���� �� ������ ecomp.r29.143691

������ ���������:
  ������ liblccopt: ecomp.r29.143691 - ecomp.r28.143691

VERSION

@@ -1 +1 @@
-ecomp.version : ecomp.r29.143635
+ecomp.version : ecomp.r29.143691

tools/lccrt_s/src/lccrt_vec.c

@@ -10,6 +10,82 @@
 #include <string.h>
 #include <assert.h>
 
+static uint64_t
+__lccrt_get_bits( uint64_t u, int off, int bits) {
+    int lshift = 64 - off - bits;
+    int rshift = 64 - bits;
+    uint64_t r = (bits > 0) ? ((u << lshift) >> rshift) : 0;
+
+    return (r);
+} /* __lccrt_get_bits */
+
+static uint64_t
+__lccrt_set_bits( uint64_t u, int off, int bits, uint64_t v) {
+    uint64_t r = u;
+
+    if ( bits > 0 ) {
+        int lshift = 64 - bits;
+        int rshift = 64 - bits - off;
+        uint64_t bitmask = ((~0ULL) << lshift) >> rshift;
+
+        r = (u & ~bitmask) | ((v << off) & bitmask);
+    }
+
+    return (r);
+} /* __lccrt_set_bits */
+
+/**
+ * ÷ ËÁÖÄÏÍ ÉÚ ÐÁÒÁÍÅÔÒÏ× u0/u1 ÈÒÁÎÉÔÓÑ ÐÏ ubytes ÂÁÊÔ ÄÁÎÎÙÈ. îÕÖÎÏ ÓËÌÅÉÔØ
+ * ÄÁÎÎÙÅ É ÐÏ ÂÉÔÏ×ÏÍÕ ÓÍÅÝÅÎÉÀ eoff ÐÒÏÞÉÔÁÔØ ÚÎÁÞÅÎÉÅ ÒÁÚÍÅÒÏÍ ebits ÂÉÔ.
+ */
+static uint64_t
+__lccrt_get_pair_bits( uint64_t u0, uint64_t u1, int ubytes, int eoff, int ebits) {
+    uint64_t r = 0;
+    int ubits = 8*ubytes;
+
+    assert( (0 <= eoff) && (0 <= ebits));
+    assert( (eoff + ebits <= 2*ubits) && (ebits <= 64));
+    if ( eoff + ebits <= ubits ) {
+        r = __lccrt_get_bits( u0, eoff, ebits);
+    } else if ( eoff >= ubits ) {
+        r = __lccrt_get_bits( u1, eoff - ubits, ebits);
+    } else {
+        int lo_len = ubits - eoff;
+        uint64_t lo = __lccrt_get_bits( u0, eoff, lo_len);
+        uint64_t hi = __lccrt_get_bits( u1, 0, ebits - lo_len);
+
+        r = (hi << lo_len) | lo;
+    }
+
+    return (r);
+} /* __lccrt_get_pair_bits */
+
+/**
+ * ÷ ËÁÖÄÏÍ ÉÚ ÐÁÒÁÍÅÔÒÏ× u0/u1 ÈÒÁÎÉÔÓÑ ÐÏ ubytes ÂÁÊÔ ÄÁÎÎÙÈ. îÕÖÎÏ
+ * ÌÏÇÉÞÅÓËÉ ÓËÌÅÉÔØ ÄÁÎÎÙÅ É ÐÏ ÂÉÔÏ×ÏÍÕ ÓÍÅÝÅÎÉÀ eoff ÚÁÐÉÓÁÔØ ÚÎÁÞÅÎÉÅ v
+ * ÒÁÚÍÅÒÏÍ ebits ÂÉÔ.
+ */
+static void
+__lccrt_set_pair_bits( uint64_t *u0, uint64_t *u1, int ubytes, int eoff, int ebits, uint64_t v) {
+    uint64_t r = 0;
+    int ubits = 8*ubytes;
+
+    assert( (0 <= eoff) && (0 <= ebits));
+    assert( (eoff + ebits <= 2*ubits) && (ebits <= 64));
+    if ( eoff + ebits <= ubits ) {
+        u0[0] = __lccrt_set_bits( u0[0], eoff, ebits, v);
+    } else if ( eoff >= ubits ) {
+        u1[0] = __lccrt_set_bits( u1[0], eoff - ubits, ebits, v);
+    } else {
+        int lo_len = ubits - eoff;
+
+        u0[0] = __lccrt_set_bits( u0[0], eoff, lo_len, v);
+        u1[0] = __lccrt_set_bits( u1[0], 0, ebits - lo_len, v >> lo_len);
+    }
+
+    return;
+} /* __lccrt_set_pair_bits */
+
 static int
 __lccrt_bitwidth_bytesize( int bitsize) {
     int r = 0;
@@ -75,50 +151,63 @@ __lccrt_load_bytes( void *src, int64_t bytesize) {
 
 /**
  * õÐÁËÏ×ËÁ ÂÉÔÏ× ×ÅËÔÏÒÁ ÎÅÓÔÁÎÄÁÒÔÎÙÈ ÃÅÌÙÈ × (ÎÅÓÔÁÎÄÁÒÔÎÏÅ) ÃÅÌÏÅ ÚÎÁÞÅÎÉÅ.
- * óÕÍÍÁÒÎÏÅ ËÏÌÉÞÅÓÔ×Ï ÚÎÁÞÉÍÙÈ ÂÉÔ ÎÅ ÄÏÌÖÎÏ ÐÒÅ×ÙÛÁÔØ 64.
  * ðÒÉÍÅÒ: <3 x i6> -> i18.
  */
 void
 __lccrt_vecbitpack( void *dst, void *src, int64_t veclen, int64_t elembitsize) {
-    uint64_t r = 0;
     int ebytes = __lccrt_bitwidth_bytesize( elembitsize);
-    int maskshift = 64 - elembitsize;
+    int lcnt = 0;
-    int totalbitsize = veclen*elembitsize;
+    int uoff = 0;
+    uint64_t u0 = 0;
+    uint64_t u1 = 0;
 
-    assert( (0 <= veclen) && (0 < elembitsize) && (totalbitsize <= 64));
+    assert( (0 <= veclen) && (0 < elembitsize) && (elembitsize <= 64));
-    for ( int i = veclen - 1; i >= 0; --i ) {
+    for ( int i = 0; i < veclen; ++i ) {
-        uint64_t si = 0;
+        uint64_t v = __lccrt_load_bytes( (char *)src + i*ebytes, ebytes);
 
-        si = __lccrt_load_bytes( (char *)src + i*ebytes, ebytes);
+        __lccrt_set_pair_bits( &u0, &u1, ebytes, uoff, elembitsize, v);
-        si = (si << maskshift) >> maskshift;
+        uoff = uoff + elembitsize;
-        r = (r << elembitsize) | si;
+        if ( uoff >= 8*ebytes ) {
+            __lccrt_store_bytes( (char *)dst + lcnt, u0, ebytes);
+            u0 = u1;
+            u1 = 0;
+            lcnt += ebytes;
+            uoff = uoff - 8*ebytes;
+        }
     }
 
-    __lccrt_store_bytes( dst, r, __lccrt_bitwidth_bytesize( totalbitsize));
+    if ( uoff > 0 ) {
+        __lccrt_store_bytes( (char *)dst + lcnt, u0, ebytes);
+    }
 
     return;
 }
 
 /**
  * òÁÓÐÁËÏ×ËÁ ÂÉÔÏ× (ÎÅÓÔÁÎÄÁÒÔÎÏÇÏ) ÃÅÌÏÇÏ × ×ÅËÔÏÒ ÎÅÓÔÁÎÄÁÒÔÎÙÈ ÃÅÌÙÈ.
- * óÕÍÍÁÒÎÏÅ ËÏÌÉÞÅÓÔ×Ï ÚÎÁÞÉÍÙÈ ÂÉÔ ÎÅ ÄÏÌÖÎÏ ÐÒÅ×ÙÛÁÔØ 64.
  * ðÒÉÍÅÒ: i18 -> <3 x i6>.
  */
 void
 __lccrt_vecbitunpack( void *dst, void *src, int64_t veclen, int64_t elembitsize) {
     int ebytes = __lccrt_bitwidth_bytesize( elembitsize);
-    int maskshift = 64 - elembitsize;
+    int lcnt = ebytes;
-    int totalbitsize = veclen*elembitsize;
+    int uoff = 8*ebytes;
-    uint64_t v0 = __lccrt_load_bytes( src, __lccrt_bitwidth_bytesize( totalbitsize));
+    uint64_t u0 = 0;
-    uint64_t v = v0;
+    uint64_t u1 = __lccrt_load_bytes( src, ebytes);
 
-    assert( (0 <= veclen) && (0 < elembitsize) && (totalbitsize <= 64));
+    assert( (0 <= veclen) && (0 < elembitsize) && (elembitsize <= 64));
     for ( int i = 0; i < veclen; ++i ) {
-        uint64_t si = 0;
+        uint64_t v = __lccrt_get_pair_bits( u0, u1, ebytes, uoff, elembitsize);
 
-        si = (v << maskshift) >> maskshift;
+        __lccrt_store_bytes( (char *)dst + i*ebytes, v, ebytes);
-        v = v >> elembitsize;
+
-        __lccrt_store_bytes( (char *)dst + i*ebytes, si, ebytes);
+        uoff = uoff + elembitsize;
+        if ( uoff + elembitsize > 2*8*ebytes ) {
+            u0 = u1;
+            u1 = __lccrt_load_bytes( (char *)src + lcnt, ebytes);
+            lcnt += ebytes;
+            uoff = uoff - 8*ebytes;
+        }
     }
 
     return;