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

1 files changed, 30 insertions, 22 deletions

diff --git a/src/text_run.c b/src/text_run.c
index 3901cbe..9df7ac0 100644
--- a/src/text_run.c
+++ b/src/text_run.c
@@ -53,11 +53,19 @@ static TERM_MAYBE_UNUSED int width_run_find(uint32_t cp)
 static TERM_MAYBE_UNUSED uint16_t width_cache_covers(const struct width_cache *wc,
                                                      const uint32_t *cp)
 {
+    uint32_t lo0 = wc->lo[0], span0 = wc->span[0];
+    uint32_t lo1 = wc->lo[1], span1 = wc->span[1];
+    uint32_t lo2 = wc->lo[2], span2 = wc->span[2];
+    uint32_t lo3 = wc->lo[3], span3 = wc->span[3];
     uint16_t covered = 0;
-    for (int j = 0; j < TERM_WIDTH_SLOTS; j++)
-        for (int l = 0; l < 16; l++)
-            if (cp[l] - wc->lo[j] <= wc->span[j])
-                covered |= (uint16_t)(1u << l);
+    for (int l = 0; l < 16; l++) {
+        uint32_t c = cp[l];
+        bool hit = ((c - lo0 <= span0) |
+                    (c - lo1 <= span1) |
+                    (c - lo2 <= span2) |
+                    (c - lo3 <= span3));
+        covered |= (uint16_t)(hit << l);
+    }
     return covered;
 }
 
@@ -70,10 +78,8 @@ static TERM_MAYBE_UNUSED bool width_cache_extend(struct width_cache *wc, const u
     uint8_t k = width_run_class[r];
     if (k != 1 && k != 2) return false;
     if (k != wc->klass) {
-        for (int j = 0; j < TERM_WIDTH_SLOTS; j++) {
-            wc->lo[j] = 0xFFFFFFFFu;
-            wc->span[j] = 0;
-        }
+        memset(wc->lo, 0xFF, sizeof(wc->lo));
+        memset(wc->span, 0, sizeof(wc->span));
         wc->klass = k;
         wc->next = 0;
     }
@@ -140,10 +146,14 @@ TERM_TARGET_AVX512
 static inline struct width_ranges width_ranges_load(const struct width_cache *wc)
 {
     struct width_ranges r;
-    for (int j = 0; j < TERM_WIDTH_SLOTS; j++) {
-        r.lo[j] = _mm512_set1_epi32((int)wc->lo[j]);
-        r.span[j] = _mm512_set1_epi32((int)wc->span[j]);
-    }
+    r.lo[0] = _mm512_set1_epi32((int)wc->lo[0]);
+    r.span[0] = _mm512_set1_epi32((int)wc->span[0]);
+    r.lo[1] = _mm512_set1_epi32((int)wc->lo[1]);
+    r.span[1] = _mm512_set1_epi32((int)wc->span[1]);
+    r.lo[2] = _mm512_set1_epi32((int)wc->lo[2]);
+    r.span[2] = _mm512_set1_epi32((int)wc->span[2]);
+    r.lo[3] = _mm512_set1_epi32((int)wc->lo[3]);
+    r.span[3] = _mm512_set1_epi32((int)wc->span[3]);
     return r;
 }
 
@@ -519,16 +529,14 @@ static inline int decode_3byte_quads(uint64_t **dst_io, const char *buf, int len
         uint8_t k2 = unicode_fast_class[cp[2]], k3 = unicode_fast_class[cp[3]];
         if (!k0 || !k1 || !k2 || !k3)
             break;
-        const uint8_t k[4] = { k0, k1, k2, k3 };
-        for (int j = 0; j < 4; j++) {
-            if (k[j] == 1) {
-                dst = store_cell(dst, b, cp[j]);
-                cx++;
-            } else {
-                dst = store_cell_wide(dst, b, cp[j]);
-                cx += 2;
-            }
-        }
+        if (k0 == 1) { dst = store_cell(dst, b, cp[0]); cx++; }
+        else { dst = store_cell_wide(dst, b, cp[0]); cx += 2; }
+        if (k1 == 1) { dst = store_cell(dst, b, cp[1]); cx++; }
+        else { dst = store_cell_wide(dst, b, cp[1]); cx += 2; }
+        if (k2 == 1) { dst = store_cell(dst, b, cp[2]); cx++; }
+        else { dst = store_cell_wide(dst, b, cp[2]); cx += 2; }
+        if (k3 == 1) { dst = store_cell(dst, b, cp[3]); cx++; }
+        else { dst = store_cell_wide(dst, b, cp[3]); cx += 2; }
         *last = cp[3];
         i += 12;
     }