Merge remote-tracking branch 'origin/dev' into HEAD

picolibc 1.8.10 (and its main) read them as hex digits, so "0x1:" was 26 and
parse_number<float> accepted it; newlib and the C standard stop there.

esphome/core/helpers.cpp

@@ -337,57 +337,6 @@ size_t truncate_name_to(char *buffer, size_t buffer_size, const char *name, size
 
 // Parsing & formatting
 
-// A plain decimal converts exactly with one division: the digits accumulate
-// into an integer that stays below 2^24 and the fraction's scale is a power of
-// ten up to 1e9 (2^9 * 5^9), so both convert to float exactly and IEEE division
-// rounds the quotient correctly, the same result strtof produces.
-static constexpr uint32_t PARSE_FLOAT_MANTISSA_LIMIT = 1000000;  // one more digit stays below 2^24
-static constexpr uint32_t PARSE_FLOAT_MAX_FRACTION_DIGITS = 9;
-
-optional<float> parse_float(const char *str) {
-  const char *p = str;
-  const bool negative = *p == '-';
-  if (negative || *p == '+')
-    p++;
-  const char *digits = p;
-  const char *dot = nullptr;
-  uint32_t mantissa = 0;
-  for (;; p++) {
-    const uint32_t d = static_cast<unsigned char>(*p) - '0';
-    if (d > 9) {
-      if (*p == '.' && dot == nullptr) {
-        dot = p;
-        continue;
-      }
-      break;
-    }
-    if (mantissa >= PARSE_FLOAT_MANTISSA_LIMIT)
-      break;
-    mantissa = mantissa * 10 + d;
-  }
-  const uint32_t fraction_digits = dot != nullptr ? p - dot - 1 : 0;
-  const bool any_digit = p - digits > (dot != nullptr ? 1 : 0);
-  if (*p == '\0' && any_digit && fraction_digits <= PARSE_FLOAT_MAX_FRACTION_DIGITS) {
-    // The mantissa is exact in a float, so the sign goes on before the one rounding step
-    float value = static_cast<float>(mantissa);
-    if (negative)
-      value = -value;
-    if (fraction_digits != 0) {
-      uint32_t scale = 10;
-      for (uint32_t i = 1; i < fraction_digits; i++)
-        scale *= 10;
-      value /= static_cast<float>(scale);
-    }
-    return value;
-  }
-  // Exponents, hex, inf, nan, leading whitespace and longer inputs take the libc path.
-  char *end = nullptr;
-  const float value = ::strtof(str, &end);
-  if (end == str || *end != '\0' || value == HUGE_VALF)
-    return {};
-  return value;
-}
-
 size_t parse_hex(const char *str, size_t length, uint8_t *data, size_t count) {
   size_t chars = std::min(length, 2 * count);
   for (size_t i = 2 * count - chars; i < 2 * count; i++, str++) {

esphome/core/helpers.h

@@ -25,6 +25,7 @@
 #include <strings.h>
 
 #include "esphome/core/optional.h"
+#include "esphome/core/parse_float.h"
 #include "esphome/core/time_conversion.h"
 
 #ifdef USE_ESP8266
@@ -1033,6 +1034,26 @@ inline float pow10_int(int8_t exp) {
   return result;
 }
 
+/// 10^n, exact while it fits in T.
+template<typename T> constexpr T pow10_exact(uint32_t n) {
+  T result = 1;
+  T base = 10;
+  for (;;) {
+    if ((n & 1) != 0)
+      result *= base;
+    n >>= 1;
+    if (n == 0)
+      return result;
+    base *= base;
+  }
+}
+
+/// Whether c is '0' to '9'.
+constexpr bool is_digit(char c) { return static_cast<uint8_t>(c - '0') <= 9; }
+
+/// Whether c is C-locale white space, without isspace's ctype table.
+constexpr bool is_space_c(char c) { return c == ' ' || (c >= '\t' && c <= '\r'); }
+
 /// Remap \p value from the range (\p min, \p max) to (\p min_out, \p max_out).
 template<typename T, typename U> T remap(U value, U min, U max, T min_out, T max_out) {
   return (value - min) * (max_out - min_out) / (max - min) + min_out;
@@ -1573,8 +1594,6 @@ template<typename T, enable_if_t<(std::is_integral<T>::value && std::is_signed<T
 optional<T> parse_number(const std::string &str) {
   return parse_number<T>(str.c_str());
 }
-/// Parse a decimal floating-point number; plain decimals skip the libc parser.
-optional<float> parse_float(const char *str);
 /// Parse a decimal floating-point number from a null-terminated string.
 template<typename T, enable_if_t<(std::is_same<T, float>::value), int> = 0> optional<T> parse_number(const char *str) {
   return parse_float(str);

esphome/core/parse_float.cpp

@@ -0,0 +1,372 @@
+// picolibc 1.8.10's strtof, bit for bit, without libc's strtod: decimals round to 9 digits, then to the nearest float.
+// Float bits are built from integers, so chips that flush denormals give the same result.
+
+#include "esphome/core/parse_float.h"
+#include "esphome/core/hal.h"
+#include "esphome/core/helpers.h"
+
+#include <algorithm>
+#include <bit>
+
+namespace esphome {
+
+namespace {
+
+constexpr uint32_t FLOAT_SIGN_BIT = 0x80000000, FLOAT_INF_BITS = 0x7F800000;
+constexpr uint32_t FLOAT_NAN_BITS = 0x7FC00000;  // nanf(""), which strtof returns for every NaN
+
+constexpr uint32_t DECIMAL_DIGITS = 9, HEX_DIGITS = 8;  // digits kept before the rest only round
+constexpr int32_t EXPONENT_CAP = 224;  // picolibc stops adding exponent digits once the exponent reaches this
+// 0.<digits> * 10^exp10 is zero at or below MIN_EXP10 and infinity at or above MAX_EXP10, as in picolibc
+constexpr int32_t MIN_EXP10 = -46, MAX_EXP10 = 40;
+// Past this, one more digit could pass 2^24
+constexpr uint32_t PLAIN_MANTISSA_LIMIT = 1677720;
+
+// Bits of the float nearest to (m + sticky fraction) * 2^e, ties to even; m > 0
+uint32_t round_to_float_bits(uint64_t m, int32_t e, bool sticky) {
+  const int32_t msb = e + static_cast<int32_t>(std::bit_width(m)) - 1;  // the value is in [2^msb, 2^(msb + 1))
+  if (msb >= 128)
+    return FLOAT_INF_BITS;
+  if (msb < -150)  // below half the smallest float
+    return 0;
+  // The unit of the float's last bit: 2^(msb - 23), or 2^-149 for a denormal
+  const int32_t shift = std::max<int32_t>(msb - 23, -149) - e;  // at most 64
+  uint64_t q;
+  if (shift <= 0) {
+    q = m << -shift;
+  } else {
+    q = shift < 64 ? m >> shift : 0;
+    const uint64_t rem = shift < 64 ? m & ((uint64_t{1} << shift) - 1) : m;
+    const uint64_t half = uint64_t{1} << (shift - 1);
+    if (rem > half || (rem == half && (sticky || (q & 1) != 0)))
+      q++;
+  }
+  if (msb < -126)
+    return static_cast<uint32_t>(q);  // 2^23 here is the smallest normal float, as it should be
+  // Rounding up past the largest float gives exactly FLOAT_INF_BITS
+  return (static_cast<uint32_t>(msb + 127) << 23) + static_cast<uint32_t>(q - (uint64_t{1} << 23));
+}
+
+// The decimal exponents that reach decimal_to_float_bits: 1 to 9 digits and MIN_EXP10 < digits + e10 < MAX_EXP10
+constexpr int32_t MIN_E10 = MIN_EXP10 + 1 - static_cast<int32_t>(DECIMAL_DIGITS), MAX_E10 = MAX_EXP10 - 2;
+// Top 64 bits of 10^(8k), rounded down: 10^(8k) = entry * 2^(floor_log2_pow10(8k) - 63), exact for k = 0..3
+constexpr int32_t POW10_BY_8_FIRST = MIN_E10 >> 3, POW10_BY_8_LAST = MAX_E10 >> 3;
+constexpr uint64_t POW10_BY_8[] PROGMEM = {
+    0xFB158592BE068D2E, 0xBB127C53B17EC159, 0x8B61313BBABCE2C6, 0xCFB11EAD453994BA,
+    0x9ABE14CD44753B52, 0xE69594BEC44DE15B, 0xABCC77118461CEFC, 0x8000000000000000,
+    0xBEBC200000000000, 0x8E1BC9BF04000000, 0xD3C21BCECCEDA100, 0x9DC5ADA82B70B59D,
+};
+static_assert(POW10_BY_8_FIRST == -7 && POW10_BY_8_LAST == 4 &&
+                  sizeof(POW10_BY_8) / sizeof(POW10_BY_8[0]) == POW10_BY_8_LAST - POW10_BY_8_FIRST + 1,
+              "POW10_BY_8 must cover MIN_E10..MAX_E10");
+
+// floor(q * log2(10)), for |q| < 100
+constexpr int32_t floor_log2_pow10(int32_t q) { return (q * 217706) >> 16; }
+
+// The top 64 bits of a * b, rounded down
+uint64_t mul_high(uint64_t a, uint64_t b) {
+  const uint64_t a_lo = static_cast<uint32_t>(a), a_hi = a >> 32;
+  const uint64_t b_lo = static_cast<uint32_t>(b), b_hi = b >> 32;
+  const uint64_t hi_lo = a_hi * b_lo, lo_hi = a_lo * b_hi;
+  const uint64_t middle = ((a_lo * b_lo) >> 32) + static_cast<uint32_t>(hi_lo) + static_cast<uint32_t>(lo_hi);
+  return a_hi * b_hi + (hi_lo >> 32) + (lo_hi >> 32) + (middle >> 32);
+}
+
+// Cold: a decimal within 16 units of a midpoint is an exact tie when m10 * 5^e10 is an integer
+__attribute__((noinline, cold)) uint32_t tie_float_bits(uint32_t m10, int32_t e10, uint32_t bits) {
+  if (e10 >= 0) {
+    if (e10 <= 14)  // m10 * 5^14 stays below 2^64
+      return round_to_float_bits(m10 * (pow10_exact<uint64_t>(e10) >> e10), e10, false);
+  } else if (e10 >= -13) {
+    const auto pow5 = static_cast<uint32_t>(pow10_exact<uint64_t>(-e10) >> -e10);
+    if (m10 % pow5 == 0)
+      return round_to_float_bits(m10 / pow5, e10, false);
+  }
+  return bits;
+}
+
+// The float nearest to m10 * 10^e10 (0 < m10 <= 10^9). The product is under 4 units low; checked for all 85 billion
+// (m10, e10), only exact ties and one non-tie (6 units off, keeps its float) come within 34 units of a midpoint.
+uint32_t decimal_to_float_bits(uint32_t m10, int32_t e10) {
+  // Both operands are exact floats, so the one float operation gives the nearest float
+  const uint32_t power = e10 < 0 ? -e10 : e10;
+  if (m10 < (1u << 24) && power <= 10) {
+    const auto value = static_cast<float>(m10);
+    // 10^10 overflows uint32_t
+    const float factor = power == 10 ? 1e10f : static_cast<float>(pow10_exact<uint32_t>(power));
+    return bit_cast<uint32_t>(e10 < 0 ? value / factor : value * factor);
+  }
+  const int32_t k8 = e10 >> 3;  // rounds down
+  const uint64_t x = uint64_t{m10} * pow10_exact<uint32_t>(e10 & 7);
+  const auto zeros = static_cast<int32_t>(std::countl_zero(x));
+  uint64_t m = x << zeros;
+  int32_t e = -zeros;
+  if (k8 == 0)
+    return round_to_float_bits(m, e, false);  // exact
+  m = mul_high(m, POW10_BY_8[k8 - POW10_BY_8_FIRST]);
+  e += floor_log2_pow10(8 * k8) + 1;
+  if ((m >> 63) == 0) {
+    m <<= 1;
+    e--;
+  }
+  const uint32_t bits = round_to_float_bits(m, e, false);
+  // The bits below the float's last one, and the midpoint among them
+  const int32_t shift = std::max<int32_t>(40, -149 - e);
+  if (shift <= 64) {
+    const uint64_t low = shift < 64 ? m & ((uint64_t{1} << shift) - 1) : m;
+    const uint64_t half = uint64_t{1} << (shift - 1);
+    if (low - half + 16 > 32)  // |low - half| > 16
+      return bits;
+  }
+  return tie_float_bits(m10, e10, bits);
+}
+
+uint32_t no_number(const char *str, const char **end, bool *negative) {
+  *end = str;
+  *negative = false;
+  return 0;
+}
+
+// The rest of an "nan" or "inf" after its first letter at p; returns the end, or nullptr for no number
+const char *read_inf_nan(const char *p) {
+  const auto lower = [](char c) { return static_cast<char>(c | 0x20); };
+  if (lower(*p) == 'n') {
+    if (lower(p[1]) != 'a' || lower(p[2]) != 'n')
+      return nullptr;
+    p += 3;
+    if (*p != '(')
+      return p;
+    // "nan(" must be closed after letters, digits and '_'
+    do {
+      p++;
+    } while (static_cast<uint8_t>(lower(*p) - 'a') < 26 || is_digit(*p) || *p == '_');
+    return *p == ')' ? p + 1 : nullptr;
+  }
+  if (lower(p[1]) != 'n' || lower(p[2]) != 'f')
+    return nullptr;
+  p += 3;
+  if (lower(*p) != 'i')
+    return p;
+  // Past "infi", only "infinity" is a number
+  return lower(p[1]) == 'n' && lower(p[2]) == 'i' && lower(p[3]) == 't' && lower(p[4]) == 'y' ? p + 5 : nullptr;
+}
+
+// Reads a binary exponent "p" or decimal exponent "e" at p into *exp; returns the end of the number
+const char *read_exponent(const char *p, char letter, int32_t *exp) {
+  if ((*p | 0x20) != letter)
+    return p;
+  const char *q = p + 1;
+  const bool negative = *q == '-';
+  if (negative || *q == '+')
+    q++;
+  if (!is_digit(*q))
+    return p;
+  int32_t value = 0;
+  for (; is_digit(*q); q++) {
+    if (value < EXPONENT_CAP)
+      value = value * 10 + (*q - '0');
+  }
+  *exp += negative ? -value : value;
+  return q;
+}
+
+// Hex digits after the "x" at p, with the point and decimal exponent read so far (no non-zero digit yet)
+__attribute__((noinline, cold)) uint32_t read_hex(const char *p, bool dot, int32_t exp, const char **end) {
+  uint32_t m = 0, kept = 0;  // kept: digits in m from its first non-zero one
+  bool sticky = false;
+  for (;; p++) {
+    const char lower = static_cast<char>(*p | 0x20);
+    uint32_t digit = static_cast<uint8_t>(*p - '0');
+    if (static_cast<uint8_t>(lower - 'a') <= 5) {
+      digit = static_cast<uint8_t>(lower - 'a') + 10;
+    } else if (digit > 9) {
+      digit = 16;  // picolibc takes ':' to '?' as 10 to 15; we stop there like newlib
+    }
+    if (digit < 16) {
+      if (kept < HEX_DIGITS) {
+        if (dot)
+          exp--;
+        m = m * 16 + digit;
+        kept += m != 0 ? 1 : 0;
+      } else {
+        if (!dot)
+          exp++;
+        sticky |= digit != 0;
+      }
+    } else if (*p == '.' && !dot) {
+      dot = true;
+    } else if (lower != 'x' || m != 0) {  // picolibc takes another "x" while the value is still zero
+      break;
+    }
+  }
+  exp *= 4;
+  *end = read_exponent(p, 'p', &exp);
+  return m == 0 ? 0 : round_to_float_bits(m, exp, sticky);
+}
+
+// picolibc's conv_flt for a float from the first digit or point at p: returns the bits; *end is str for no number
+__attribute__((noinline)) uint32_t read_decimal(const char *str, const char *p, const char **end) {
+  uint32_t m = 0, kept = 0;  // the first (up to 9) significant digits, and how many from the first non-zero one
+  int32_t exp = 0;           // the value is m * 10^exp
+  bool dot = false, any = false;
+  uint32_t overflow = 0;  // digits past the kept ones: 0 none, 1 below half, 2 half, 3 above half
+  bool full = false;      // a digit past the kept ones was read
+  for (;; p++) {
+    const uint32_t digit = static_cast<uint8_t>(*p - '0');
+    if (digit <= 9) {
+      any = true;
+      if (kept < DECIMAL_DIGITS) {
+        if (dot)
+          exp--;
+        m = m * 10 + digit;
+        kept += m != 0 ? 1 : 0;
+      } else {
+        if (!dot)
+          exp++;
+        if (!full) {
+          full = true;
+          overflow = digit > 5 ? 3 : digit == 5 ? 2 : digit != 0 ? 1 : 0;  // the first one decides the half
+        } else {
+          overflow |= digit != 0 ? 1 : 0;
+        }
+      }
+    } else if (*p == '.' && !dot) {
+      dot = true;
+    } else {
+      break;
+    }
+  }
+  if ((*p | 0x20) == 'x' && any && m == 0)
+    return read_hex(p + 1, dot, exp, end);
+  if (!any) {
+    *end = str;
+    return 0;
+  }
+  *end = read_exponent(p, 'e', &exp);
+  if (m == 0)
+    return 0;
+  // Ties to even on the decimal digits
+  overflow |= m & 1;
+  m += (overflow + 1) >> 2;
+  const int32_t exp10 = static_cast<int32_t>(kept) + exp;
+  if (exp10 <= MIN_EXP10)
+    return 0;
+  if (exp10 >= MAX_EXP10)
+    return FLOAT_INF_BITS;
+  return decimal_to_float_bits(m, exp);
+}
+
+// picolibc's strtof from the start: returns the bits without the sign, which goes in *negative
+__attribute__((noinline)) uint32_t parse_float_full(const char *str, const char **end, bool *negative) {
+  // No number starts with a letter but i (inf) or n (nan): "unavailable", "on" and "unknown" stop here
+  const char first = static_cast<char>(*str | 0x20);
+  if (first >= 'a' && first <= 'z' && first != 'i' && first != 'n')
+    return no_number(str, end, negative);
+  const char *p = str;
+  while (is_space_c(*p))
+    p++;
+  *negative = *p == '-';
+  if (*negative || *p == '+')
+    p++;
+  const char lower = static_cast<char>(*p | 0x20);
+  if (lower == 'i' || lower == 'n') {
+    const char *stop = read_inf_nan(p);
+    if (stop == nullptr)
+      return no_number(str, end, negative);
+    *end = stop;
+    return lower == 'i' ? FLOAT_INF_BITS : FLOAT_NAN_BITS;
+  }
+  const uint32_t bits = read_decimal(str, p, end);
+  if (*end == str)
+    *negative = false;
+  return bits;
+}
+
+// A plain decimal below 2^24 with up to 9 digits after the point, in one float operation; nullptr otherwise
+ESPHOME_ALWAYS_INLINE inline const char *parse_plain_decimal(const char *str, uint32_t *bits, bool *negative) {
+  const char *p = str;
+  *negative = *p == '-';
+  if (*negative || *p == '+')
+    p++;
+  const char *digits = p;
+  const char *point = nullptr;
+  uint32_t mantissa = 0;
+  for (;; p++) {
+    const uint32_t d = static_cast<uint8_t>(*p) - '0';
+    if (d > 9) {
+      if (*p == '.' && point == nullptr) {
+        point = p;
+        continue;
+      }
+      break;
+    }
+    if (mantissa > PLAIN_MANTISSA_LIMIT)
+      return nullptr;
+    mantissa = mantissa * 10 + d;
+  }
+  const uint32_t fraction = point != nullptr ? p - point - 1 : 0;
+  const char c = static_cast<char>(*p | 0x20);
+  if (p - digits == (point != nullptr ? 1 : 0) || fraction > 9 || c == 'e' || c == 'x')
+    return nullptr;
+  float value = static_cast<float>(mantissa);
+  if (fraction != 0)
+    value /= static_cast<float>(pow10_exact<uint32_t>(fraction));  // 10^9 is an exact float
+  *bits = bit_cast<uint32_t>(value);
+  return p;
+}
+
+}  // namespace
+
+optional<float> parse_float(const char *str) {
+  // Own copy of the plain-decimal loop: sharing parse_plain_decimal costs 12 to 18% more instructions
+  const char *p = str;
+  const bool negative = *p == '-';
+  if (negative || *p == '+')
+    p++;
+  const char *digits = p;
+  const char *point = nullptr;
+  uint32_t mantissa = 0;
+  for (;; p++) {
+    const uint32_t d = static_cast<uint8_t>(*p) - '0';
+    if (d > 9) {
+      if (*p == '.' && point == nullptr) {
+        point = p;
+        continue;
+      }
+      break;
+    }
+    if (mantissa > PLAIN_MANTISSA_LIMIT)
+      break;
+    mantissa = mantissa * 10 + d;
+  }
+  const uint32_t fraction = point != nullptr ? p - point - 1 : 0;
+  if (*p == '\0' && p - digits > (point != nullptr ? 1 : 0) && fraction <= 9) {
+    float value = negative ? -static_cast<float>(mantissa) : static_cast<float>(mantissa);
+    if (fraction != 0)
+      value /= static_cast<float>(pow10_exact<uint32_t>(fraction));
+    return value;
+  }
+  const char *end;
+  bool full_negative;
+  const uint32_t bits = parse_float_full(str, &end, &full_negative);
+  // HUGE_VALF, positive overflow or "inf", is rejected like before
+  if (end == str || *end != '\0' || (!full_negative && bits == FLOAT_INF_BITS))
+    return {};
+  return bit_cast<float>(full_negative ? bits | FLOAT_SIGN_BIT : bits);
+}
+
+float str_to_float(const char *str, char **end) {
+  const char *stop;
+  bool negative;
+  uint32_t bits;
+  if (const char *plain_end = parse_plain_decimal(str, &bits, &negative)) {
+    stop = plain_end;
+  } else {
+    bits = parse_float_full(str, &stop, &negative);
+  }
+  if (end != nullptr)
+    *end = const_cast<char *>(stop);  // strtof's signature
+  return bit_cast<float>(negative ? bits | FLOAT_SIGN_BIT : bits);
+}
+
+}  // namespace esphome

esphome/core/parse_float.h

@@ -0,0 +1,17 @@
+#pragma once
+
+#include <cstdint>
+
+#include "esphome/core/optional.h"
+
+namespace esphome {
+
+/// The whole string as str_to_float's syntax (leading space, hex, inf, nan); trailing characters, "inf" and positive
+/// overflow are rejected, while -inf, negative overflow and NaN are accepted.
+optional<float> parse_float(const char *str);
+
+/// strtof with picolibc's results on every platform, but it never sets errno (overflow gives +-HUGE_VALF, underflow 0);
+/// use it instead of strtof, or strtod/atof for a float result.
+float str_to_float(const char *str, char **end);
+
+}  // namespace esphome

esphome/core/string_ref.h

@@ -7,6 +7,7 @@
 #include <memory>
 #include <string>
 #include "esphome/core/defines.h"
+#include "esphome/core/parse_float.h"
 
 #ifdef USE_JSON
 #include "esphome/components/json/json_util.h"
@@ -293,7 +294,7 @@ inline bool operator!=(const std::string &lhs, const ProgmemStringRef &rhs) { re
 #endif
 
 // String conversion functions for ADL compatibility (allows stoi(x) where x is StringRef)
-// Must be in esphome namespace for ADL to find them. Uses strtol/strtod directly to avoid heap allocation.
+// Must be in esphome namespace for ADL to find them. Uses strtol, strtod and str_to_float to avoid heap allocation.
 namespace internal {
 // NOLINTBEGIN(google-runtime-int)
 template<typename R, typename F> inline R parse_number(const StringRef &str, size_t *pos, F conv) {
@@ -334,7 +335,7 @@ inline long stol(const StringRef &str, size_t *pos = nullptr, int base = 10) {
   return internal::parse_number<long>(str, pos, base, std::strtol);
 }
 inline float stof(const StringRef &str, size_t *pos = nullptr) {
-  return internal::parse_number<float>(str, pos, std::strtof);
+  return internal::parse_number<float>(str, pos, str_to_float);
 }
 inline double stod(const StringRef &str, size_t *pos = nullptr) {
   return internal::parse_number<double>(str, pos, std::strtod);

tests/benchmarks/core/bench_parse_float.cpp

@@ -0,0 +1,40 @@
+#include <benchmark/benchmark.h>
+
+#include "esphome/core/helpers.h"
+
+namespace esphome::benchmarks {
+
+// Inner iteration count to amortize CodSpeed instrumentation overhead.
+static constexpr int kInnerIterations = 2000;
+
+static const char *const PLAIN[] = {"0.5", "255", "-12.75", "21.3"};
+static const char *const EXPONENT[] = {"1e3", "2.5e-2", "-6.02E5", "1.5e+1"};
+static const char *const PRECISE[] = {"3.14159265358979", "0.000123456789", "123456789.125", "-0.1000000000000000055"};
+// Home Assistant states that are not numbers
+static const char *const REJECT[] = {"unavailable", "unknown", "on", "off"};
+
+template<size_t N, typename F> static void run(benchmark::State &state, const char *const (&inputs)[N], F parse) {
+  for (auto _ : state) {
+    float sum = 0.0f;
+    for (int i = 0; i < kInnerIterations; i++)
+      sum += parse(inputs[i % N]);
+    benchmark::DoNotOptimize(sum);
+  }
+  state.SetItemsProcessed(state.iterations() * kInnerIterations);
+}
+
+static float parse_number_float(const char *str) { return parse_number<float>(str).value_or(0.0f); }
+
+static void ParseNumberFloat_Plain(benchmark::State &state) { run(state, PLAIN, parse_number_float); }
+BENCHMARK(ParseNumberFloat_Plain);
+
+static void ParseNumberFloat_Exponent(benchmark::State &state) { run(state, EXPONENT, parse_number_float); }
+BENCHMARK(ParseNumberFloat_Exponent);
+
+static void ParseNumberFloat_Precise(benchmark::State &state) { run(state, PRECISE, parse_number_float); }
+BENCHMARK(ParseNumberFloat_Precise);
+
+static void ParseNumberFloat_Reject(benchmark::State &state) { run(state, REJECT, parse_number_float); }
+BENCHMARK(ParseNumberFloat_Reject);
+
+}  // namespace esphome::benchmarks

tests/components/core/test_parse_float.cpp

@@ -1,34 +1,74 @@
 #include <gtest/gtest.h>
 
+#include <cmath>
 #include <cstdint>
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
+#include <random>
 #include <string>
 
 #include "esphome/core/helpers.h"
+#include "esphome/core/parse_float.h"
 
+#ifdef USE_HOST
 namespace esphome::core::testing {
 
-static uint32_t bits(float value) {
-  uint32_t out;
-  memcpy(&out, &value, sizeof(out));
-  return out;
+static uint32_t bits(float value) { return bit_cast<uint32_t>(value); }
+
+// picolibc rounds a decimal to 9 significant digits and converts those to the nearest float, so up to 9
+// significant digits its result is the correctly rounded float the host's strtof (glibc, macOS) gives. Longer
+// input and the syntax where libcs differ are checked against values taken from picolibc 1.8.10's own strtof.
+static float reference_strtof(const char *str, char **end) {
+  const float value = ::strtof(str, end);
+  if (std::isnan(value))
+    return std::signbit(value) ? -NAN : NAN;
+  return value;
 }
 
-// strtof is the reference: every accepted input must give the same bits.
+// Every accepted input must give the reference's bits.
 static void expect_matches_strtof(const char *str) {
-  char *end = nullptr;
-  const float expected = ::strtof(str, &end);
+  const float expected = reference_strtof(str, nullptr);
   const auto parsed = parse_number<float>(str);
   if (!parsed.has_value()) {
     ADD_FAILURE() << str << " was rejected";
     return;
   }
-  const float actual = parsed.value();
-  EXPECT_EQ(bits(actual), bits(expected)) << str << " parsed " << actual << " expected " << expected;
+  EXPECT_EQ(bits(parsed.value()), bits(expected)) << str;
 }
 
+// The same bits and end as the host's strtof; for decimals of up to 9 significant digits
+static void expect_str_to_float_matches(const char *str) {
+  char *expected_end = nullptr;
+  const float expected = reference_strtof(str, &expected_end);
+  char *end = nullptr;
+  const float actual = str_to_float(str, &end);
+  EXPECT_EQ(bits(actual), bits(expected)) << '"' << str << '"';
+  EXPECT_EQ(end - str, expected_end - str) << '"' << str << '"';
+}
+
+struct PicolibcCase {
+  const char *str;
+  uint32_t bits;
+  int32_t end;
+};
+
+// Bits and end from picolibc 1.8.10's strtof
+static void expect_picolibc(const PicolibcCase &c) {
+  char *end = nullptr;
+  const float value = str_to_float(c.str, &end);
+  EXPECT_EQ(bits(value), c.bits) << '"' << c.str << '"';
+  EXPECT_EQ(end - c.str, c.end) << '"' << c.str << '"';
+}
+
+static void expect_picolibc(const std::string &str, uint32_t expected) {
+  char *end = nullptr;
+  EXPECT_EQ(bits(str_to_float(str.c_str(), &end)), expected) << str.substr(0, 40);
+  EXPECT_EQ(static_cast<size_t>(end - str.c_str()), str.size()) << str.substr(0, 40);
+}
+
+static float float_from(uint32_t value) { return bit_cast<float>(value); }
+
 TEST(ParseFloat, PlainDecimalsMatchStrtof) {
   for (const char *str :
        {"0",     "-0",      "+0",        "1",        "-1",          "100.50",      "99.75",       "0.1", "0.3",
@@ -46,6 +86,14 @@ TEST(ParseFloat, SweepOfTwoDecimalValuesMatchesStrtof) {
   }
 }
 
+// Eight-digit mantissas below 2^24 stay on the one-division path; the next digit past the limit falls back
+TEST(ParseFloat, EightDigitPlainDecimalsMatchStrtof) {
+  for (const char *str : {"12345678", "-12345678", "1234.5678", "0.12345678", "16777209", "1677720.9", "0.16777209",
+                          "16777210", "16777215", "1677721.5", "99999999", "9999999.9", "16777209."}) {
+    expect_matches_strtof(str);
+  }
+}
+
 // A fraction longer than the fast path accepts must fall back, however long it is.
 TEST(ParseFloat, LongFractionsMatchStrtof) {
   std::string str = "0.";
@@ -68,10 +116,293 @@ TEST(ParseFloat, LibcPathInputsMatchStrtof) {
 // Positive overflow and a positive literal infinity both come back as HUGE_VALF from strtof, and the parser
 // rejects both; the negative forms pass through, as they did before the fast path.
 TEST(ParseFloat, RejectsInvalidInput) {
-  for (const char *str :
-       {"", "-", "+", ".", "abc", "12abc", "1.2.3", "1,5", "12 ", "--1", "1e", "1e400", "inf", "infinity"}) {
+  for (const char *str : {"", "-", "+", ".", "abc", "12abc", "1.2.3", "1,5", "12 ", "--1", "1e", "1e400", "inf",
+                          "infinity", "infi", "nan(12 ab)", "nan(12"}) {
     EXPECT_FALSE(parse_number<float>(str).has_value()) << str;
   }
 }
 
+// Above the midpoint 1 + 2^-24 by less than half a double: newlib's strtof rounds to the midpoint and then to
+// 1.0f; picolibc's 9 digits, 1.00000006, give the next float.
+TEST(ParseFloat, FullPathRoundsLikePicolibc) {
+  EXPECT_EQ(bits(parse_number<float>("1.0000000596046448").value()), 0x3F800001u);
+  EXPECT_EQ(bits(parse_number<float>("-1.0000000596046448").value()), 0xBF800001u);
+}
+
+// picolibc takes "nan(" with letters, digits and '_' up to a ")", and nothing at all otherwise
+TEST(ParseFloat, NanForms) {
+  EXPECT_TRUE(std::isnan(parse_number<float>("nan(12)").value()));
+  EXPECT_TRUE(std::isnan(parse_number<float>("nan(xyz_1)").value()));
+  EXPECT_FALSE(parse_number<float>("nan(x y)").has_value());
+}
+
+// The callers' overflow check: a positive overflow is HUGE_VALF and rejected, a negative one is -inf and accepted
+TEST(ParseFloat, OverflowIsHugeValf) {
+  EXPECT_FALSE(parse_number<float>("1e39").has_value());
+  EXPECT_FALSE(parse_number<float>("3.40282357e38").has_value());
+  EXPECT_FALSE(parse_number<float>("1e99999").has_value());
+  EXPECT_EQ(bits(parse_number<float>("-1e39").value()), 0xFF800000u);
+  EXPECT_EQ(bits(parse_number<float>("3.4028235e38").value()), 0x7F7FFFFFu);
+}
+
+// Up to 9 significant digits, with leading and trailing zeros, a point anywhere and an exponent
+TEST(StrToFloat, RandomDecimalsMatchStrtof) {
+  std::mt19937_64 rng(1);  // NOLINT(cert-msc32-c,cert-msc51-cpp,bugprone-random-generator-seed) reproducible
+  for (int i = 0; i < 50000; i++) {
+    std::string str;
+    if (rng() % 3 == 0)
+      str += (rng() & 1) ? "-" : "+";
+    const int zeros = static_cast<int>(rng() % 4);
+    const int significant = 1 + static_cast<int>(rng() % 9);
+    const int trailing = static_cast<int>(rng() % 4);
+    std::string digits(zeros, '0');
+    digits += static_cast<char>('1' + rng() % 9);
+    for (int j = 1; j < significant; j++)
+      digits += static_cast<char>('0' + rng() % 10);
+    digits.append(trailing, '0');
+    const int point = static_cast<int>(rng() % (digits.size() + 2)) - 1;
+    for (int j = 0; j < static_cast<int>(digits.size()); j++) {
+      if (j == point)
+        str += '.';
+      str += digits[j];
+    }
+    if (rng() & 1) {
+      str += (rng() & 1) ? 'e' : 'E';
+      char exp[8];
+      snprintf(exp, sizeof(exp), "%d", static_cast<int>(rng() % 110) - 60);
+      str += exp;
+    }
+    expect_str_to_float_matches(str.c_str());
+  }
+}
+
+// Every 16411th float, printed with 6 and 9 digits
+TEST(StrToFloat, FloatsMatchStrtof) {
+  char buf[64];
+  for (uint32_t u = 0; u < 0x7F800000; u += 16411) {
+    for (int precision : {6, 9}) {
+      snprintf(buf, sizeof(buf), "%.*e", precision - 1, static_cast<double>(float_from(u)));
+      expect_str_to_float_matches(buf);
+    }
+  }
+}
+
+TEST(StrToFloat, SyntaxMatchesStrtof) {
+  // Tokens whose combinations the libcs read alike: no "x", "nan(" or "infi"
+  static const char *const TOKENS[] = {"0", "1", "9", ".", "e", "E", "+", "-", "a", "N", " ", "\t", "y", "00", "#", ""};
+  for (const char *a : TOKENS) {
+    for (const char *b : TOKENS) {
+      for (const char *c : TOKENS) {
+        for (const char *d : TOKENS) {
+          const std::string str = std::string(a) + b + c + d;
+          expect_str_to_float_matches(str.c_str());
+        }
+      }
+    }
+  }
+}
+
+// Where newlib and glibc read the text differently from picolibc
+TEST(StrToFloat, SyntaxMatchesPicolibc) {
+  static const PicolibcCase CASES[] = {
+      // "x" after any zero digits starts hex, even after a point or a second "0x"
+      {"000x", 0x00000000, 4},
+      {"0x", 0x00000000, 2},
+      {"-0x", 0x80000000, 3},
+      {"00x1", 0x3F800000, 4},
+      {"0x0x1", 0x3F800000, 5},
+      {"0.0x1", 0x3B800000, 5},
+      {".x1", 0x00000000, 0},
+      {"0x.p1", 0x00000000, 5},
+      {"0x.8", 0x3F000000, 4},
+      {"0X1.8", 0x3FC00000, 5},
+      {"0x1P-1", 0x3F000000, 6},
+      {"0x1p+", 0x3F800000, 3},
+      {"1x5", 0x3F800000, 1},
+      // "inf" stops before anything but "i"; past "infi" only "infinity" is a number
+      {"inf", 0x7F800000, 3},
+      {"-inf", 0xFF800000, 4},
+      {"INFINITY", 0x7F800000, 8},
+      {"+infi0", 0x00000000, 0},
+      {"infi", 0x00000000, 0},
+      {"infinit", 0x00000000, 0},
+      // "nan(" takes letters, digits and '_' up to ")", and is no number without the ")"
+      {"NaN", 0x7FC00000, 3},
+      {"nan()", 0x7FC00000, 5},
+      {"nan(xyz)", 0x7FC00000, 8},
+      {"nan(_a1)", 0x7FC00000, 8},
+      {"-nan(1)", 0xFFC00000, 7},
+      {"nan( )", 0x00000000, 0},
+      {"nan(12 ab)", 0x00000000, 0},
+      {"nan(12", 0x00000000, 0},
+      // An exponent needs a digit
+      {"1e", 0x3F800000, 1},
+      {"1e+", 0x3F800000, 1},
+      {"1e-x", 0x3F800000, 1},
+      {"\t 12", 0x41400000, 4},
+      {"- 1", 0x00000000, 0},
+      {"1.2.3", 0x3F99999A, 3},
+  };
+  for (const PicolibcCase &c : CASES)
+    expect_picolibc(c);
+}
+
+// The inputs where newlib's strtof (decimal to double to float) and picolibc's were found to differ; every platform
+// now gives picolibc's result, which is also the correctly rounded float for these
+TEST(StrToFloat, RoundsLikePicolibcWhereNewlibDoubleRounds) {
+  expect_picolibc({"1.0000000596046448", 0x3F800001, 18});
+  expect_picolibc({"0x1.100b0d0000000c", 0x3F880587, 18});
+  expect_picolibc("1.000000059604644775390625" + std::string(800, '0') + "1", 0x3F800001);
+}
+
+// picolibc rounds a decimal to 9 significant digits, ties to even, before converting it: past 9 digits its float can
+// be one away from the nearest, and an exact tie between two floats can round up. Values from picolibc.
+TEST(StrToFloat, NineDigitRoundingLikePicolibc) {
+  static const PicolibcCase CASES[] = {
+      {"1.0856183295439e10", 0x5021C50E, 18},          // nearest 0x5021C50D
+      {"3.97496754550e-4", 0x39D0671D, 16},            // nearest 0x39D0671C
+      {"9.76178673657108930e23", 0x674EB6C1, 22},      // nearest 0x674EB6C0
+      {"156.6562236e-26", 0x17F269ED, 15},             // nearest 0x17F269EC
+      {"1.000000059604644775390625", 0x3F800001, 26},  // the tie 1 + 2^-24: nearest-even 0x3F800000
+      {"1.000000059", 0x3F800001, 11},                 // 1.00000006 after rounding
+      {"7.0064923216240854e-46", 0x00000000, 22},      // 7.00649232e-46, below half the smallest float
+      {"21.299999999999997", 0x41AA6666, 18},
+      {"0.30000000000000004", 0x3E99999A, 19},
+      {"0.1000000000000000055511151231257827", 0x3DCCCCCD, 36},
+  };
+  for (const PicolibcCase &c : CASES)
+    expect_picolibc(c);
+}
+
+// Nine digits converted exactly: ties to even between floats, where the 64-bit product cannot decide alone
+TEST(StrToFloat, NineDigitTiesGoToEven) {
+  static const PicolibcCase CASES[] = {
+      {"16777217", 0x4B800000, 8},  {"16777219", 0x4B800002, 8},    {"16777217.0", 0x4B800000, 10},
+      {"8388608.5", 0x4B000000, 9}, {"8388609.5", 0x4B000002, 9},   {"3e10", 0x50DF8476, 4},
+      {"5e10", 0x513A43B7, 4},      {"1.00000006", 0x3F800001, 10},
+  };
+  for (const PicolibcCase &c : CASES)
+    expect_picolibc(c);
+}
+
+// RP2350 flushes float denormals in its double to float conversion; #22677 bench. The parser builds the bits
+// with integer operations, so every denormal parses to its own value.
+TEST(StrToFloat, DenormalsParseWithoutFlushing) {
+  char buf[64];
+  const auto check = [&buf](uint32_t k) {
+    snprintf(buf, sizeof(buf), "%.8e", std::ldexp(static_cast<double>(k), -149));
+    EXPECT_EQ(bits(str_to_float(buf, nullptr)), k) << buf;
+  };
+  for (uint32_t k = 0; k <= 0x800000; k += 4099)
+    check(k);
+  // The largest denormal and FLT_MIN, where rounding carries into the exponent
+  for (uint32_t k : {1u, 2u, 0x7FFFFEu, 0x7FFFFFu, 0x800000u})
+    check(k);
+}
+
+// Pico W and Pico 2 W bench, #22677: the RP2040 and RP2350 libcs flush float denormals to 0 in their double to
+// float conversion, and the RP2040 one turns a NaN into +infinity
+TEST(StrToFloat, Rp2040LibcFlushesAndNanToInfinity) {
+  static const PicolibcCase CASES[] = {
+      {"1e-40", 0x000116C2, 5},          {"1.4e-45", 0x00000001, 7},  {"1e-38", 0x006CE3EE, 5},
+      {"1.1754942e-38", 0x007FFFFF, 13}, {"0x1p-149", 0x00000001, 8}, {"1.17549435e-38", 0x00800000, 14},
+      {"nan(12)", 0x7FC00000, 7},        {"nan", 0x7FC00000, 3},
+  };
+  for (const PicolibcCase &c : CASES)
+    expect_picolibc(c);
+}
+
+// RP2350 bench: libc returned 0 and the first parser stalled stepping up from a flushed guess; #22677 bench
+TEST(StrToFloat, DenormalsAndFltMinMatchStrtof) {
+  for (const char *str : {"1.4e-45", "-1.4e-45", "1e-40", "1e-38", "1.17549435e-38", "1.1754942e-38", "1.17549421e-38",
+                          "1.1754945e-38", "7.0064923e-46", "7.0064924e-46"}) {
+    expect_str_to_float_matches(str);
+  }
+}
+
+// Home Assistant's non-numeric states; CodSpeed HomeassistantSensor_StateResponse_Unavailable on #22677
+TEST(StrToFloat, NonNumericStatesAreRejected) {
+  for (const char *str : {"unavailable", "unknown", "on", "off", "", "-", ".", "none", "inactive"}) {
+    expect_str_to_float_matches(str);
+    EXPECT_FALSE(parse_number<float>(str).has_value()) << str;
+  }
+}
+
+// The one float operation path takes up to 2^24 - 1 and 10^10; #22677 review
+TEST(StrToFloat, PlainPathBoundaries) {
+  for (const char *top : {"16777215", "16777216", "16777217"}) {
+    for (const char *exp : {"e10", "e-10", "e11", "e-11", "e9", "e-9", ".5", ".25"}) {
+      const std::string str = std::string(top) + exp;
+      expect_str_to_float_matches(str.c_str());
+    }
+  }
+}
+
+// The midpoint between FLT_MAX and 2^128 rounds to infinity; review
+TEST(StrToFloat, OverflowMidpoint) {
+  expect_picolibc({"340282356779733661637539395458142568448", 0x7F800000, 39});
+  expect_picolibc({"3.40282356e38", 0x7F7FFFFF, 13});
+  expect_picolibc({"3.40282357e38", 0x7F800000, 13});
+  EXPECT_EQ(bits(str_to_float("0x1.ffffffp127", nullptr)), 0x7F800000u);
+}
+
+// picolibc stops adding exponent digits at 224, so the digits of the mantissa can move it no further
+TEST(StrToFloat, HugeExponents) {
+  for (const char *str : {"1e-99999", "1e99999", "-1e99999", "0e99999", "1e9999999999999999999",
+                          "1e-9999999999999999999", "1e-999999999999", "0.000e99999", "1e224", "1e2240", "1e-2240"}) {
+    expect_str_to_float_matches(str);
+  }
+  // 10^-3000 * 10^3000 is 1, but the exponent counts only as 300
+  expect_picolibc("0." + std::string(2999, '0') + "1e3000", 0x00000000);
+  expect_picolibc("1" + std::string(300, '0') + "e-300", 0x3F800000);
+}
+
+TEST(StrToFloat, LongMantissas) {
+  expect_picolibc("1" + std::string(799, '0') + "e-799", 0x3F800000);
+  expect_picolibc("1." + std::string(3000, '0') + "1e-3001", 0x00000000);
+  expect_picolibc("0." + std::string(400, '0') + "34028235" + std::string(500, '7') + "e440", 0x7F800000);
+  expect_picolibc(std::string(300, '9') + "e-261", 0x7F800000);
+  expect_picolibc("3" + std::string(799, '7') + "e-761", 0x7F800000);
+}
+
+// picolibc takes ':' to '?' as hex digits 10 to 15 ("0x1:" is 26); these stop at them, as newlib does. #22677 review
+TEST(StrToFloat, HexStopsAtPunctuationAfterNine) {
+  struct Case {
+    const char *str;
+    uint32_t bits;
+    int32_t end;
+  };
+  const Case cases[] = {
+      {"0x1:", 0x3F800000, 3}, {"0x1;", 0x3F800000, 3}, {"0x1<", 0x3F800000, 3},  {"0x1=", 0x3F800000, 3},
+      {"0x1>", 0x3F800000, 3}, {"0x1?", 0x3F800000, 3}, {"0x1.:", 0x3F800000, 4}, {"0x:", 0x00000000, 2},
+      {"0x9?", 0x41100000, 3}, {"0xA:", 0x41200000, 3},
+  };
+  for (const auto &c : cases) {
+    char *end = nullptr;
+    const float value = str_to_float(c.str, &end);
+    EXPECT_EQ(bits(value), c.bits) << '"' << c.str << '"';
+    EXPECT_EQ(end - c.str, c.end) << '"' << c.str << '"';
+    EXPECT_FALSE(parse_number<float>(c.str).has_value()) << c.str;
+  }
+}
+
+TEST(StrToFloat, HexDenormalsAndOverflow) {
+  struct Case {
+    const char *str;
+    uint32_t bits;
+  };
+  static const Case CASES[] = {
+      {"0x1p-149", 0x00000001},         {"0x1p-150", 0x00000000},
+      {"0x1.8p-149", 0x00000002},       {"0x1.fffffcp-127", 0x007FFFFF},
+      {"0x1.fffffep-127", 0x00800000},  {"-0x1p-149", 0x80000001},
+      {"0x1p128", 0x7F800000},          {"0x1.fffffep127", 0x7F7FFFFF},
+      {"0x1.fffffefp127", 0x7F7FFFFF},  {"0x1.0000000000001p-150", 0x00000001},
+      {"0x1.000001p0", 0x3F800000},     {"0x1.000003p0", 0x3F800002},
+      {"0x1p-99999999999", 0x00000000}, {"0x1p99999999999", 0x7F800000},
+  };
+  for (const Case &c : CASES)
+    EXPECT_EQ(bits(str_to_float(c.str, nullptr)), c.bits) << c.str;
+}
+
 }  // namespace esphome::core::testing
+#endif  // USE_HOST