Optimize chunkcopy_rvv:

1. Skip aligning memcpy when dist >= len.
Obviously aligning memcpy is redundant when dist >= len which
contains extra very slow load&store instrutions. And I noticed
that dist is way larger than len in most cases by adding printf in
chunkcopy_rvv with apt install (very narrow situation but makes
sense). So I tend to move the comparing before aligning memcpy
since it is only needed by the overlap situation.

2. Make the largest copy while len > dist.
Chunkcopy_rvv only copies as much memory as possible once after
aligning memcpy then uses sizeof(chunk_t) to finish the rest
copying. However, we should do the largest copy as long as
len < dist.

diff --git a/arch/riscv/chunkset_rvv.c b/arch/riscv/chunkset_rvv.c
index e0915dfc9666e5a788fdc1c7307990a383b3b94e..45ff8d333e5f25d8452d88f977defcf45fb34afc 100644 (file)
--- a/arch/riscv/chunkset_rvv.c
+++ b/arch/riscv/chunkset_rvv.c
@@ -86,11 +86,6 @@ static inline void storechunk(uint8_t *out, chunk_t *chunk) {
  */
 static inline uint8_t* CHUNKCOPY(uint8_t *out, uint8_t const *from, unsigned len) {
     Assert(len > 0, "chunkcopy should never have a length 0");
-    int32_t align = ((len - 1) % sizeof(chunk_t)) + 1;
-    memcpy(out, from, sizeof(chunk_t));
-    out += align;
-    from += align;
-    len -= align;
     ptrdiff_t dist = out - from;
     if (dist < 0 || dist >= len) {
         memcpy(out, from, len);
@@ -98,18 +93,24 @@ static inline uint8_t* CHUNKCOPY(uint8_t *out, uint8_t const *from, unsigned len
         from += len;
         return out;
     }
-    if (dist >= sizeof(chunk_t)) {
-        dist = (dist / sizeof(chunk_t)) * sizeof(chunk_t);
-        memcpy(out, from, dist);
-        out += dist;
-        from += dist;
-        len -= dist;
+
+    int32_t align = ((len - 1) % sizeof(chunk_t)) + 1;
+    memcpy(out, from, sizeof(chunk_t));
+    out += align;
+    from += align;
+    len -= align;
+
+    size_t vl = (dist / sizeof(chunk_t)) * sizeof(chunk_t);
+    while (len > dist) {
+        memcpy(out, from, vl);
+        out += vl;
+        from += vl;
+        len -= vl;
     }
-    while (len > 0) {
-        memcpy(out, from, sizeof(chunk_t));
-        out += sizeof(chunk_t);
-        from += sizeof(chunk_t);
-        len -= sizeof(chunk_t);
+
+    if (len > 0) {
+        memcpy(out, from, len);
+        out += len;
     }
     return out;
 }
@@ -118,4 +119,4 @@ static inline uint8_t* CHUNKCOPY(uint8_t *out, uint8_t const *from, unsigned len
 
 #define INFLATE_FAST     inflate_fast_rvv
 
-#include "inffast_tpl.h"
+#include "inffast_tpl.h"
\ No newline at end of file