//=-lib/fp_extend_impl.inc - low precision -> high precision conversion -*-- -// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // // This file implements a fairly generic conversion from a narrower to a wider // IEEE-754 floating-point type. The constants and types defined following the // includes below parameterize the conversion. // // It does not support types that don't use the usual IEEE-754 interchange // formats; specifically, some work would be needed to adapt it to // (for example) the Intel 80-bit format or PowerPC double-double format. // // Note please, however, that this implementation is only intended to support // *widening* operations; if you need to convert to a *narrower* floating-point // type (e.g. double -> float), then this routine will not do what you want it // to. // // It also requires that integer types at least as large as both formats // are available on the target platform; this may pose a problem when trying // to add support for quad on some 32-bit systems, for example. You also may // run into trouble finding an appropriate CLZ function for wide source types; // you will likely need to roll your own on some platforms. // // Finally, the following assumptions are made: // // 1. Floating-point types and integer types have the same endianness on the // target platform. // // 2. Quiet NaNs, if supported, are indicated by the leading bit of the // significand field being set. // //===----------------------------------------------------------------------===// #include "fp_extend.h" // The source type may use a usual IEEE-754 interchange format or Intel 80-bit // format. In particular, for the source type srcSigFracBits may be not equal to // srcSigBits. The destination type is assumed to be one of IEEE-754 standard // types. static __inline dst_t __extendXfYf2__(src_t a) { // Various constants whose values follow from the type parameters. // Any reasonable optimizer will fold and propagate all of these. const int srcInfExp = (1 << srcExpBits) - 1; const int srcExpBias = srcInfExp >> 1; const int dstInfExp = (1 << dstExpBits) - 1; const int dstExpBias = dstInfExp >> 1; // Break a into a sign and representation of the absolute value. const src_rep_t aRep = srcToRep(a); const src_rep_t srcSign = extract_sign_from_src(aRep); const src_rep_t srcExp = extract_exp_from_src(aRep); const src_rep_t srcSigFrac = extract_sig_frac_from_src(aRep); dst_rep_t dstSign = srcSign; dst_rep_t dstExp; dst_rep_t dstSigFrac; if (srcExp >= 1 && srcExp < (src_rep_t)srcInfExp) { // a is a normal number. dstExp = (dst_rep_t)srcExp + (dst_rep_t)(dstExpBias - srcExpBias); dstSigFrac = (dst_rep_t)srcSigFrac << (dstSigFracBits - srcSigFracBits); } else if (srcExp == srcInfExp) { // a is NaN or infinity. dstExp = dstInfExp; dstSigFrac = (dst_rep_t)srcSigFrac << (dstSigFracBits - srcSigFracBits); } else if (srcSigFrac) { // a is denormal. if (srcExpBits == dstExpBits) { // The exponent fields are identical and this is a denormal number, so all // the non-significand bits are zero. In particular, this branch is always // taken when we extend a denormal F80 to F128. dstExp = 0; dstSigFrac = ((dst_rep_t)srcSigFrac) << (dstSigFracBits - srcSigFracBits); } else { #ifndef src_rep_t_clz // If src_rep_t_clz is not defined this branch must be unreachable. __builtin_unreachable(); #else // Renormalize the significand and clear the leading bit. // For F80 -> F128 this codepath is unused. const int scale = clz_in_sig_frac(srcSigFrac) + 1; dstExp = dstExpBias - srcExpBias - scale + 1; dstSigFrac = (dst_rep_t)srcSigFrac << (dstSigFracBits - srcSigFracBits + scale); const dst_rep_t dstMinNormal = DST_REP_C(1) << (dstBits - 1 - dstExpBits); dstSigFrac ^= dstMinNormal; #endif } } else { // a is zero. dstExp = 0; dstSigFrac = 0; } const dst_rep_t result = construct_dst_rep(dstSign, dstExp, dstSigFrac); return dstFromRep(result); }