casaroli commented on code in PR #21491:
URL: https://github.com/apache/nuttx/pull/21491#discussion_r4587409365


##########
include/nuttx/lib/math32.h:
##########
@@ -195,176 +206,205 @@ extern "C"
 
 /* Default C implementation for umul64_const()
  *
- * Prototype: uint64_t umul64_const(uint64_t retval, uint64_t m,
- *                                  uint64_t n, bool bias);
- * Semantic:  retval = ((bias ? m : 0) + m * n) >> 64
+ * Prototype: uint64_t umul64_const(uint64_t m, uint64_t n, int bias);
+ * Semantic:  return ((bias ? m : 0) + m * n) >> 64
  *
  * The product is a 128-bit value, scaled down to 64 bits.
  * Assuming constant propagation to optimize away unused conditional code.
  * Architectures may provide their own optimized assembly implementation.
  */
 
-#  ifdef up_umul64_const
-#    define umul64_const(res, m, n, bias) \
-      (res) = up_umul64_const(m, n, bias)
-#  else
-#    define umul64_const(res, m, n, bias) \
-      do \
-        { \
-          uint32_t __m_lo = (m) & 0xffffffff; \
-          uint32_t __m_hi = (m) >> 32; \
-          uint32_t __n_lo = (n) & 0xffffffff; \
-          uint32_t __n_hi = (n) >> 32; \
-          uint32_t __res_lo; \
-          uint32_t __res_hi; \
-          uint32_t __tmp; \
-          \
-          if (!(bias)) \
-            { \
-              (res) = ((uint64_t)__m_lo * __n_lo) >> 32; \
-            } \
-          else if (!((m) & ((1ULL << 63) | (1ULL << 31)))) \
-            { \
-              (res) = ((m) + (uint64_t)__m_lo * __n_lo) >> 32; \
-            } \
-          else \
-            { \
-              (res) = (m) + (uint64_t)__m_lo * __n_lo; \
-              __res_lo = (res) >> 32; \
-              __res_hi = (__res_lo < __m_hi); \
-              (res) = __res_lo | ((uint64_t)__res_hi << 32); \
-            } \
-          \
-          if (!((m) & ((1ULL << 63) | (1ULL << 31)))) \
-            { \
-              (res) += (uint64_t)__m_lo * __n_hi; \
-              (res) += (uint64_t)__m_hi * __n_lo; \
-              (res) >>= 32; \
-            } \
-          else \
-            { \
-              (res) += (uint64_t)__m_lo * __n_hi; \
-              __tmp = (res) >> 32; \
-              (res) += (uint64_t)__m_hi * __n_lo; \
-              __res_lo = (res) >> 32; \
-              __res_hi = (__res_lo < __tmp); \
-              (res) = __res_lo | ((uint64_t)__res_hi << 32); \
-            } \
-          \
-          (res) += (uint64_t)__m_hi * __n_hi; \
-        } \
-      while (0)
-#  endif
+/* Due to compilation issues, use inline functions instead of macros. */
 
-#  define div64_const32(n, b) \
-    do \
-      { \
-        uint64_t ___res; \
-        uint64_t ___x; \
-        uint64_t ___t; \
-        uint64_t ___m; \
-        uint64_t ___n = (n); \
-        uint32_t ___p; \
-        uint32_t ___bias; \
-        uint32_t ___b = (b); \
-        ___p = 1 << LOG2_FLOOR(___b); \
-        ___m = (~0ULL / ___b) * ___p; \
-        ___m += (((~0ULL % ___b + 1) * ___p) + ___b - 1) / ___b; \
-        ___x = ~0ULL / ___b * ___b - 1; \
-        ___res = ((___m & 0xffffffff) * (___x & 0xffffffff)) >> 32; \
-        ___t = ___res += (___m & 0xffffffff) * (___x >> 32); \
-        ___res += (___x & 0xffffffff) * (___m >> 32); \
-        ___t = (___res < ___t) ? (1ULL << 32) : 0; \
-        ___res = (___res >> 32) + ___t; \
-        ___res += (___m >> 32) * (___x >> 32); \
-        ___res /= ___p; \
-        if (~0ULL % (___b / (___b & -___b)) == 0) \
-          { \
-            ___n /= (___b & -___b); \
-            ___m = ~0ULL / (___b / (___b & -___b)); \
-            ___p = 1; \
-            ___bias = 1; \
-          } \
-        else if (___res != ___x / ___b) \
-          { \
-            ___bias = 1; \
-            ___m = (~0ULL / ___b) * ___p; \
-            ___m += ((~0ULL % ___b + 1) * ___p) / ___b; \
-          } \
-        else \
-          { \
-            uint32_t ___bits = -(___m & -___m); \
-            ___bits |= ___m >> 32; \
-            ___bits = (~___bits) << 1; \
-            if (!___bits) \
-              { \
-                ___p /= (___m & -___m); \
-                ___m /= (___m & -___m); \
-              } \
-            else \
-              { \
-                ___p >>= LOG2_FLOOR(___bits); \
-                ___m >>= LOG2_FLOOR(___bits); \
-              } \
-            ___bias = 0; \
-          } \
-        umul64_const(___res, ___m, ___n, ___bias); \
-        \
-        ___res /= ___p; \
-        (n) = ___res; \
-      } \
-    while (0)
-
-#if defined(CONFIG_HAVE_EXPRESSION_STATEMENT)
-#  define div64_const(n, base) \
-    ({ \
-      uint64_t __n = (n); \
-      uint32_t __base = (base); \
-      if (IS_POWER_OF_2(__base)) \
-        { \
-          (__n) >>= LOG2_FLOOR(__base); \
-        } \
-      else if (UINTPTR_MAX == UINT32_MAX) \
-        { \
-          div64_const32(__n, __base); \
-        } \
-      else \
-        { \
-          __n /= __base; \
-        } \
-        __n; \
-      })
-
-#  define div_const(n, base) \
-    ((sizeof(typeof(n)) == sizeof(uint64_t)) ? div64_const(n, base) : ((n) / 
(base)))
-#  define div_const_roundup(n, base) \
-    ((sizeof(typeof(n)) == sizeof(uint64_t)) ? div64_const((n) + (base) - 1, 
base) : \
-     (((n) + (base) - 1) / (base)))
-#  define div_const_roundnearest(n, base) \
-    ((sizeof(typeof(n)) == sizeof(uint64_t)) ? div64_const((n) + ((base) / 2), 
base) : \
-     (((n) + ((base) / 2)) / (base)))
+#ifdef up_umul64_const
+#  define umul64_const(m, n, bias) up_umul64_const(m, n, bias)
 #else
-#  define div_const(n, base) ((n) / (base))
-#  define div_const_roundup(n, base) (((n) + (base) - 1) / (base))
-#  define div_const_roundnearest(n, base) (((n) + ((base) / 2)) / (base))
+static inline_function uint64_t umul64_const(uint64_t m, uint64_t n,
+                                             int bias)
+{
+  uint32_t m_lo = (m) & 0xffffffff;
+  uint32_t m_hi = (m) >> 32;
+  uint32_t n_lo = (n) & 0xffffffff;
+  uint32_t n_hi = (n) >> 32;
+  uint32_t res_lo;
+  uint32_t res_hi;
+  uint32_t tmp;
+  uint64_t res;
+
+  if (!(bias))
+    {
+      res = ((uint64_t)m_lo * n_lo) >> 32;
+    }
+  else if (!((m) & ((1ULL << 63) | (1ULL << 31))))
+    {
+      res = ((m) + (uint64_t)m_lo * n_lo) >> 32;
+    }
+  else
+    {
+      res = (m) + (uint64_t)m_lo * n_lo;
+      res_lo = res >> 32;
+      res_hi = (res_lo < m_hi);
+      res = res_lo | ((uint64_t)res_hi << 32);
+    }
+
+  if (!((m) & ((1ULL << 63) | (1ULL << 31))))
+    {
+      res += (uint64_t)m_lo * n_hi;
+      res += (uint64_t)m_hi * n_lo;
+      res >>= 32;
+    }
+  else
+    {
+      res += (uint64_t)m_lo * n_hi;
+      tmp = res >> 32;
+      res += (uint64_t)m_hi * n_lo;
+      res_lo = res >> 32;
+      res_hi = (res_lo < tmp);
+      res = res_lo | ((uint64_t)res_hi << 32);
+    }
+
+  res += (uint64_t)m_hi * n_hi;
+
+  return res;
+}
 #endif
 
+static inline_function uint64_t div64_const32(uint64_t n, uint32_t b)
+{
+  uint64_t res;
+  uint64_t x;
+  uint64_t t;
+  uint64_t m;
+  uint32_t p;
+  uint32_t bias;
+
+  DEBUGASSERT(b > 1);
+
+  p = 1u << LOG2_FLOOR(b);
+  m = (~0ULL / b) * p;
+  m += (((~0ULL % b + 1u) * p) + b - 1u) / b;
+  x = ~0ULL / b * b - 1u;
+  res = ((m & 0xffffffffu) * (x & 0xffffffffu)) >> 32;
+  t = res += (m & 0xffffffffu) * (x >> 32);
+  res += (x & 0xffffffffu) * (m >> 32);
+  t = (res < t) ? (1ULL << 32) : 0u;
+  res = (res >> 32) + t;
+  res += (m >> 32) * (x >> 32);
+  res /= p;
+
+  if (~0ULL % (b / (b & -b)) == 0u)
+    {
+      n /= (b & -b);
+      m = ~0ULL / (b / (b & -b));
+      p = 1u;
+      bias = 1u;
+    }
+  else if (res != x / b)
+    {
+      bias = 1u;
+      m = (~0ULL / b) * p;
+      m += ((~0ULL % b + 1u) * p) / b;
+    }
+  else
+    {
+      uint32_t bits = -(m & -m);
+      bits |= m >> 32;
+      bits = (~bits) << 1;
+      if (!bits)
+        {
+          p /= (m & -m);
+          m /= (m & -m);
+        }
+      else
+        {
+          p >>= LOG2_FLOOR(bits);
+          m >>= LOG2_FLOOR(bits);
+        }
+
+      bias = 0u;
+    }
+
+  res = umul64_const(m, n, bias);
+  res /= p;
+
+  return res;
+}
+
+#ifdef CONFIG_DEBUG_NOOPT
+
+/* Without optimization the magic number setup is not folded at compile
+ * time but runs on every call (it even contains several divisions),
+ * which is far slower than a plain division.  Cast the result, not the
+ * operands: div_const() must return uint64_t in every build, but the
+ * division itself keeps the operands' native width.
+ */
+
+#  define div64_const(n, base) ((uint64_t)((n) / (base)))
+#else
+
+/* The divisor is expected to be a compile-time constant so that the
+ * magic multiplier setup below is folded away; for runtime-invariant
+ * divisors use the invdiv API instead.
+ */
+
+static inline_function uint64_t div64_const(uint64_t n, uint64_t base)
+{
+  DEBUGASSERT(base > 0);
+
+  if (IS_POWER_OF_2(base))
+    {
+      n >>= LOG2_FLOOR(base);
+    }
+  else if (base <= UINT32_MAX)
+    {
+      n = div64_const32(n, (uint32_t)base);
+    }
+  else
+    {
+      n /= base;
+    }
+
+  return n;
+}
+#endif
+
+/* The div_const*() family takes non-negative arguments only,
+ * negative values are outside the contract.
+ */
+
+#define div_const(n, base) \
+  ((is_constexpr(base) && !is_constexpr(n) && \
+    sizeof((n)) == sizeof(uint64_t) && UINTPTR_MAX == UINT32_MAX) ? \

Review Comment:
   This PR still makes two 32-bit configs too big for their flash. CI: 
`nucleo-g431rb:nsh` (CMake) "region `flash' overflowed by 424 bytes" in `Linux 
(arm-stm32l-u-g)`, and `lm3s6965-ek:qemu-protected` "region `kflash' overflowed 
by 40 bytes" in `Linux (arm-ti)`. Both jobs pass on other PRs today. With GCC 
13.2 (Make) on the CI base d9b891b2bea, master leaves 111 and 544 bytes free. 
This PR adds 432 bytes to nucleo-g431rb:nsh (`gmtime_r` +176, `nxsig_clockwait` 
+88, `clock_systime_timespec` +72, `sysinfo` +70) and 584 bytes to the 
lm3s6965-ek kernel (`timer_settime` +148, `timer_gettime` +144, 
`tcp_update_timer` +88, `sysinfo` +74, `clock_systime_timespec` +72). 
`__udivmoddi4` stays linked in both images.
   
   Fix: make room in these two defconfigs in this PR, or take the magic path 
only when `__OPTIMIZE_SIZE__` is not defined.



##########
libs/libc/time/lib_gmtimer.c:
##########
@@ -301,15 +307,18 @@ FAR struct tm *gmtime_r(FAR const time_t *timep, FAR 
struct tm *result)
   epoch = *timep;
   linfo("timer=%d\n", (int)epoch);
 
-  /* Convert to days, hours, minutes, and seconds since the EPOCH */
+  /* Convert to days, hours, minutes, and seconds since the EPOCH.
+   * div_const() keeps the time math out of the division library; it
+   * needs a non-negative epoch.
+   */
 
-  jdn    = epoch / SEC_PER_DAY;
+  jdn    = div_const(epoch, SEC_PER_DAY);

Review Comment:
   On a 32-bit target, `gmtime_r()` with a negative `time_t` does not return. 
`div_const()` takes -1 as 2^64 - 1, so `jdn` is 213503982334601 and the year 
loop in `clock_utc2calendar()` runs about 5.8e11 times. Host test of this file 
with gcc 16 `-m32` and a 64-bit `time_t`: master returns `1970-01-01 00:00:-1` 
for t = -1, this PR does not return in 5 s (also for -60, -3600, -86400 and 
-31536000). A negative `time_t` is a valid argument of `gmtime()` (a date 
before 1970, for example a file time or a user value), and master returns.
   
   Fix: in `gmtime_r()`, use `div_const()` only when `epoch >= 0`, and keep the 
plain `/` for a negative epoch.



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to