Deduplicate scroll and SIMD paths while preserving performance

1 files changed, 40 insertions, 62 deletions

diff --git a/src/text_run.c b/src/text_run.c
index bf0e8a9..028b160 100644
--- a/src/text_run.c
+++ b/src/text_run.c
@@ -53,17 +53,11 @@ 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 l = 0; l < 16; l++) {
-        uint32_t c = cp[l];
-        bool hit = ((c - lo0 <= span0) |
-                    (c - lo1 <= span1) |
-                    (c - lo2 <= span2) |
-                    (c - lo3 <= span3));
+        bool hit = false;
+        for (int k = 0; k < TERM_WIDTH_SLOTS; k++)
+            hit |= cp[l] - wc->lo[k] <= wc->span[k];
         covered |= (uint16_t)(hit << l);
     }
     return covered;
@@ -146,24 +140,20 @@ TERM_TARGET_AVX512
 static inline struct width_ranges width_ranges_load(const struct width_cache *wc)
 {
     struct width_ranges r;
-    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]);
+    for (int k = 0; k < TERM_WIDTH_SLOTS; k++) {
+        r.lo[k] = _mm512_set1_epi32((int)wc->lo[k]);
+        r.span[k] = _mm512_set1_epi32((int)wc->span[k]);
+    }
     return r;
 }
 
 TERM_TARGET_AVX512
 static inline __mmask16 width_ranges_cover(const struct width_ranges *r, __m512i cps)
 {
-    return _mm512_cmple_epu32_mask(_mm512_sub_epi32(cps, r->lo[0]), r->span[0])
-         | _mm512_cmple_epu32_mask(_mm512_sub_epi32(cps, r->lo[1]), r->span[1])
-         | _mm512_cmple_epu32_mask(_mm512_sub_epi32(cps, r->lo[2]), r->span[2])
-         | _mm512_cmple_epu32_mask(_mm512_sub_epi32(cps, r->lo[3]), r->span[3]);
+    __mmask16 covered = 0;
+    for (int k = 0; k < TERM_WIDTH_SLOTS; k++)
+        covered |= _mm512_cmple_epu32_mask(_mm512_sub_epi32(cps, r->lo[k]), r->span[k]);
+    return covered;
 }
 
 TERM_TARGET_AVX512
@@ -189,6 +179,31 @@ static inline void store_narrow_cells(uint64_t * restrict dst, __m512i base, __m
 #undef TERM_NARROW_QUAD
 }
 
+/* Both cell and history writers validate and classify the same UTF-8 block. */
+TERM_TARGET_AVX512
+static inline bool classify_3byte_block(__m512i in, struct width_ranges *ranges,
+                                       struct width_cache *wc, __m512i *cps)
+{
+    if (TERM_EXPECT(_mm512_test_epi8_mask(
+            _mm512_xor_si512(in, _mm512_loadu_si512(utf8_3byte_want)),
+            _mm512_loadu_si512(utf8_3byte_bits)) & UTF8_3BYTE_SPAN, 0))
+        return false;
+    *cps = utf8_3byte_decode(_mm512_permutexvar_epi8(_mm512_loadu_si512(utf8_3byte_shuffle), in));
+    if (_mm512_cmplt_epu32_mask(*cps, _mm512_set1_epi32((int)UTF8_3BYTE_MIN)) ||
+        _mm512_cmple_epu32_mask(
+            _mm512_sub_epi32(*cps, _mm512_set1_epi32((int)UTF8_SURROGATE_LO)),
+            _mm512_set1_epi32((int)(UTF8_SURROGATE_HI - UTF8_SURROGATE_LO))))
+        return false;
+    __mmask16 covered = width_ranges_cover(ranges, *cps);
+    if (TERM_EXPECT(covered != 0xFFFF || wc->klass == 0, 0)) {
+        uint32_t cp[16];
+        _mm512_storeu_si512(cp, *cps);
+        if (!width_cache_extend(wc, cp, covered)) return false;
+        *ranges = width_ranges_load(wc);
+    }
+    return true;
+}
+
 TERM_TARGET_AVX512
 static TERM_NOINLINE int
 decode_3byte_blocks(uint64_t * restrict dst, const char * restrict buf, int len, int cols_left,
@@ -197,33 +212,14 @@ decode_3byte_blocks(uint64_t * restrict dst, const char * restrict buf, int len,
                    int * restrict out_cols, uint32_t * restrict last_cp)
 {
     struct width_ranges ranges = width_ranges_load(wc);
-    __m512i shuffle = _mm512_loadu_si512(utf8_3byte_shuffle);
-    __m512i want = _mm512_loadu_si512(utf8_3byte_want);
-    __m512i bits = _mm512_loadu_si512(utf8_3byte_bits);
     uint32_t last = *last_cp;
     int i = 0, cx = 0;
 
     while (i + 48 <= len && cx + 32 <= cols_left &&
            ((unsigned char)buf[i + 45] & UTF8_LEAD3_MASK) == UTF8_LEAD3_TAG) {
         __m512i in = _mm512_maskz_loadu_epi8(UTF8_3BYTE_SPAN, buf + i);
-        if (TERM_EXPECT(_mm512_test_epi8_mask(_mm512_xor_si512(in, want), bits) &
-                             UTF8_3BYTE_SPAN, 0))
-            break;
-
-        __m512i cps = utf8_3byte_decode(_mm512_permutexvar_epi8(shuffle, in));
-        if (_mm512_cmplt_epu32_mask(cps, _mm512_set1_epi32((int)UTF8_3BYTE_MIN)) ||
-            _mm512_cmple_epu32_mask(
-                _mm512_sub_epi32(cps, _mm512_set1_epi32((int)UTF8_SURROGATE_LO)),
-                _mm512_set1_epi32((int)(UTF8_SURROGATE_HI - UTF8_SURROGATE_LO))))
-            break;
-        __mmask16 covered = width_ranges_cover(&ranges, cps);
-        if (TERM_EXPECT(covered != 0xFFFF || wc->klass == 0, 0)) {
-            uint32_t cp[16];
-            _mm512_storeu_si512(cp, cps);
-            if (!width_cache_extend(wc, cp, covered))
-                break;
-            ranges = width_ranges_load(wc);
-        }
+        __m512i cps;
+        if (!classify_3byte_block(in, &ranges, wc, &cps)) break;
 
         if (wc->klass == 2) {
             store_wide_cells(dst, base_wide_pair, cps);
@@ -258,9 +254,6 @@ text_3byte_blocks(uint8_t * restrict dst, const char * restrict buf, int len, in
          0, 36, 37, 38,   0, 39, 40, 41,   0, 42, 43, 44,   0, 45, 46, 47,
     };
     struct width_ranges ranges = width_ranges_load(wc);
-    __m512i shuffle = _mm512_loadu_si512(utf8_3byte_shuffle);
-    __m512i want = _mm512_loadu_si512(utf8_3byte_want);
-    __m512i bits = _mm512_loadu_si512(utf8_3byte_bits);
     __m512i idx = _mm512_loadu_si512(spread);
     __m512i prefix = _mm512_set1_epi8((char)0xF5);
     uint32_t last = *last_cp;
@@ -269,23 +262,8 @@ text_3byte_blocks(uint8_t * restrict dst, const char * restrict buf, int len, in
     while (i + 48 <= len && cx + 32 <= cols_left &&
            ((unsigned char)buf[i + 45] & UTF8_LEAD3_MASK) == UTF8_LEAD3_TAG) {
         __m512i in = _mm512_maskz_loadu_epi8(UTF8_3BYTE_SPAN, buf + i);
-        if (TERM_EXPECT(_mm512_test_epi8_mask(_mm512_xor_si512(in, want), bits) &
-                             UTF8_3BYTE_SPAN, 0))
-            break;
-        __m512i cps = utf8_3byte_decode(_mm512_permutexvar_epi8(shuffle, in));
-        if (_mm512_cmplt_epu32_mask(cps, _mm512_set1_epi32((int)UTF8_3BYTE_MIN)) ||
-            _mm512_cmple_epu32_mask(
-                _mm512_sub_epi32(cps, _mm512_set1_epi32((int)UTF8_SURROGATE_LO)),
-                _mm512_set1_epi32((int)(UTF8_SURROGATE_HI - UTF8_SURROGATE_LO))))
-            break;
-        __mmask16 covered = width_ranges_cover(&ranges, cps);
-        if (TERM_EXPECT(covered != 0xFFFF || wc->klass == 0, 0)) {
-            uint32_t cp[16];
-            _mm512_storeu_si512(cp, cps);
-            if (!width_cache_extend(wc, cp, covered))
-                break;
-            ranges = width_ranges_load(wc);
-        }
+        __m512i cps;
+        if (!classify_3byte_block(in, &ranges, wc, &cps)) break;
         if (wc->klass == 2) {
             __m512i out = _mm512_mask_mov_epi8(_mm512_permutexvar_epi8(idx, in),
                                                0x1111111111111111ull, prefix);