Optimize hot loops with direct access, LUTs, and closed-form formulas

1 files changed, 52 insertions, 21 deletions

diff --git a/src/palette.c b/src/palette.c
index a8add4e..bfaba6c 100644
--- a/src/palette.c
+++ b/src/palette.c
@@ -133,13 +133,19 @@ static uint32_t *palette_slot(struct term_theme *p, const char *key,
     const char *digits = key + 5;
     if (!*digits || (*digits == '0' && digits[1]))
         return NULL;
-    int i = 0;
-    for (const char *d = digits; *d; d++) {
-        if (*d < '0' || *d > '9')
+    int i;
+    if (!digits[1]) {
+        if (digits[0] < '0' || digits[0] > '9')
             return NULL;
-        i = i * 10 + (*d - '0');
+        i = digits[0] - '0';
+    } else if (!digits[2]) {
+        if (digits[0] < '0' || digits[0] > '9' || digits[1] < '0' || digits[1] > '9')
+            return NULL;
+        i = (digits[0] - '0') * 10 + (digits[1] - '0');
         if (i > 15)
             return NULL;
+    } else {
+        return NULL;
     }
     *fallback = builtin_theme.ansi[i];
     return &p->ansi[i];
@@ -166,19 +172,36 @@ bool term_theme_set_key(struct term_theme *p, const char *key, const char *val)
 
 static bool parse_hex_channel(const char *s, size_t len, uint8_t *out)
 {
-    if (len == 0 || len > 4)
-        return false;
-    unsigned v = 0;
-    for (size_t i = 0; i < len; i++) {
-        int d = term_hex_nib(s[i]);
+    if (len == 2) {
+        int d0 = term_hex_nib(s[0]), d1 = term_hex_nib(s[1]);
+        if ((d0 | d1) < 0)
+            return false;
+        *out = (uint8_t)((d0 << 4) | d1);
+        return true;
+    }
+    if (len == 1) {
+        int d = term_hex_nib(s[0]);
         if (d < 0)
             return false;
-        v = (v << 4) | (unsigned)d;
+        *out = (uint8_t)(d * 17);
+        return true;
     }
-    static const unsigned max_for_len[5] = {0, 15, 255, 4095, 65535};
-    unsigned max = max_for_len[len];
-    *out = (uint8_t)((unsigned long)v * 255 / max);
-    return true;
+    if (len == 4) {
+        int d0 = term_hex_nib(s[0]), d1 = term_hex_nib(s[1]);
+        int d2 = term_hex_nib(s[2]), d3 = term_hex_nib(s[3]);
+        if ((d0 | d1 | d2 | d3) < 0)
+            return false;
+        *out = (uint8_t)((((unsigned)d0 << 12) | ((unsigned)d1 << 8) | ((unsigned)d2 << 4) | (unsigned)d3) * 255 / 65535);
+        return true;
+    }
+    if (len == 3) {
+        int d0 = term_hex_nib(s[0]), d1 = term_hex_nib(s[1]), d2 = term_hex_nib(s[2]);
+        if ((d0 | d1 | d2) < 0)
+            return false;
+        *out = (uint8_t)((((unsigned)d0 << 8) | ((unsigned)d1 << 4) | (unsigned)d2) * 255 / 4095);
+        return true;
+    }
+    return false;
 }
 
 static bool parse_fraction(const char *s, size_t len, uint8_t *out)
@@ -224,6 +247,8 @@ static bool parse_fraction(const char *s, size_t len, uint8_t *out)
 
 static const char *skip_leading_ws(const char *s)
 {
+    if (*s != ' ' && *s != '\t')
+        return s;
     while (*s == ' ' || *s == '\t')
         s++;
     return s;
@@ -232,6 +257,8 @@ static const char *skip_leading_ws(const char *s)
 static size_t rtrimmed_len(const char *s)
 {
     size_t n = strlen(s);
+    if (n == 0 || (s[n - 1] != ' ' && s[n - 1] != '\t'))
+        return n;
     while (n > 0 && (s[n - 1] == ' ' || s[n - 1] == '\t'))
         n--;
     return n;
@@ -294,15 +321,19 @@ bool term_color_parse(const char *input, uint32_t *out)
         }
         if (ci != 3)
             return false;
-        uint8_t c[3];
-        for (int k = 0; k < 3; k++) {
-            bool ok = intensity
-                ? parse_fraction(channels[k], chlens[k], &c[k])
-                : parse_hex_channel(channels[k], chlens[k], &c[k]);
-            if (!ok)
+        uint8_t c0, c1, c2;
+        if (intensity) {
+            if (!parse_fraction(channels[0], chlens[0], &c0) ||
+                !parse_fraction(channels[1], chlens[1], &c1) ||
+                !parse_fraction(channels[2], chlens[2], &c2))
+                return false;
+        } else {
+            if (!parse_hex_channel(channels[0], chlens[0], &c0) ||
+                !parse_hex_channel(channels[1], chlens[1], &c1) ||
+                !parse_hex_channel(channels[2], chlens[2], &c2))
                 return false;
         }
-        *out = term_rgb(c[0], c[1], c[2]);
+        *out = term_rgb(c0, c1, c2);
         return true;
     }