE_STANDARD}, or\n\ +@code{EXCESS_PRECISION_TYPE_FLOAT16}. For\n\ @code{EXCESS_PRECISION_TYPE_IMPLICIT}, the target should return which\n\ precision and range operations will be implictly evaluated in regardless\n\ of the excess precision explicitly added. For\n\ -@code{EXCESS_PRECISION_TYPE_STANDARD} and\n\ +@code{EXCESS_PRECISION_TYPE_STANDARD}, \n\ +@code{EXCESS_PRECISION_TYPE_FLOAT16}, and\n\ @code{EXCESS_PRECISION_TYPE_FAST}, the target should return the\n\ explicit excess precision that should be added depending on the\n\ value set for @option{-fexcess-precision=@r{[}standard@r{|}fast@r{]}}.\n\ Note that unpredictable explicit excess precision does not make sense,\n\ so a target should never return @code{FLT_EVAL_METHOD_UNPREDICTABLE}\n\ -when @var{type} is @code{EXCESS_PRECISION_TYPE_STANDARD} or\n\ +when @var{type} is @code{EXCESS_PRECISION_TYPE_STANDARD},\n\ +@code{EXCESS_PRECISION_TYPE_FLOAT16} or\n\ @code{EXCESS_PRECISION_TYPE_FAST}.", enum flt_eval_method, (enum excess_precision_type type), default_excess_precision) diff --git a/gcc/testsuite/gcc.target/i386/float16-6.c b/gcc/testsuite/gcc.target/i386/float16-6.c new file mode 100644 index 00000000000..3d2503ce5e3 --- /dev/null +++ b/gcc/testsuite/gcc.target/i386/float16-6.c @@ -0,0 +1,8 @@ +/* { dg-do compile } */ +/* { dg-options "-msse2 -O2 -mfpmath=sse -fdump-tree-gimple -fexcess-precision=16" } */ +/* { dg-final { scan-tree-dump-not "\\(float\\)" "gimple" } } */ +_Float16 +foo (_Float16 a, _Float16 b, _Float16 c) +{ + return a + b + c; +} diff --git a/gcc/testsuite/gcc.target/i386/float16-7.c b/gcc/testsuite/gcc.target/i386/float16-7.c new file mode 100644 index 00000000000..86641afeba9 --- /dev/null +++ b/gcc/testsuite/gcc.target/i386/float16-7.c @@ -0,0 +1,9 @@ +/* { dg-do compile } */ +/* { dg-options "-O2 -mfpmath=387 -fexcess-precision=16" } */ +/* { dg-excess-errors "'-fexcess-precision=16' is not compatible with '-mfpmath=387'" } */ +_Float16 +foo (_Float16 a, _Float16 b) +{ + return a + b;/* { dg-error "'-fexcess-precision=16' is not compatible with '-mfpmath=387'" } */ +} + diff --git a/gcc/tree.c b/gcc/tree.c index a89d50a30ad..486cdb0e7c7 100644 --- a/gcc/tree.c +++ b/gcc/tree.c @@ -7637,7 +7637,8 @@ excess_precision_type (tree type) enum excess_precision_type requested_type = (flag_excess_precision == EXCESS_PRECISION_FAST ? EXCESS_PRECISION_TYPE_FAST - : EXCESS_PRECISION_TYPE_STANDARD); + : (flag_excess_precision == EXCESS_PRECISION_FLOAT16 + ? EXCESS_PRECISION_TYPE_FLOAT16 :EXCESS_PRECISION_TYPE_STANDARD)); enum flt_eval_method target_flt_eval_method = targetm.c.excess_precision (requested_type);