From c2565a31c1622ab0926aeef4a6579413e121b9f9 Mon Sep 17 00:00:00 2001 From: Jakub Jelinek Date: Fri, 14 Oct 2022 09:37:01 +0200 Subject: middle-end, c++, i386, libgcc: std::bfloat16_t and __bf16 arithmetic support Here is a complete patch to add std::bfloat16_t support on x86 (AArch64 and ARM left for later). Almost no BFmode optabs are added by the patch, so for binops/unops it extends to SFmode first and then truncates back to BFmode. For {HF,SF,DF,XF,TF}mode -> BFmode conversions libgcc has implementations of all those conversions so that we avoid double rounding, for BFmode -> {DF,XF,TF}mode conversions to avoid growing libgcc too much it emits BFmode -> SFmode conversion first and then converts to the even wider mode, neither step should be imprecise. For BFmode -> HFmode, it first emits a precise BFmode -> SFmode conversion and then SFmode -> HFmode, because neither format is subset or superset of the other, while SFmode is superset of both. expr.cc then contains a -ffast-math optimization of the BF -> SF and SF -> BF conversions if we don't optimize for space (and for the latter if -frounding-math isn't enabled either). For x86, perhaps truncsfbf2 optab could be defined for TARGET_AVX512BF16 but IMNSHO should FAIL if !flag_finite_math || flag_rounding_math || !flag_unsafe_math_optimizations, because I think the insn doesn't raise on sNaNs, hardcodes round to nearest and flushes denormals to zero. By default (unless x86 -fexcess-precision=16) we use float excess precision for BFmode, so truncate only on explicit casts and assignments. The patch introduces a single __bf16 builtin - __builtin_nansf16b, because (__bf16) __builtin_nansf ("") will drop the sNaN into qNaN, and uses f16b suffix instead of bf16 because there would be ambiguity on log vs. logb - __builtin_logbf16 could be either log with bf16 suffix or logb with f16 suffix. In other cases libstdc++ should mostly use __builtin_*f for std::bfloat16_t overloads (we have a problem with std::nextafter though but that one we have also for std::float16_t). 2022-10-14 Jakub Jelinek gcc/ * tree-core.h (enum tree_index): Add TI_BFLOAT16_TYPE. * tree.h (bfloat16_type_node): Define. * tree.cc (excess_precision_type): Promote bfloat16_type_mode like float16_type_mode. (build_common_tree_nodes): Initialize bfloat16_type_node if BFmode is supported. * expmed.h (maybe_expand_shift): Declare. * expmed.cc (maybe_expand_shift): No longer static. * expr.cc (convert_mode_scalar): Don't ICE on BF -> HF or HF -> BF conversions. If there is no optab, handle BF -> {DF,XF,TF,HF} conversions as separate BF -> SF -> {DF,XF,TF,HF} conversions, add -ffast-math generic implementation for BF -> SF and SF -> BF conversions. * builtin-types.def (BT_BFLOAT16, BT_FN_BFLOAT16_CONST_STRING): New. * builtins.def (BUILT_IN_NANSF16B): New builtin. * fold-const-call.cc (fold_const_call): Handle CFN_BUILT_IN_NANSF16B. * config/i386/i386.cc (classify_argument): Handle E_BCmode. (ix86_libgcc_floating_mode_supported_p): Also return true for BFmode for -msse2. (ix86_mangle_type): Mangle BFmode as DF16b. (ix86_invalid_conversion, ix86_invalid_unary_op, ix86_invalid_binary_op): Remove. (TARGET_INVALID_CONVERSION, TARGET_INVALID_UNARY_OP, TARGET_INVALID_BINARY_OP): Don't redefine. * config/i386/i386-builtins.cc (ix86_bf16_type_node): Remove. (ix86_register_bf16_builtin_type): Use bfloat16_type_node rather than ix86_bf16_type_node, only create it if still NULL. * config/i386/i386-builtin-types.def (BFLOAT16): Likewise. * config/i386/i386.md (cbranchbf4, cstorebf4): New expanders. gcc/c-family/ * c-cppbuiltin.cc (c_cpp_builtins): If bfloat16_type_node, predefine __BFLT16_*__ macros and for C++23 also __STDCPP_BFLOAT16_T__. Predefine bfloat16_type_node related macros for -fbuilding-libgcc. * c-lex.cc (interpret_float): Handle CPP_N_BFLOAT16. gcc/c/ * c-typeck.cc (convert_arguments): Don't promote __bf16 to double. gcc/cp/ * cp-tree.h (extended_float_type_p): Return true for bfloat16_type_node. * typeck.cc (cp_compare_floating_point_conversion_ranks): Set extended{1,2} if mv{1,2} is bfloat16_type_node. Adjust comment. gcc/testsuite/ * lib/target-supports.exp (check_effective_target_bfloat16, check_effective_target_bfloat16_runtime, add_options_for_bfloat16): New. * gcc.dg/torture/bfloat16-basic.c: New test. * gcc.dg/torture/bfloat16-builtin.c: New test. * gcc.dg/torture/bfloat16-builtin-issignaling-1.c: New test. * gcc.dg/torture/bfloat16-complex.c: New test. * gcc.dg/torture/builtin-issignaling-1.c: Allow to be includable from bfloat16-builtin-issignaling-1.c. * gcc.dg/torture/floatn-basic.h: Allow to be includable from bfloat16-basic.c. * gcc.target/i386/vect-bfloat16-typecheck_2.c: Adjust expected diagnostics. * gcc.target/i386/sse2-bfloat16-scalar-typecheck.c: Likewise. * gcc.target/i386/vect-bfloat16-typecheck_1.c: Likewise. * g++.target/i386/bfloat_cpp_typecheck.C: Likewise. libcpp/ * include/cpplib.h (CPP_N_BFLOAT16): Define. * expr.cc (interpret_float_suffix): Handle bf16 and BF16 suffixes for C++. libgcc/ * config/i386/t-softfp (softfp_extensions): Add bfsf. (softfp_truncations): Add tfbf xfbf dfbf sfbf hfbf. (CFLAGS-extendbfsf2.c, CFLAGS-truncsfbf2.c, CFLAGS-truncdfbf2.c, CFLAGS-truncxfbf2.c, CFLAGS-trunctfbf2.c, CFLAGS-trunchfbf2.c): Add -msse2. * config/i386/libgcc-glibc.ver (GCC_13.0.0): Export __extendbfsf2 and __trunc{s,d,x,t,h}fbf2. * config/i386/sfp-machine.h (_FP_NANSIGN_B): Define. * config/i386/64/sfp-machine.h (_FP_NANFRAC_B): Define. * config/i386/32/sfp-machine.h (_FP_NANFRAC_B): Define. * soft-fp/brain.h: New file. * soft-fp/truncsfbf2.c: New file. * soft-fp/truncdfbf2.c: New file. * soft-fp/truncxfbf2.c: New file. * soft-fp/trunctfbf2.c: New file. * soft-fp/trunchfbf2.c: New file. * soft-fp/truncbfhf2.c: New file. * soft-fp/extendbfsf2.c: New file. libiberty/ * cp-demangle.h (D_BUILTIN_TYPE_COUNT): Increment. * cp-demangle.c (cplus_demangle_builtin_types): Add std::bfloat16_t entry. (cplus_demangle_type): Demangle DF16b. * testsuite/demangle-expected (_Z3xxxDF16b): New test. --- gcc/expr.cc | 150 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 149 insertions(+), 1 deletion(-) (limited to 'gcc/expr.cc') diff --git a/gcc/expr.cc b/gcc/expr.cc index b897b6d..4c892d6 100644 --- a/gcc/expr.cc +++ b/gcc/expr.cc @@ -344,7 +344,11 @@ convert_mode_scalar (rtx to, rtx from, int unsignedp) gcc_assert ((GET_MODE_PRECISION (from_mode) != GET_MODE_PRECISION (to_mode)) || (DECIMAL_FLOAT_MODE_P (from_mode) - != DECIMAL_FLOAT_MODE_P (to_mode))); + != DECIMAL_FLOAT_MODE_P (to_mode)) + || (REAL_MODE_FORMAT (from_mode) == &arm_bfloat_half_format + && REAL_MODE_FORMAT (to_mode) == &ieee_half_format) + || (REAL_MODE_FORMAT (to_mode) == &arm_bfloat_half_format + && REAL_MODE_FORMAT (from_mode) == &ieee_half_format)); if (GET_MODE_PRECISION (from_mode) == GET_MODE_PRECISION (to_mode)) /* Conversion between decimal float and binary float, same size. */ @@ -364,6 +368,150 @@ convert_mode_scalar (rtx to, rtx from, int unsignedp) return; } +#ifdef HAVE_SFmode + if (REAL_MODE_FORMAT (from_mode) == &arm_bfloat_half_format + && REAL_MODE_FORMAT (SFmode) == &ieee_single_format) + { + if (GET_MODE_PRECISION (to_mode) > GET_MODE_PRECISION (SFmode)) + { + /* To cut down on libgcc size, implement + BFmode -> {DF,XF,TF}mode conversions by + BFmode -> SFmode -> {DF,XF,TF}mode conversions. */ + rtx temp = gen_reg_rtx (SFmode); + convert_mode_scalar (temp, from, unsignedp); + convert_mode_scalar (to, temp, unsignedp); + return; + } + if (REAL_MODE_FORMAT (to_mode) == &ieee_half_format) + { + /* Similarly, implement BFmode -> HFmode as + BFmode -> SFmode -> HFmode conversion where SFmode + has superset of BFmode values. We don't need + to handle sNaNs by raising exception and turning + into into qNaN though, as that can be done in the + SFmode -> HFmode conversion too. */ + rtx temp = gen_reg_rtx (SFmode); + int save_flag_finite_math_only = flag_finite_math_only; + flag_finite_math_only = true; + convert_mode_scalar (temp, from, unsignedp); + flag_finite_math_only = save_flag_finite_math_only; + convert_mode_scalar (to, temp, unsignedp); + return; + } + if (to_mode == SFmode + && !HONOR_NANS (from_mode) + && !HONOR_NANS (to_mode) + && optimize_insn_for_speed_p ()) + { + /* If we don't expect sNaNs, for BFmode -> SFmode we can just + shift the bits up. */ + machine_mode fromi_mode, toi_mode; + if (int_mode_for_size (GET_MODE_BITSIZE (from_mode), + 0).exists (&fromi_mode) + && int_mode_for_size (GET_MODE_BITSIZE (to_mode), + 0).exists (&toi_mode)) + { + start_sequence (); + rtx fromi = lowpart_subreg (fromi_mode, from, from_mode); + rtx tof = NULL_RTX; + if (fromi) + { + rtx toi = gen_reg_rtx (toi_mode); + convert_mode_scalar (toi, fromi, 1); + toi + = maybe_expand_shift (LSHIFT_EXPR, toi_mode, toi, + GET_MODE_PRECISION (to_mode) + - GET_MODE_PRECISION (from_mode), + NULL_RTX, 1); + if (toi) + { + tof = lowpart_subreg (to_mode, toi, toi_mode); + if (tof) + emit_move_insn (to, tof); + } + } + insns = get_insns (); + end_sequence (); + if (tof) + { + emit_insn (insns); + return; + } + } + } + } + if (REAL_MODE_FORMAT (from_mode) == &ieee_single_format + && REAL_MODE_FORMAT (to_mode) == &arm_bfloat_half_format + && !HONOR_NANS (from_mode) + && !HONOR_NANS (to_mode) + && !flag_rounding_math + && optimize_insn_for_speed_p ()) + { + /* If we don't expect qNaNs nor sNaNs and can assume rounding + to nearest, we can expand the conversion inline as + (fromi + 0x7fff + ((fromi >> 16) & 1)) >> 16. */ + machine_mode fromi_mode, toi_mode; + if (int_mode_for_size (GET_MODE_BITSIZE (from_mode), + 0).exists (&fromi_mode) + && int_mode_for_size (GET_MODE_BITSIZE (to_mode), + 0).exists (&toi_mode)) + { + start_sequence (); + rtx fromi = lowpart_subreg (fromi_mode, from, from_mode); + rtx tof = NULL_RTX; + do + { + if (!fromi) + break; + int shift = (GET_MODE_PRECISION (from_mode) + - GET_MODE_PRECISION (to_mode)); + rtx temp1 + = maybe_expand_shift (RSHIFT_EXPR, fromi_mode, fromi, + shift, NULL_RTX, 1); + if (!temp1) + break; + rtx temp2 + = expand_binop (fromi_mode, and_optab, temp1, const1_rtx, + NULL_RTX, 1, OPTAB_DIRECT); + if (!temp2) + break; + rtx temp3 + = expand_binop (fromi_mode, add_optab, fromi, + gen_int_mode ((HOST_WIDE_INT_1U + << (shift - 1)) - 1, + fromi_mode), NULL_RTX, + 1, OPTAB_DIRECT); + if (!temp3) + break; + rtx temp4 + = expand_binop (fromi_mode, add_optab, temp3, temp2, + NULL_RTX, 1, OPTAB_DIRECT); + if (!temp4) + break; + rtx temp5 = maybe_expand_shift (RSHIFT_EXPR, fromi_mode, + temp4, shift, NULL_RTX, 1); + if (!temp5) + break; + rtx temp6 = lowpart_subreg (toi_mode, temp5, fromi_mode); + if (!temp6) + break; + tof = lowpart_subreg (to_mode, force_reg (toi_mode, temp6), + toi_mode); + if (tof) + emit_move_insn (to, tof); + } + while (0); + insns = get_insns (); + end_sequence (); + if (tof) + { + emit_insn (insns); + return; + } + } + } +#endif + /* Otherwise use a libcall. */ libcall = convert_optab_libfunc (tab, to_mode, from_mode); -- cgit v1.1