From patchwork Thu Sep 14 14:06:38 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Adhemerval Zanella Netto X-Patchwork-Id: 722542 Delivered-To: patch@linaro.org Received: by 2002:ab3:5684:0:b0:23f:8cfb:5ad7 with SMTP id d4csp472219ltc; Thu, 14 Sep 2023 07:07:03 -0700 (PDT) X-Google-Smtp-Source: AGHT+IFUU+VdBbnAHDih3bPZYgIv3K/wO2pvS7S/yCHNEjS8uJ+ASq0ojdp2tcBJosw4vEtC9byD X-Received: by 2002:a17:906:11a:b0:9ad:a486:1f7e with SMTP id 26-20020a170906011a00b009ada4861f7emr4238094eje.72.1694700422939; Thu, 14 Sep 2023 07:07:02 -0700 (PDT) ARC-Seal: i=1; a=rsa-sha256; t=1694700422; cv=none; d=google.com; s=arc-20160816; b=mPAiEzdiH5R7paggDM/6EvC9j5Aa+SF9tK22mYZCoClsRd6TS0p3jzRC6uhTpPMHkT ELgI9Ty0P0rR6DI8BPbKkoM3zn/WacRj4O0P/GR2+sg+A9Q3SqE9dCjdEKnYzCUSGE4u /UiGYF4YMk91smvDBcouORd49X71XL6KjwYsxsE1tpPzq6zLEL0bLACOSW+ppgSG7kUy 5zd/+rYDRswr0XG0fAAx9tYTcl2MkNDonTKl59+vfq1ZusFPGGZIQuGm207LLZYMZurI TwvcyYddWMfo5OuvGGg5VK/nmJpktiZFmjjsBIYS5yzwL5qx6YI8i0TWz5VeHfSR6Ci0 fqUA== ARC-Message-Signature: i=1; a=rsa-sha256; c=relaxed/relaxed; d=google.com; s=arc-20160816; h=errors-to:list-subscribe:list-help:list-post:list-archive :list-unsubscribe:list-id:precedence:content-transfer-encoding :mime-version:references:in-reply-to:message-id:date:subject:to:from :dkim-signature:dmarc-filter:delivered-to; bh=0gCwOHXPs8LvUsXG2rIlV/288vq5/8of7Pe15Yo1JKw=; fh=Ji/BrrBV5vvGE0PSJoc8P6vA3gQfM/DANnlOBrK0hiY=; b=LZlf7cpMY16Fr2F6yH8InyDHS3KfYZY8wBfvB38c3hb1Auh4k2ZiGrwhBLg51zYb7i Fq3TMQ3L5Qh3GxntbG6mz0kgIR8EbJOycZar+qy9N77iNLGlBZhLJ/gUOFnannbwKzRn CfUwl/5r/36uEEIx435OE/fIZfT/1XHctkSdKr37MMGLSBo/IoI8+ZvoD+0FmbnEZoFw R0wSen6cNIeI7J4ql4hw7wmfs92IsGYwEtmqHfw7L+M216vRE/CRHN0osnd5bhnFS3SR l6jRAj/Hw/xrNj/JjDjzQuSjHybN8+I4Um8e9u2nsZT9zYwa/qWPsXAfwkapra19BAvF 0ccA== ARC-Authentication-Results: i=1; mx.google.com; dkim=pass header.i=@linaro.org header.s=google header.b=nFY1z3H5; spf=pass (google.com: domain of libc-alpha-bounces+patch=linaro.org@sourceware.org designates 8.43.85.97 as permitted sender) smtp.mailfrom="libc-alpha-bounces+patch=linaro.org@sourceware.org"; dmarc=pass (p=NONE sp=NONE dis=NONE) header.from=linaro.org Return-Path: Received: from server2.sourceware.org (ip-8-43-85-97.sourceware.org. [8.43.85.97]) by mx.google.com with ESMTPS id v16-20020a1709060b5000b0099d9804740csi1251292ejg.858.2023.09.14.07.07.02 for (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Thu, 14 Sep 2023 07:07:02 -0700 (PDT) Received-SPF: pass (google.com: domain of libc-alpha-bounces+patch=linaro.org@sourceware.org designates 8.43.85.97 as permitted sender) client-ip=8.43.85.97; Authentication-Results: mx.google.com; dkim=pass header.i=@linaro.org header.s=google header.b=nFY1z3H5; spf=pass (google.com: domain of libc-alpha-bounces+patch=linaro.org@sourceware.org designates 8.43.85.97 as permitted sender) smtp.mailfrom="libc-alpha-bounces+patch=linaro.org@sourceware.org"; dmarc=pass (p=NONE sp=NONE dis=NONE) header.from=linaro.org Received: from server2.sourceware.org (localhost [IPv6:::1]) by sourceware.org (Postfix) with ESMTP id AB059385CC90 for ; Thu, 14 Sep 2023 14:07:01 +0000 (GMT) X-Original-To: libc-alpha@sourceware.org Delivered-To: libc-alpha@sourceware.org Received: from mail-oi1-x236.google.com (mail-oi1-x236.google.com [IPv6:2607:f8b0:4864:20::236]) by sourceware.org (Postfix) with ESMTPS id 40D1C3858C52 for ; Thu, 14 Sep 2023 14:06:52 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.2 sourceware.org 40D1C3858C52 Authentication-Results: sourceware.org; dmarc=pass (p=none dis=none) header.from=linaro.org Authentication-Results: sourceware.org; spf=pass smtp.mailfrom=linaro.org Received: by mail-oi1-x236.google.com with SMTP id 5614622812f47-3ab2a0391c0so594973b6e.1 for ; Thu, 14 Sep 2023 07:06:52 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=linaro.org; s=google; t=1694700411; x=1695305211; darn=sourceware.org; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:to:from:from:to:cc:subject:date:message-id :reply-to; bh=0gCwOHXPs8LvUsXG2rIlV/288vq5/8of7Pe15Yo1JKw=; b=nFY1z3H5W/ai45exvGihDwRA4AIhgIeXN5Yt8eS1JzpN7T4cHlRKBs19wcqC8xJBCM LhtX+PWAI31QBgIdh3OWfgq+YriIkf8JrzIsImJ/jrPZvtlDM93FMcMsaoT37dzgMr82 nrHfPkZaSqcxVls27YWNwIc6pdY6tfNN+qDL2y2oeCgUwwtJFszRQjRPONFsgQkAr49l +aODJCA/gB3KnR/akqawInv3uYSJ/klw866CKi12wZZXGePZzMYL0tPGrqu7KyKPoYOe WURl8fD8h+qywDkRjW3RMc2k/lhWtaZZjkXviSBLwyoYTM313w5WSTglJapcIxDYNshl kgYw== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20230601; t=1694700411; x=1695305211; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:to:from:x-gm-message-state:from:to:cc :subject:date:message-id:reply-to; bh=0gCwOHXPs8LvUsXG2rIlV/288vq5/8of7Pe15Yo1JKw=; b=M3npHkTaQfB+J8SpOwm8zZBIsezbkdt6f0GI5VQz4P33JPlpJSezvLyN8KcMekpNmx mdthUxaUQ52GQ62WIjjTgYNrEIMc3Zvx5bi1N3gFufu1lhdVl4iVXq/FxYNdsGCw1c1L Zjj/z85tL9k7R0At40xeSmcJkh8eTxmwqyyXct38wusKPKXDcBq/RqmRSJ+/mZ+ew7vD NhVUJ2LrCFp5kXqDK0TFGcEAd7CM4dKrVXjfTla2G4PKe4V5SrUkAEYYGl15apHlfJJG gUw8nLbgN6r7WQTEFfCaO1RBGFvRa2xizbGT7gLRvnPPC0lHsqLGq1JfRF+gRy7cAhZV f5/Q== X-Gm-Message-State: AOJu0YxxmlSJ1UeqC1v737mBLHcF7GpAElK+xBGSJDy4VycDQucVBalX X6lpnIX8lgJvvhOT2U4aPwlPjDwVuDGO3JRHT4W50Q== X-Received: by 2002:a05:6808:13ce:b0:3a3:76c6:a46f with SMTP id d14-20020a05680813ce00b003a376c6a46fmr7157509oiw.38.1694700410506; Thu, 14 Sep 2023 07:06:50 -0700 (PDT) Received: from mandiga.. ([2804:1b3:a7c0:91cb:5173:d70:d8d:4d2c]) by smtp.gmail.com with ESMTPSA id s26-20020a056808009a00b003a862e70bcbsm620417oic.13.2023.09.14.07.06.48 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Thu, 14 Sep 2023 07:06:49 -0700 (PDT) From: Adhemerval Zanella To: libc-alpha@sourceware.org, Joe Simmons-Talbott Subject: [PATCH v2 1/2] stdio: Remove __printf_fp_buffer_1 alloca usage Date: Thu, 14 Sep 2023 11:06:38 -0300 Message-Id: <20230914140639.1912983-2-adhemerval.zanella@linaro.org> X-Mailer: git-send-email 2.34.1 In-Reply-To: <20230914140639.1912983-1-adhemerval.zanella@linaro.org> References: <20230914140639.1912983-1-adhemerval.zanella@linaro.org> MIME-Version: 1.0 X-Spam-Status: No, score=-12.3 required=5.0 tests=BAYES_00, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, DKIM_VALID_EF, GIT_PATCH_0, RCVD_IN_DNSWL_NONE, SPF_HELO_NONE, SPF_PASS, TXREP autolearn=ham autolearn_force=no version=3.4.6 X-Spam-Checker-Version: SpamAssassin 3.4.6 (2021-04-09) on server2.sourceware.org X-BeenThere: libc-alpha@sourceware.org X-Mailman-Version: 2.1.30 Precedence: list List-Id: Libc-alpha mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: libc-alpha-bounces+patch=linaro.org@sourceware.org The function is refactored so double and long double / Float128 are handled in different functions and thus different stack requirements. For double, the code now allocated the maximum required size for all possible inputs (about ~480 bytes). For long double / _Float128, which would about 6240 bytes, the required memory is allocated using a stratch_buffer based on the required expoenent (so most inputs should still use the default scratch_buffer buffer). Checked on x86_64-linux-gnu. --- stdio-common/printf_fp.c | 605 +++++++++++++++++++++------------------ 1 file changed, 333 insertions(+), 272 deletions(-) diff --git a/stdio-common/printf_fp.c b/stdio-common/printf_fp.c index 6f22985ba1..9ddf61fd6d 100644 --- a/stdio-common/printf_fp.c +++ b/stdio-common/printf_fp.c @@ -44,6 +44,7 @@ #include #include #include +#include #include @@ -52,6 +53,15 @@ An MP variable occupies a varying number of entries in its array. We keep track of this number for efficiency reasons. Otherwise we would always have to process the whole array. */ + +/* When _Float128 is enabled in the library and ABI-distinct from long + double, we need mp_limbs enough for any of them. */ +#if __HAVE_DISTINCT_FLOAT128 +# define GREATER_MANT_DIG FLT128_MANT_DIG +#else +# define GREATER_MANT_DIG LDBL_MANT_DIG +#endif + #define MPN_VAR(name) mp_limb_t *name; mp_size_t name##size #define MPN_ASSIGN(dst,src) \ @@ -81,6 +91,17 @@ struct hack_digit_param MPN_VAR(scale); /* Temporary bignum value. */ MPN_VAR(tmp); + + int is_neg; + /* "NaN" or "Inf" for the special cases. */ + const char *special; + /* We need to shift the contents of fp_input by this amount of bits. */ + int to_shift; + + /* We need just a few limbs for the input before shifting to the right + position. */ + mp_limb_t fp_input[(GREATER_MANT_DIG + BITS_PER_MP_LIMB - 1) + / BITS_PER_MP_LIMB]; }; static char @@ -135,47 +156,15 @@ hack_digit (struct hack_digit_param *p) temporary buffer. To prepare for that, THOUSANDS_SEP_LENGTH is the final length of the thousands separator. */ static void -__printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, - char thousands_sep, char decimal, +__printf_fp_buffer_2 (struct hack_digit_param *p, struct __printf_buffer *buf, + locale_t loc, char thousands_sep, char decimal, unsigned int thousands_sep_length, const struct printf_info *info, const void *const *args) { - /* The floating-point value to output. */ - union - { - double dbl; - long double ldbl; -#if __HAVE_DISTINCT_FLOAT128 - _Float128 f128; -#endif - } - fpnum; - - /* "NaN" or "Inf" for the special cases. */ - const char *special = NULL; - /* Used to determine grouping rules. */ int lc_category = info->extra ? LC_MONETARY : LC_NUMERIC; - /* When _Float128 is enabled in the library and ABI-distinct from long - double, we need mp_limbs enough for any of them. */ -#if __HAVE_DISTINCT_FLOAT128 -# define GREATER_MANT_DIG FLT128_MANT_DIG -#else -# define GREATER_MANT_DIG LDBL_MANT_DIG -#endif - /* We need just a few limbs for the input before shifting to the right - position. */ - mp_limb_t fp_input[(GREATER_MANT_DIG + BITS_PER_MP_LIMB - 1) - / BITS_PER_MP_LIMB]; - /* We need to shift the contents of fp_input by this amount of bits. */ - int to_shift = 0; - - struct hack_digit_param p; - /* Sign of float number. */ - int is_neg = 0; - /* General helper (carry limb). */ mp_limb_t cy; @@ -184,72 +173,27 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, /* Flag whether wbuffer and buffer are malloc'ed or not. */ int buffer_malloced = 0; - p.expsign = 0; - -#define PRINTF_FP_FETCH(FLOAT, VAR, SUFFIX, MANT_DIG) \ - { \ - (VAR) = *(const FLOAT *) args[0]; \ - \ - /* Check for special values: not a number or infinity. */ \ - if (isnan (VAR)) \ - { \ - is_neg = signbit (VAR); \ - if (isupper (info->spec)) \ - special = "NAN"; \ - else \ - special = "nan"; \ - } \ - else if (isinf (VAR)) \ - { \ - is_neg = signbit (VAR); \ - if (isupper (info->spec)) \ - special = "INF"; \ - else \ - special = "inf"; \ - } \ - else \ - { \ - p.fracsize = __mpn_extract_##SUFFIX \ - (fp_input, array_length (fp_input), \ - &p.exponent, &is_neg, VAR); \ - to_shift = 1 + p.fracsize * BITS_PER_MP_LIMB - MANT_DIG; \ - } \ - } - - /* Fetch the argument value. */ -#if __HAVE_DISTINCT_FLOAT128 - if (info->is_binary128) - PRINTF_FP_FETCH (_Float128, fpnum.f128, float128, FLT128_MANT_DIG) - else -#endif -#ifndef __NO_LONG_DOUBLE_MATH - if (info->is_long_double && sizeof (long double) > sizeof (double)) - PRINTF_FP_FETCH (long double, fpnum.ldbl, long_double, LDBL_MANT_DIG) - else -#endif - PRINTF_FP_FETCH (double, fpnum.dbl, double, DBL_MANT_DIG) - -#undef PRINTF_FP_FETCH + p->expsign = 0; - if (special) + if (p->special) { int width = info->width; - if (is_neg || info->showsign || info->space) + if (p->is_neg || info->showsign || info->space) --width; width -= 3; if (!info->left) __printf_buffer_pad (buf, ' ', width); - if (is_neg) + if (p->is_neg) __printf_buffer_putc (buf, '-'); else if (info->showsign) __printf_buffer_putc (buf, '+'); else if (info->space) __printf_buffer_putc (buf, ' '); - __printf_buffer_puts (buf, special); + __printf_buffer_puts (buf, p->special); if (info->left) __printf_buffer_pad (buf, ' ', width); @@ -257,27 +201,11 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, return; } - - /* We need three multiprecision variables. Now that we have the p.exponent - of the number we can allocate the needed memory. It would be more - efficient to use variables of the fixed maximum size but because this - would be really big it could lead to memory problems. */ - { - mp_size_t bignum_size = ((abs (p.exponent) + BITS_PER_MP_LIMB - 1) - / BITS_PER_MP_LIMB - + (GREATER_MANT_DIG / BITS_PER_MP_LIMB > 2 - ? 8 : 4)) - * sizeof (mp_limb_t); - p.frac = (mp_limb_t *) alloca (bignum_size); - p.tmp = (mp_limb_t *) alloca (bignum_size); - p.scale = (mp_limb_t *) alloca (bignum_size); - } - /* We now have to distinguish between numbers with positive and negative exponents because the method used for the one is not applicable/efficient for the other. */ - p.scalesize = 0; - if (p.exponent > 2) + p->scalesize = 0; + if (p->exponent > 2) { /* |FP| >= 8.0. */ int scaleexpo = 0; @@ -294,23 +222,23 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, const struct mp_power *powers = &_fpioconst_pow10[explog + 1]; int cnt_h, cnt_l, i; - if ((p.exponent + to_shift) % BITS_PER_MP_LIMB == 0) + if ((p->exponent + p->to_shift) % BITS_PER_MP_LIMB == 0) { - MPN_COPY_DECR (p.frac + (p.exponent + to_shift) / BITS_PER_MP_LIMB, - fp_input, p.fracsize); - p.fracsize += (p.exponent + to_shift) / BITS_PER_MP_LIMB; + MPN_COPY_DECR (p->frac + (p->exponent + p->to_shift) / BITS_PER_MP_LIMB, + p->fp_input, p->fracsize); + p->fracsize += (p->exponent + p->to_shift) / BITS_PER_MP_LIMB; } else { - cy = __mpn_lshift (p.frac - + (p.exponent + to_shift) / BITS_PER_MP_LIMB, - fp_input, p.fracsize, - (p.exponent + to_shift) % BITS_PER_MP_LIMB); - p.fracsize += (p.exponent + to_shift) / BITS_PER_MP_LIMB; + cy = __mpn_lshift (p->frac + + (p->exponent + p->to_shift) / BITS_PER_MP_LIMB, + p->fp_input, p->fracsize, + (p->exponent + p->to_shift) % BITS_PER_MP_LIMB); + p->fracsize += (p->exponent + p->to_shift) / BITS_PER_MP_LIMB; if (cy) - p.frac[p.fracsize++] = cy; + p->frac[p->fracsize++] = cy; } - MPN_ZERO (p.frac, (p.exponent + to_shift) / BITS_PER_MP_LIMB); + MPN_ZERO (p->frac, (p->exponent + p->to_shift) / BITS_PER_MP_LIMB); assert (powers > &_fpioconst_pow10[0]); do @@ -319,9 +247,9 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, /* The number of the product of two binary numbers with n and m bits respectively has m+n or m+n-1 bits. */ - if (p.exponent >= scaleexpo + powers->p_expo - 1) + if (p->exponent >= scaleexpo + powers->p_expo - 1) { - if (p.scalesize == 0) + if (p->scalesize == 0) { #if __HAVE_DISTINCT_FLOAT128 if ((FLT128_MANT_DIG @@ -333,14 +261,14 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, - _FPIO_CONST_OFFSET) /* 64bit const offset is not enough for IEEE 854 quad long double (_Float128). */ - p.tmpsize = powers->arraysize + _FLT128_FPIO_CONST_SHIFT; - memcpy (p.tmp + _FLT128_FPIO_CONST_SHIFT, + p->tmpsize = powers->arraysize + _FLT128_FPIO_CONST_SHIFT; + memcpy (p->tmp + _FLT128_FPIO_CONST_SHIFT, &__tens[powers->arrayoff], - p.tmpsize * sizeof (mp_limb_t)); - MPN_ZERO (p.tmp, _FLT128_FPIO_CONST_SHIFT); - /* Adjust p.exponent, as scaleexpo will be this much + p->tmpsize * sizeof (mp_limb_t)); + MPN_ZERO (p->tmp, _FLT128_FPIO_CONST_SHIFT); + /* Adjust p->exponent, as scaleexpo will be this much bigger too. */ - p.exponent += _FLT128_FPIO_CONST_SHIFT * BITS_PER_MP_LIMB; + p->exponent += _FLT128_FPIO_CONST_SHIFT * BITS_PER_MP_LIMB; } else #endif /* __HAVE_DISTINCT_FLOAT128 */ @@ -353,62 +281,62 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, - _FPIO_CONST_OFFSET) /* 64bit const offset is not enough for IEEE quad long double. */ - p.tmpsize = powers->arraysize + _FPIO_CONST_SHIFT; - memcpy (p.tmp + _FPIO_CONST_SHIFT, + p->tmpsize = powers->arraysize + _FPIO_CONST_SHIFT; + memcpy (p->tmp + _FPIO_CONST_SHIFT, &__tens[powers->arrayoff], - p.tmpsize * sizeof (mp_limb_t)); - MPN_ZERO (p.tmp, _FPIO_CONST_SHIFT); - /* Adjust p.exponent, as scaleexpo will be this much + p->tmpsize * sizeof (mp_limb_t)); + MPN_ZERO (p->tmp, _FPIO_CONST_SHIFT); + /* Adjust p->exponent, as scaleexpo will be this much bigger too. */ - p.exponent += _FPIO_CONST_SHIFT * BITS_PER_MP_LIMB; + p->exponent += _FPIO_CONST_SHIFT * BITS_PER_MP_LIMB; } else #endif { - p.tmpsize = powers->arraysize; - memcpy (p.tmp, &__tens[powers->arrayoff], - p.tmpsize * sizeof (mp_limb_t)); + p->tmpsize = powers->arraysize; + memcpy (p->tmp, &__tens[powers->arrayoff], + p->tmpsize * sizeof (mp_limb_t)); } } else { - cy = __mpn_mul (p.tmp, p.scale, p.scalesize, + cy = __mpn_mul (p->tmp, p->scale, p->scalesize, &__tens[powers->arrayoff + _FPIO_CONST_OFFSET], powers->arraysize - _FPIO_CONST_OFFSET); - p.tmpsize = p.scalesize + p->tmpsize = p->scalesize + powers->arraysize - _FPIO_CONST_OFFSET; if (cy == 0) - --p.tmpsize; + --p->tmpsize; } - if (MPN_GE (p.frac, p.tmp)) + if (MPN_GE (p->frac, p->tmp)) { int cnt; - MPN_ASSIGN (p.scale, p.tmp); - count_leading_zeros (cnt, p.scale[p.scalesize - 1]); - scaleexpo = (p.scalesize - 2) * BITS_PER_MP_LIMB - cnt - 1; + MPN_ASSIGN (p->scale, p->tmp); + count_leading_zeros (cnt, p->scale[p->scalesize - 1]); + scaleexpo = (p->scalesize - 2) * BITS_PER_MP_LIMB - cnt - 1; exp10 |= 1 << explog; } } --explog; } while (powers > &_fpioconst_pow10[0]); - p.exponent = exp10; + p->exponent = exp10; /* Optimize number representations. We want to represent the numbers with the lowest number of bytes possible without losing any bytes. Also the highest bit in the scaling factor has to be set (this is a requirement of the MPN division routines). */ - if (p.scalesize > 0) + if (p->scalesize > 0) { /* Determine minimum number of zero bits at the end of both numbers. */ - for (i = 0; p.scale[i] == 0 && p.frac[i] == 0; i++) + for (i = 0; p->scale[i] == 0 && p->frac[i] == 0; i++) ; /* Determine number of bits the scaling factor is misplaced. */ - count_leading_zeros (cnt_h, p.scale[p.scalesize - 1]); + count_leading_zeros (cnt_h, p->scale[p->scalesize - 1]); if (cnt_h == 0) { @@ -416,27 +344,27 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, we only have to remove the trailing empty limbs. */ if (i > 0) { - MPN_COPY_INCR (p.scale, p.scale + i, p.scalesize - i); - p.scalesize -= i; - MPN_COPY_INCR (p.frac, p.frac + i, p.fracsize - i); - p.fracsize -= i; + MPN_COPY_INCR (p->scale, p->scale + i, p->scalesize - i); + p->scalesize -= i; + MPN_COPY_INCR (p->frac, p->frac + i, p->fracsize - i); + p->fracsize -= i; } } else { - if (p.scale[i] != 0) + if (p->scale[i] != 0) { - count_trailing_zeros (cnt_l, p.scale[i]); - if (p.frac[i] != 0) + count_trailing_zeros (cnt_l, p->scale[i]); + if (p->frac[i] != 0) { int cnt_l2; - count_trailing_zeros (cnt_l2, p.frac[i]); + count_trailing_zeros (cnt_l2, p->frac[i]); if (cnt_l2 < cnt_l) cnt_l = cnt_l2; } } else - count_trailing_zeros (cnt_l, p.frac[i]); + count_trailing_zeros (cnt_l, p->frac[i]); /* Now shift the numbers to their optimal position. */ if (i == 0 && BITS_PER_MP_LIMB - cnt_h > cnt_l) @@ -444,10 +372,10 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, /* We cannot save any memory. So just roll both numbers so that the scaling factor has its highest bit set. */ - (void) __mpn_lshift (p.scale, p.scale, p.scalesize, cnt_h); - cy = __mpn_lshift (p.frac, p.frac, p.fracsize, cnt_h); + (void) __mpn_lshift (p->scale, p->scale, p->scalesize, cnt_h); + cy = __mpn_lshift (p->frac, p->frac, p->fracsize, cnt_h); if (cy != 0) - p.frac[p.fracsize++] = cy; + p->frac[p->fracsize++] = cy; } else if (BITS_PER_MP_LIMB - cnt_h <= cnt_l) { @@ -455,32 +383,32 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, and by packing the non-zero limbs which gain another free one. */ - (void) __mpn_rshift (p.scale, p.scale + i, p.scalesize - i, + (void) __mpn_rshift (p->scale, p->scale + i, p->scalesize - i, BITS_PER_MP_LIMB - cnt_h); - p.scalesize -= i + 1; - (void) __mpn_rshift (p.frac, p.frac + i, p.fracsize - i, + p->scalesize -= i + 1; + (void) __mpn_rshift (p->frac, p->frac + i, p->fracsize - i, BITS_PER_MP_LIMB - cnt_h); - p.fracsize -= p.frac[p.fracsize - i - 1] == 0 ? i + 1 : i; + p->fracsize -= p->frac[p->fracsize - i - 1] == 0 ? i + 1 : i; } else { /* We can only save the memory of the limbs which are zero. The non-zero parts occupy the same number of limbs. */ - (void) __mpn_rshift (p.scale, p.scale + (i - 1), - p.scalesize - (i - 1), + (void) __mpn_rshift (p->scale, p->scale + (i - 1), + p->scalesize - (i - 1), BITS_PER_MP_LIMB - cnt_h); - p.scalesize -= i; - (void) __mpn_rshift (p.frac, p.frac + (i - 1), - p.fracsize - (i - 1), + p->scalesize -= i; + (void) __mpn_rshift (p->frac, p->frac + (i - 1), + p->fracsize - (i - 1), BITS_PER_MP_LIMB - cnt_h); - p.fracsize -= - p.frac[p.fracsize - (i - 1) - 1] == 0 ? i : i - 1; + p->fracsize -= + p->frac[p->fracsize - (i - 1) - 1] == 0 ? i : i - 1; } } } } - else if (p.exponent < 0) + else if (p->exponent < 0) { /* |FP| < 1.0. */ int exp10 = 0; @@ -496,48 +424,48 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, const struct mp_power *powers = &_fpioconst_pow10[explog + 1]; /* Now shift the input value to its right place. */ - cy = __mpn_lshift (p.frac, fp_input, p.fracsize, to_shift); - p.frac[p.fracsize++] = cy; - assert (cy == 1 || (p.frac[p.fracsize - 2] == 0 && p.frac[0] == 0)); + cy = __mpn_lshift (p->frac, p->fp_input, p->fracsize, p->to_shift); + p->frac[p->fracsize++] = cy; + assert (cy == 1 || (p->frac[p->fracsize - 2] == 0 && p->frac[0] == 0)); - p.expsign = 1; - p.exponent = -p.exponent; + p->expsign = 1; + p->exponent = -p->exponent; assert (powers != &_fpioconst_pow10[0]); do { --powers; - if (p.exponent >= powers->m_expo) + if (p->exponent >= powers->m_expo) { int i, incr, cnt_h, cnt_l; mp_limb_t topval[2]; /* The __mpn_mul function expects the first argument to be bigger than the second. */ - if (p.fracsize < powers->arraysize - _FPIO_CONST_OFFSET) - cy = __mpn_mul (p.tmp, &__tens[powers->arrayoff + if (p->fracsize < powers->arraysize - _FPIO_CONST_OFFSET) + cy = __mpn_mul (p->tmp, &__tens[powers->arrayoff + _FPIO_CONST_OFFSET], powers->arraysize - _FPIO_CONST_OFFSET, - p.frac, p.fracsize); + p->frac, p->fracsize); else - cy = __mpn_mul (p.tmp, p.frac, p.fracsize, + cy = __mpn_mul (p->tmp, p->frac, p->fracsize, &__tens[powers->arrayoff + _FPIO_CONST_OFFSET], powers->arraysize - _FPIO_CONST_OFFSET); - p.tmpsize = p.fracsize + powers->arraysize - _FPIO_CONST_OFFSET; + p->tmpsize = p->fracsize + powers->arraysize - _FPIO_CONST_OFFSET; if (cy == 0) - --p.tmpsize; + --p->tmpsize; - count_leading_zeros (cnt_h, p.tmp[p.tmpsize - 1]); - incr = (p.tmpsize - p.fracsize) * BITS_PER_MP_LIMB + count_leading_zeros (cnt_h, p->tmp[p->tmpsize - 1]); + incr = (p->tmpsize - p->fracsize) * BITS_PER_MP_LIMB + BITS_PER_MP_LIMB - 1 - cnt_h; assert (incr <= powers->p_expo); - /* If we increased the p.exponent by exactly 3 we have to test + /* If we increased the p->exponent by exactly 3 we have to test for overflow. This is done by comparing with 10 shifted to the right position. */ - if (incr == p.exponent + 3) + if (incr == p->exponent + 3) { if (cnt_h <= BITS_PER_MP_LIMB - 4) { @@ -559,32 +487,32 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, against 10.0. If it is greater or equal to 10.0 the multiplication was not valid. This is because we cannot determine the number of bits in the result in advance. */ - if (incr < p.exponent + 3 - || (incr == p.exponent + 3 - && (p.tmp[p.tmpsize - 1] < topval[1] - || (p.tmp[p.tmpsize - 1] == topval[1] - && p.tmp[p.tmpsize - 2] < topval[0])))) + if (incr < p->exponent + 3 + || (incr == p->exponent + 3 + && (p->tmp[p->tmpsize - 1] < topval[1] + || (p->tmp[p->tmpsize - 1] == topval[1] + && p->tmp[p->tmpsize - 2] < topval[0])))) { /* The factor is right. Adapt binary and decimal exponents. */ - p.exponent -= incr; + p->exponent -= incr; exp10 |= 1 << explog; /* If this factor yields a number greater or equal to 1.0, we must not shift the non-fractional digits down. */ - if (p.exponent < 0) - cnt_h += -p.exponent; + if (p->exponent < 0) + cnt_h += -p->exponent; /* Now we optimize the number representation. */ - for (i = 0; p.tmp[i] == 0; ++i); + for (i = 0; p->tmp[i] == 0; ++i); if (cnt_h == BITS_PER_MP_LIMB - 1) { - MPN_COPY (p.frac, p.tmp + i, p.tmpsize - i); - p.fracsize = p.tmpsize - i; + MPN_COPY (p->frac, p->tmp + i, p->tmpsize - i); + p->fracsize = p->tmpsize - i; } else { - count_trailing_zeros (cnt_l, p.tmp[i]); + count_trailing_zeros (cnt_l, p->tmp[i]); /* Now shift the numbers to their optimal position. */ if (i == 0 && BITS_PER_MP_LIMB - 1 - cnt_h > cnt_l) @@ -593,16 +521,16 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, number so that the leading digit is in a separate limb. */ - cy = __mpn_lshift (p.frac, p.tmp, p.tmpsize, + cy = __mpn_lshift (p->frac, p->tmp, p->tmpsize, cnt_h + 1); - p.fracsize = p.tmpsize + 1; - p.frac[p.fracsize - 1] = cy; + p->fracsize = p->tmpsize + 1; + p->frac[p->fracsize - 1] = cy; } else if (BITS_PER_MP_LIMB - 1 - cnt_h <= cnt_l) { - (void) __mpn_rshift (p.frac, p.tmp + i, p.tmpsize - i, + (void) __mpn_rshift (p->frac, p->tmp + i, p->tmpsize - i, BITS_PER_MP_LIMB - 1 - cnt_h); - p.fracsize = p.tmpsize - i; + p->fracsize = p->tmpsize - i; } else { @@ -610,41 +538,41 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, are zero. The non-zero parts occupy the same number of limbs. */ - (void) __mpn_rshift (p.frac, p.tmp + (i - 1), - p.tmpsize - (i - 1), + (void) __mpn_rshift (p->frac, p->tmp + (i - 1), + p->tmpsize - (i - 1), BITS_PER_MP_LIMB - 1 - cnt_h); - p.fracsize = p.tmpsize - (i - 1); + p->fracsize = p->tmpsize - (i - 1); } } } } --explog; } - while (powers != &_fpioconst_pow10[1] && p.exponent > 0); + while (powers != &_fpioconst_pow10[1] && p->exponent > 0); /* All factors but 10^-1 are tested now. */ - if (p.exponent > 0) + if (p->exponent > 0) { int cnt_l; - cy = __mpn_mul_1 (p.tmp, p.frac, p.fracsize, 10); - p.tmpsize = p.fracsize; - assert (cy == 0 || p.tmp[p.tmpsize - 1] < 20); + cy = __mpn_mul_1 (p->tmp, p->frac, p->fracsize, 10); + p->tmpsize = p->fracsize; + assert (cy == 0 || p->tmp[p->tmpsize - 1] < 20); - count_trailing_zeros (cnt_l, p.tmp[0]); - if (cnt_l < MIN (4, p.exponent)) + count_trailing_zeros (cnt_l, p->tmp[0]); + if (cnt_l < MIN (4, p->exponent)) { - cy = __mpn_lshift (p.frac, p.tmp, p.tmpsize, - BITS_PER_MP_LIMB - MIN (4, p.exponent)); + cy = __mpn_lshift (p->frac, p->tmp, p->tmpsize, + BITS_PER_MP_LIMB - MIN (4, p->exponent)); if (cy != 0) - p.frac[p.tmpsize++] = cy; + p->frac[p->tmpsize++] = cy; } else - (void) __mpn_rshift (p.frac, p.tmp, p.tmpsize, MIN (4, p.exponent)); - p.fracsize = p.tmpsize; + (void) __mpn_rshift (p->frac, p->tmp, p->tmpsize, MIN (4, p->exponent)); + p->fracsize = p->tmpsize; exp10 |= 1; - assert (p.frac[p.fracsize - 1] < 10); + assert (p->frac[p->fracsize - 1] < 10); } - p.exponent = exp10; + p->exponent = exp10; } else { @@ -652,13 +580,13 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, numbers are in the range of 1.0 <= |fp| < 8.0. We simply shift it to the right place and divide it by 1.0 to get the leading digit. (Of course this division is not really made.) */ - assert (0 <= p.exponent && p.exponent < 3 - && p.exponent + to_shift < BITS_PER_MP_LIMB); + assert (0 <= p->exponent && p->exponent < 3 + && p->exponent + p->to_shift < BITS_PER_MP_LIMB); /* Now shift the input value to its right place. */ - cy = __mpn_lshift (p.frac, fp_input, p.fracsize, (p.exponent + to_shift)); - p.frac[p.fracsize++] = cy; - p.exponent = 0; + cy = __mpn_lshift (p->frac, p->fp_input, p->fracsize, (p->exponent + p->to_shift)); + p->frac[p->fracsize++] = cy; + p->exponent = 0; } { @@ -675,7 +603,7 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, if (spec == 'e') { - p.type = info->spec; + p->type = info->spec; intdig_max = 1; fracdig_min = fracdig_max = info->prec < 0 ? 6 : info->prec; chars_needed = 1 + 1 + (size_t) fracdig_max + 1 + 1 + 4; @@ -685,15 +613,15 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, } else if (spec == 'f') { - p.type = 'f'; + p->type = 'f'; fracdig_min = fracdig_max = info->prec < 0 ? 6 : info->prec; dig_max = INT_MAX; /* Unlimited. */ significant = 1; /* Does not matter here. */ - if (p.expsign == 0) + if (p->expsign == 0) { - intdig_max = p.exponent + 1; + intdig_max = p->exponent + 1; /* This can be really big! */ /* XXX Maybe malloc if too big? */ - chars_needed = (size_t) p.exponent + 1 + 1 + (size_t) fracdig_max; + chars_needed = (size_t) p->exponent + 1 + 1 + (size_t) fracdig_max; } else { @@ -704,27 +632,27 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, else { dig_max = info->prec < 0 ? 6 : (info->prec == 0 ? 1 : info->prec); - if ((p.expsign == 0 && p.exponent >= dig_max) - || (p.expsign != 0 && p.exponent > 4)) + if ((p->expsign == 0 && p->exponent >= dig_max) + || (p->expsign != 0 && p->exponent > 4)) { if ('g' - 'G' == 'e' - 'E') - p.type = 'E' + (info->spec - 'G'); + p->type = 'E' + (info->spec - 'G'); else - p.type = isupper (info->spec) ? 'E' : 'e'; + p->type = isupper (info->spec) ? 'E' : 'e'; fracdig_max = dig_max - 1; intdig_max = 1; chars_needed = 1 + 1 + (size_t) fracdig_max + 1 + 1 + 4; } else { - p.type = 'f'; - intdig_max = p.expsign == 0 ? p.exponent + 1 : 0; + p->type = 'f'; + intdig_max = p->expsign == 0 ? p->exponent + 1 : 0; fracdig_max = dig_max - intdig_max; /* We need space for the significant digits and perhaps for leading zeros when < 1.0. The number of leading zeros can be as many as would be required for exponential notation with a negative two-digit - p.exponent, which is 4. */ + p->exponent, which is 4. */ chars_needed = (size_t) dig_max + 1 + 4; } fracdig_min = info->alt ? fracdig_max : 0; @@ -760,26 +688,26 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, wcp = wstartp = wbuffer + 2; /* Let room for rounding. */ /* Do the real work: put digits in allocated buffer. */ - if (p.expsign == 0 || p.type != 'f') + if (p->expsign == 0 || p->type != 'f') { - assert (p.expsign == 0 || intdig_max == 1); + assert (p->expsign == 0 || intdig_max == 1); while (intdig_no < intdig_max) { ++intdig_no; - *wcp++ = hack_digit (&p); + *wcp++ = hack_digit (p); } significant = 1; if (info->alt || fracdig_min > 0 - || (fracdig_max > 0 && (p.fracsize > 1 || p.frac[0] != 0))) + || (fracdig_max > 0 && (p->fracsize > 1 || p->frac[0] != 0))) *wcp++ = decimal; } else { - /* |fp| < 1.0 and the selected p.type is 'f', so put "0." + /* |fp| < 1.0 and the selected p->type is 'f', so put "0." in the buffer. */ *wcp++ = '0'; - --p.exponent; + --p->exponent; *wcp++ = decimal; } @@ -787,10 +715,10 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, int fracdig_no = 0; int added_zeros = 0; while (fracdig_no < fracdig_min + added_zeros - || (fracdig_no < fracdig_max && (p.fracsize > 1 || p.frac[0] != 0))) + || (fracdig_no < fracdig_max && (p->fracsize > 1 || p->frac[0] != 0))) { ++fracdig_no; - *wcp = hack_digit (&p); + *wcp = hack_digit (p); if (*wcp++ != '0') significant = 1; else if (significant == 0) @@ -803,26 +731,26 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, /* Do rounding. */ char last_digit = wcp[-1] != decimal ? wcp[-1] : wcp[-2]; - char next_digit = hack_digit (&p); + char next_digit = hack_digit (p); bool more_bits; if (next_digit != '0' && next_digit != '5') more_bits = true; - else if (p.fracsize == 1 && p.frac[0] == 0) + else if (p->fracsize == 1 && p->frac[0] == 0) /* Rest of the number is zero. */ more_bits = false; - else if (p.scalesize == 0) + else if (p->scalesize == 0) { /* Here we have to see whether all limbs are zero since no normalization happened. */ - size_t lcnt = p.fracsize; - while (lcnt >= 1 && p.frac[lcnt - 1] == 0) + size_t lcnt = p->fracsize; + while (lcnt >= 1 && p->frac[lcnt - 1] == 0) --lcnt; more_bits = lcnt > 0; } else more_bits = true; int rounding_mode = get_rounding_mode (); - if (round_away (is_neg, (last_digit - '0') & 1, next_digit >= '5', + if (round_away (p->is_neg, (last_digit - '0') & 1, next_digit >= '5', more_bits, rounding_mode)) { char *wtp = wcp; @@ -842,7 +770,7 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, if (*wtp != decimal) /* Round up. */ (*wtp)++; - else if (__builtin_expect (spec == 'g' && p.type == 'f' && info->alt + else if (__builtin_expect (spec == 'g' && p->type == 'f' && info->alt && wtp == wstartp + 1 && wstartp[0] == '0', 0)) @@ -867,20 +795,20 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, else /* It is more critical. All digits were 9's. */ { - if (p.type != 'f') + if (p->type != 'f') { *wstartp = '1'; - p.exponent += p.expsign == 0 ? 1 : -1; + p->exponent += p->expsign == 0 ? 1 : -1; - /* The above p.exponent adjustment could lead to 1.0e-00, - e.g. for 0.999999999. Make sure p.exponent 0 always + /* The above p->exponent adjustment could lead to 1.0e-00, + e.g. for 0.999999999. Make sure p->exponent 0 always uses + sign. */ - if (p.exponent == 0) - p.expsign = 0; + if (p->exponent == 0) + p->expsign = 0; } else if (intdig_no == dig_max) { - /* This is the case where for p.type %g the number fits + /* This is the case where for p->type %g the number fits really in the range for %f output but after rounding the number of digits is too big. */ *--wstartp = decimal; @@ -896,9 +824,9 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, fracdig_no += intdig_no; intdig_no = 1; fracdig_max = intdig_max - intdig_no; - ++p.exponent; - /* Now we must print the p.exponent. */ - p.type = isupper (info->spec) ? 'E' : 'e'; + ++p->exponent; + /* Now we must print the p->exponent. */ + p->type = isupper (info->spec) ? 'E' : 'e'; } else { @@ -931,14 +859,14 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, if (fracdig_no == 0 && !info->alt && *(wcp - 1) == decimal) --wcp; - /* Write the p.exponent if it is needed. */ - if (p.type != 'f') + /* Write the p->exponent if it is needed. */ + if (p->type != 'f') { - if (__glibc_unlikely (p.expsign != 0 && p.exponent == 4 && spec == 'g')) + if (__glibc_unlikely (p->expsign != 0 && p->exponent == 4 && spec == 'g')) { - /* This is another special case. The p.exponent of the number is + /* This is another special case. The p->exponent of the number is really smaller than -4, which requires the 'e'/'E' format. - But after rounding the number has an p.exponent of -4. */ + But after rounding the number has an p->exponent of -4. */ assert (wcp >= wstartp + 1); assert (wstartp[0] == '1'); memcpy (wstartp, "0.0001", 6); @@ -953,26 +881,26 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, } else { - *wcp++ = p.type; - *wcp++ = p.expsign ? '-' : '+'; + *wcp++ = p->type; + *wcp++ = p->expsign ? '-' : '+'; - /* Find the magnitude of the p.exponent. */ + /* Find the magnitude of the p->exponent. */ expscale = 10; - while (expscale <= p.exponent) + while (expscale <= p->exponent) expscale *= 10; - if (p.exponent < 10) + if (p->exponent < 10) /* Exponent always has at least two digits. */ *wcp++ = '0'; else do { expscale /= 10; - *wcp++ = '0' + (p.exponent / expscale); - p.exponent %= expscale; + *wcp++ = '0' + (p->exponent / expscale); + p->exponent %= expscale; } while (expscale > 10); - *wcp++ = '0' + p.exponent; + *wcp++ = '0' + p->exponent; } } @@ -984,7 +912,7 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, /* Compute number of characters which must be filled with the padding character. */ - if (is_neg || info->showsign || info->space) + if (p->is_neg || info->showsign || info->space) --width; /* To count bytes, we would have to use __translated_number_width for info->i18n && !info->wide. See bug 28943. */ @@ -996,7 +924,7 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, if (!info->left && info->pad != '0') __printf_buffer_pad (buf, info->pad, width); - if (is_neg) + if (p->is_neg) __printf_buffer_putc (buf, '-'); else if (info->showsign) __printf_buffer_putc (buf, '+'); @@ -1029,6 +957,139 @@ __printf_fp_buffer_1 (struct __printf_buffer *buf, locale_t loc, free (wbuffer); } +#define PRINTF_FP_FETCH(P, FLOAT, SUFFIX, MANT_DIG) \ + ({ \ + FLOAT __number = *(const FLOAT *) args[0]; \ + \ + /* Check for special values: not a number or infinity. */ \ + if (isnan (__number)) \ + { \ + P.is_neg = signbit (__number); \ + if (isupper (info->spec)) \ + p.special = "NAN"; \ + else \ + p.special = "nan"; \ + } \ + else if (isinf (__number)) \ + { \ + P.is_neg = signbit (__number); \ + if (isupper (info->spec)) \ + p.special = "INF"; \ + else \ + p.special = "inf"; \ + } \ + else \ + { \ + P.fracsize = __mpn_extract_##SUFFIX \ + (P.fp_input, array_length (P.fp_input), \ + &P.exponent, &P.is_neg, __number); \ + P.to_shift = 1 + p.fracsize * BITS_PER_MP_LIMB - MANT_DIG; \ + } \ + }) + +static void +__printf_fp_buffer_double (struct __printf_buffer *buf, locale_t loc, + char thousands_sep, char decimal, + unsigned int thousands_sep_length, + const struct printf_info *info, + const void *const *args) +{ + /* For double values, the required extra memory is limited to total of + 432/480 bytes (3 times DBL_EXP_LIMB_SIZE on 32/64 bits), so it is + feasible to allocate the maximum size on stack. */ + enum + { + DBL_EXP_LIMB_SIZE = ((DBL_MAX_EXP + BITS_PER_MP_LIMB - 1) / BITS_PER_MP_LIMB + + (DBL_MANT_DIG / BITS_PER_MP_LIMB > 2 ? 8 : 4)) + }; + mp_limb_t frac[DBL_EXP_LIMB_SIZE]; + mp_limb_t tmp[DBL_EXP_LIMB_SIZE]; + mp_limb_t scale[DBL_EXP_LIMB_SIZE]; + struct hack_digit_param p = { + .frac = frac, .tmp = tmp, .scale = scale, .is_neg = 0, .to_shift = 0, + }; + + PRINTF_FP_FETCH (p, double, double, DBL_MANT_DIG); + __printf_fp_buffer_2 (&p, buf, loc, thousands_sep, decimal, + thousands_sep_length, info, args); +} + +#if __HAVE_DISTINCT_FLOAT128 || !defined __NO_LONG_DOUBLE_MATH +static void +__printf_fp_buffer_max (struct __printf_buffer *buf, locale_t loc, + char thousands_sep, char decimal, + unsigned int thousands_sep_length, + const struct printf_info *info, + const void *const *args) +{ + /* For long double / _Float128 the required extra memory can be up to + 6240 bytes (it is based on GREATER_MANT_DIG, which mihg beed to handle + _Float128), and this would require a large stack requirement. In this + case, first get the input exponent and calculate the required total + memory. The default stack scratch_buffer should be suffice for most + numbers. */ + + struct hack_digit_param p = { + .is_neg = 0, .to_shift = 0, + }; +#if __HAVE_DISTINCT_FLOAT128 + if (info->is_binary128) + PRINTF_FP_FETCH (p, _Float128, float128, FLT128_MANT_DIG); + else +#endif +#ifndef __NO_LONG_DOUBLE_MATH + if (info->is_long_double && sizeof (long double) > sizeof (double)) + PRINTF_FP_FETCH (p, long double, long_double, LDBL_MANT_DIG); +#endif + + mp_size_t bignum_size = ((abs (p.exponent) + BITS_PER_MP_LIMB - 1) + / BITS_PER_MP_LIMB + + (GREATER_MANT_DIG / BITS_PER_MP_LIMB > 2 + ? 8 : 4)) + * sizeof (mp_limb_t); + + struct scratch_buffer scbuf; + scratch_buffer_init (&scbuf); + if (!scratch_buffer_set_array_size (&scbuf, 3, bignum_size)) + { + __printf_buffer_mark_failed (buf); + return; + } + + p.frac = scbuf.data; + p.tmp = scbuf.data + bignum_size; + p.scale = scbuf.data + 2 * bignum_size; + + __printf_fp_buffer_2 (&p, buf, loc, thousands_sep, decimal, + thousands_sep_length, info, args); + + scratch_buffer_free (&scbuf); +} +#endif + +static void +__printf_fp_buffer_1 (struct __printf_buffer *buf, + locale_t loc, char thousands_sep, char decimal, + unsigned int thousands_sep_length, + const struct printf_info *info, + const void *const *args) +{ +#if __HAVE_DISTINCT_FLOAT128 + if (info->is_binary128) + return __printf_fp_buffer_max (buf, loc, thousands_sep, decimal, + thousands_sep_length, info, args); + else +#endif +#ifndef __NO_LONG_DOUBLE_MATH + if (info->is_long_double && sizeof (long double) > sizeof (double)) + return __printf_fp_buffer_max (buf, loc, thousands_sep, decimal, + thousands_sep_length, info, args); + else +#endif + return __printf_fp_buffer_double (buf, loc, thousands_sep, decimal, + thousands_sep_length, info, args); +} + /* ASCII to localization translation. Multibyte version. */ struct __printf_buffer_fp { From patchwork Thu Sep 14 14:06:39 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Adhemerval Zanella Netto X-Patchwork-Id: 722541 Delivered-To: patch@linaro.org Received: by 2002:ab3:5684:0:b0:23f:8cfb:5ad7 with SMTP id d4csp472192ltc; Thu, 14 Sep 2023 07:07:01 -0700 (PDT) X-Google-Smtp-Source: AGHT+IEzrW7n6EDE6AxGFXJfrsxTvkDHbz5wfVDWr/UYsKT2TIJjBVH7TPxKCiqBwzpqb9NchnkW X-Received: by 2002:a2e:8519:0:b0:2bc:ee89:6976 with SMTP id j25-20020a2e8519000000b002bcee896976mr4645795lji.24.1694700420990; Thu, 14 Sep 2023 07:07:00 -0700 (PDT) ARC-Seal: i=1; a=rsa-sha256; t=1694700420; cv=none; d=google.com; s=arc-20160816; b=05kENOKeC8r1P5bbENLY6rj7BoopGAMBrKhsR1ab7eT3z4TUDOs6VbeY7ad5wowL8N Ly8CavggQIimgc46pugcAHfLTlrn2nvQ1eVGnucTSnu5N1nZ5SvbPdHkwetv72P1OfMR fa4TLq8D6KnCiLnTx+7Ru/ucf9xRQNA1exBzuvLtr3uUqhgm2E6rPHoaee3zSLhgJXmB 3g9XAB5CxH6cF1nBGamU9LNUygVYwVmKcHyg4ydgQwJ02GNX7Zi2QIjYZpUibvA6WVhU m1Zy3px71aFC3/Ju3iXs2WU5XTJ7NWkpjoGhJ0Oe9b0hwsJQ+lHOuTbNdTtpUcUh/oNf evzg== ARC-Message-Signature: i=1; a=rsa-sha256; c=relaxed/relaxed; d=google.com; s=arc-20160816; h=errors-to:list-subscribe:list-help:list-post:list-archive :list-unsubscribe:list-id:precedence:content-transfer-encoding :mime-version:references:in-reply-to:message-id:date:subject:to:from :dkim-signature:dmarc-filter:delivered-to; bh=ioXOjBCRsRt6YbwiLABc+BgkQN2VttAD9+nXvIe9okM=; fh=Ji/BrrBV5vvGE0PSJoc8P6vA3gQfM/DANnlOBrK0hiY=; b=Bf2z9D+gxnvuwKngZ4tgR2OYCnChyPrsvGUL/T2W6PZBDK90HqLX5lsmuMD11fICUS giHMUhABoJsqj9xW/E+cHNmdoqDkASOelDDUN6TSpaDYmlUWUWk3RYQiS59XeXD5u8KA +TlrBMSefE8BgnDSAGfFu+AAKh0EfcDfqe87xzF2bAQOKFzfcsQ9OL+eiJWzw9BekpXh PG9rvJhgg4dgABtHo+Wan/TPBxKj6K+CPyh/dKI3VfUMzpl3Fno0zFKX8TMkeaaKj1vk 8wzVjD9USqW6FX7sN1PuyZJGIpjiekd1b3p33enhXrlADc5jx3KqnoOV1M530JkirfZI cwpw== ARC-Authentication-Results: i=1; mx.google.com; dkim=pass header.i=@linaro.org header.s=google header.b=y57sn1Os; spf=pass (google.com: domain of libc-alpha-bounces+patch=linaro.org@sourceware.org designates 8.43.85.97 as permitted sender) smtp.mailfrom="libc-alpha-bounces+patch=linaro.org@sourceware.org"; dmarc=pass (p=NONE sp=NONE dis=NONE) header.from=linaro.org Return-Path: Received: from server2.sourceware.org (ip-8-43-85-97.sourceware.org. [8.43.85.97]) by mx.google.com with ESMTPS id s27-20020a170906169b00b009aa146cba1csi1342543ejd.723.2023.09.14.07.07.00 for (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Thu, 14 Sep 2023 07:07:00 -0700 (PDT) Received-SPF: pass (google.com: domain of libc-alpha-bounces+patch=linaro.org@sourceware.org designates 8.43.85.97 as permitted sender) client-ip=8.43.85.97; Authentication-Results: mx.google.com; dkim=pass header.i=@linaro.org header.s=google header.b=y57sn1Os; spf=pass (google.com: domain of libc-alpha-bounces+patch=linaro.org@sourceware.org designates 8.43.85.97 as permitted sender) smtp.mailfrom="libc-alpha-bounces+patch=linaro.org@sourceware.org"; dmarc=pass (p=NONE sp=NONE dis=NONE) header.from=linaro.org Received: from server2.sourceware.org (localhost [IPv6:::1]) by sourceware.org (Postfix) with ESMTP id E42743855591 for ; Thu, 14 Sep 2023 14:06:59 +0000 (GMT) X-Original-To: libc-alpha@sourceware.org Delivered-To: libc-alpha@sourceware.org Received: from mail-oi1-x22a.google.com (mail-oi1-x22a.google.com [IPv6:2607:f8b0:4864:20::22a]) by sourceware.org (Postfix) with ESMTPS id C76F53857C41 for ; Thu, 14 Sep 2023 14:06:53 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.2 sourceware.org C76F53857C41 Authentication-Results: sourceware.org; dmarc=pass (p=none dis=none) header.from=linaro.org Authentication-Results: sourceware.org; spf=pass smtp.mailfrom=linaro.org Received: by mail-oi1-x22a.google.com with SMTP id 5614622812f47-3a9f88b657eso571224b6e.3 for ; Thu, 14 Sep 2023 07:06:53 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=linaro.org; s=google; t=1694700412; x=1695305212; darn=sourceware.org; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:to:from:from:to:cc:subject:date:message-id :reply-to; bh=ioXOjBCRsRt6YbwiLABc+BgkQN2VttAD9+nXvIe9okM=; b=y57sn1OsfkVMnYhLtNJJoyaNWfsXioRIMH23Q5UUsb9KQqExXjEQS/zjkz7hcDAoWD h7FrK8TzboV9chOqs9RXCco4vOLyUoSe7sLVpkpO3PLir/KfcnU63lUGEBOmD0boSo9C IszelB9eEZGvcS7zUAzSLS609UttkGFSwdGS7MrGhXaakfyS/6ztIu6Mc80D/S+3uVSR qzz35EDYirR5giW09h7U0Gi7iyAJomsSH3AjB3LwXFeQnqKk6f+K0EJacPDtmaV+jZUB 8o0vr/Gl/D/VwxflgWS8uYFdKQXNZjpBiK0jZ2OjnBCn3lTMUwZ7ZTDYhM+0mZ5EIH05 aAiA== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20230601; t=1694700412; x=1695305212; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:to:from:x-gm-message-state:from:to:cc :subject:date:message-id:reply-to; bh=ioXOjBCRsRt6YbwiLABc+BgkQN2VttAD9+nXvIe9okM=; b=Ss18Jqv5AtOGDoRjDKmzDbdtSQw8I3qpt43YWrzJSsMpW1LhueKDPxeh96Xm1e7Tje cSUwrMqwjV9LciFkrRSBDOBE6axlAvNXdVezR/+nj4IZSnaEcSzKZ6wsiw008hqNn8rE 9xR3ZKbRhWgkkhlkTVGuSNrBUlmcJBQCYGzBcug5P2v8VHqbzTITqukbkaD7uVQm9Mgl lwvo1DtBhR4CDnMWb0B1xpkLL6LTPu0ZBgfUK498R+5xuKDGVuoDz+aHeWP8g0uVhi8L 7jexxRuH988naAPPjWcFssckssx/qqQ0M8IubUsmPMgxnL3Ryn+XgEn9GUlM9kbOGbn/ 3yQw== X-Gm-Message-State: AOJu0YwJlfZPbL1+diYmi8di+BgQMlPnerYT0ytCS2FUcHdbCvswYOnK 8vuC+Z4m9eldEEXQ0gRQjJzvC2/Iwv2VnkWLUHSAfQ== X-Received: by 2002:a05:6808:1a1c:b0:3a9:eb53:3856 with SMTP id bk28-20020a0568081a1c00b003a9eb533856mr7412291oib.39.1694700412467; Thu, 14 Sep 2023 07:06:52 -0700 (PDT) Received: from mandiga.. ([2804:1b3:a7c0:91cb:5173:d70:d8d:4d2c]) by smtp.gmail.com with ESMTPSA id s26-20020a056808009a00b003a862e70bcbsm620417oic.13.2023.09.14.07.06.50 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Thu, 14 Sep 2023 07:06:51 -0700 (PDT) From: Adhemerval Zanella To: libc-alpha@sourceware.org, Joe Simmons-Talbott Subject: [PATCH v2 2/2] stdio: Remove __printf_fp_buffer_2 alloca usage Date: Thu, 14 Sep 2023 11:06:39 -0300 Message-Id: <20230914140639.1912983-3-adhemerval.zanella@linaro.org> X-Mailer: git-send-email 2.34.1 In-Reply-To: <20230914140639.1912983-1-adhemerval.zanella@linaro.org> References: <20230914140639.1912983-1-adhemerval.zanella@linaro.org> MIME-Version: 1.0 X-Spam-Status: No, score=-12.1 required=5.0 tests=BAYES_00, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, DKIM_VALID_EF, GIT_PATCH_0, KAM_SHORT, RCVD_IN_DNSWL_NONE, SPF_HELO_NONE, SPF_PASS, TXREP autolearn=ham autolearn_force=no version=3.4.6 X-Spam-Checker-Version: SpamAssassin 3.4.6 (2021-04-09) on server2.sourceware.org X-BeenThere: libc-alpha@sourceware.org X-Mailman-Version: 2.1.30 Precedence: list List-Id: Libc-alpha mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: libc-alpha-bounces+patch=linaro.org@sourceware.org And replace with a scratch_buffer. --- Changes from v1: * Fixed stdio-common regressions. --- stdio-common/printf_fp.c | 36 ++++++++++-------------------------- 1 file changed, 10 insertions(+), 26 deletions(-) diff --git a/stdio-common/printf_fp.c b/stdio-common/printf_fp.c index 9ddf61fd6d..7c85edb8c9 100644 --- a/stdio-common/printf_fp.c +++ b/stdio-common/printf_fp.c @@ -17,12 +17,8 @@ License along with the GNU C Library; if not, see . */ -/* The gmp headers need some configuration frobs. */ -#define HAVE_ALLOCA 1 - #include #include -#include #include #include #include @@ -168,11 +164,6 @@ __printf_fp_buffer_2 (struct hack_digit_param *p, struct __printf_buffer *buf, /* General helper (carry limb). */ mp_limb_t cy; - /* Buffer in which we produce the output. */ - char *wbuffer = NULL; - /* Flag whether wbuffer and buffer are malloc'ed or not. */ - int buffer_malloced = 0; - p->expsign = 0; if (p->special) @@ -600,6 +591,9 @@ __printf_fp_buffer_2 (struct hack_digit_param *p, struct __printf_buffer *buf, int dig_max; int significant; char spec = _tolower (info->spec); + struct scratch_buffer wscbuf; + + scratch_buffer_init (&wscbuf); if (spec == 'e') { @@ -662,7 +656,7 @@ __printf_fp_buffer_2 (struct hack_digit_param *p, struct __printf_buffer *buf, /* Allocate buffer for output. We need two more because while rounding it is possible that we need two more characters in front of all the other output. If the amount of memory we have to allocate is too - large use `malloc' instead of `alloca'. */ + large use `malloc' instead of a static buffer. */ if (__glibc_unlikely (chars_needed >= (size_t) -1 - 2 || chars_needed < fracdig_max)) { @@ -671,21 +665,12 @@ __printf_fp_buffer_2 (struct hack_digit_param *p, struct __printf_buffer *buf, __printf_buffer_mark_failed (buf); return; } - size_t wbuffer_to_alloc = 2 + chars_needed; - buffer_malloced = ! __libc_use_alloca (wbuffer_to_alloc); - if (__builtin_expect (buffer_malloced, 0)) + if (!scratch_buffer_set_array_size (&wscbuf, 1, 2 + chars_needed)) { - wbuffer = malloc (wbuffer_to_alloc); - if (wbuffer == NULL) - { - /* Signal an error to the caller. */ - __printf_buffer_mark_failed (buf); - return; - } + __printf_buffer_mark_failed (buf); + return; } - else - wbuffer = alloca (wbuffer_to_alloc); - wcp = wstartp = wbuffer + 2; /* Let room for rounding. */ + wcp = wstartp = wscbuf.data + 2; /* Let room for rounding. */ /* Do the real work: put digits in allocated buffer. */ if (p->expsign == 0 || p->type != 'f') @@ -951,10 +936,9 @@ __printf_fp_buffer_2 (struct hack_digit_param *p, struct __printf_buffer *buf, if (info->left) __printf_buffer_pad (buf, info->pad, width); - } - if (buffer_malloced) - free (wbuffer); + scratch_buffer_free (&wscbuf); + } } #define PRINTF_FP_FETCH(P, FLOAT, SUFFIX, MANT_DIG) \