optimized float to integer routines

Author: ZERICO2005

src/crt/dtol.src

@@ -3,8 +3,12 @@
 	section	.text
 
 	public	__dtol
+	public	__dtoul
 
+; __dtol_c correctly handles all non-UB cases for both
+; (long)long double and (unsigned long)long double
 __dtol:
+__dtoul:
 	; f64_ret_i32
 	push	af, iy
 	ld	a, b

src/crt/dtoll.src

@@ -3,8 +3,12 @@
 	section	.text
 
 	public	__dtoll
+	public	__dtoull
 
+; __dtoll_c correctly handles all non-UB cases for both
+; (long long)long double and (unsigned long long)long double
 __dtoll:
+__dtoull:
 	; f64_ret_i64
 	push	af, iy
 	ld	a, b

src/crt/dtoul.src

@@ -64,16 +64,6 @@ uint64_t _dtoull_c(long double x) {
 }
 #endif
 
-/**
- * @brief set to 0 or 1
- * If set to 1, values that truncate to `INT32_MIN`/`INT64_MIN` will be
- * handled correctly.
- * If set to 0, it can save a little bit of space by removing a comparison from
- * `_dtol_c` and `_dtoll_c`. However this will cause values that would truncate
- * to `INT32_MIN`/`INT64_MIN` to have an undefined result.
- */
-#define HANDLE_INT_MIN 1
-
 typedef struct f64_sign {
     long double flt;
     bool sign;
@@ -106,41 +96,18 @@ static uint64_t f64_to_unsigned(F64_pun val) {
     return val.bin;
 }
 
-uint64_t _dtoull_c(long double x) {
-    F64_pun val;
-    val.flt = x;
-    /* overflow || signbit(x) || isinf(x) || isnan(x) */
-    if (val.reg.BC >= ((Float64_bias + Float64_u64_max_exp) << Float64_exp_BC_shift)) {
-        /* undefined return value for negative/overflow/inf/NaN of x */
-        return 0;
-    }
-    return f64_to_unsigned(val);
-}
-
-uint32_t _dtoul_c(long double x) {
-    F64_pun val;
-    val.flt = x;
-    /* overflow || signbit(x) || isinf(x) || isnan(x) */
-    if (val.reg.BC >= ((Float64_bias + Float64_u32_max_exp) << Float64_exp_BC_shift)) {
-        /* undefined return value for negative/overflow/inf/NaN values of x */
-        return 0;
-    }
-    return (uint32_t)f64_to_unsigned(val);
-}
-
+/**
+ * @brief the exact same routine is used for (long long)long double and
+ * (unsigned long long)long double. If the input long double is out of range,
+ * then the conversion is UB anyways.
+ */
 int64_t _dtoll_c(f64_sign arg) {
     F64_pun val;
     bool x_sign = arg.sign;
     val.flt = arg.flt;
 
     /* overflow || isinf(x) || isnan(x) */
-    if (val.reg.BC >= ((Float64_bias + Float64_i64_max_exp) << Float64_exp_BC_shift)) {
-        #if HANDLE_INT_MIN != 0
-            /* if the value truncates to INT64_MIN */
-            if (x_sign && val.bin == UINT64_C(0x43E0000000000000)) {
-                return INT64_MIN;
-            }
-        #endif
+    if (val.reg.BC >= ((Float64_bias + Float64_u64_max_exp) << Float64_exp_BC_shift)) {
         /* undefined return value for underflow/overflow/inf/NaN values of x */
         return 0;
     }
@@ -150,19 +117,18 @@ int64_t _dtoll_c(f64_sign arg) {
     return ret;
 }
 
+/**
+ * @brief the exact same routine is used for (long)long double and
+ * (unsigned long)long double. If the input long double is out of range,
+ * then the conversion is UB anyways.
+ */
 int32_t _dtol_c(f64_sign arg) {
     F64_pun val;
     bool x_sign = arg.sign;
     val.flt = arg.flt;
 
     /* overflow || isinf(x) || isnan(x) */
-    if (val.reg.BC >= ((Float64_bias + Float64_i32_max_exp) << Float64_exp_BC_shift)) {
-        #if HANDLE_INT_MIN != 0
-            /* if the value truncates to INT32_MIN */
-            if (x_sign && val.bin <= UINT64_C(0x41E00000001FFFFF)) {
-                return INT32_MIN;
-            }
-        #endif
+    if (val.reg.BC >= ((Float64_bias + Float64_u32_max_exp) << Float64_exp_BC_shift)) {
         /* undefined return value for underflow/overflow/inf/NaN values of x */
         return 0;
     }

src/crt/dtoull.src

@@ -2,6 +2,11 @@
 #include <math.h>
 #include <stdint.h>
 
+/**
+ * @brief the exact same routine is used for (long long)float and
+ * (unsigned long long)float. If the input float is out of range,
+ * then the conversion is UB anyways.
+ */
 long long _ftoll_c(float x)
 {
     const union { float f; uint32_t u; struct { uint32_t mantissa: FLT_MANT_DIG - 1, exponent: 8, sign: 1; }; } parts = { .f = x };

src/crt/float64_to_int.c

@@ -3,7 +3,12 @@
 	section	.text
 
 	public	__ftoll
+	public	__ftoull
+
+; __ftoll_c correctly handles all non-UB cases for both
+; (long long)float and (unsigned long long)float
 __ftoll:
+__ftoull:
 	ld	d, a
 	push	iy, de, hl
 	call	__ftoll_c