From 9f80d0d60e1272717fbb90289b37646de5346111 Mon Sep 17 00:00:00 2001
From: Joshua Haberman <joshua@reverberate.org>
Date: Sat, 20 Jun 2009 22:45:24 -0700
Subject: Do unaligned reads if UPB_UNALIGNED_READS_OK is defined.

Also removed commented-out alternate implementations for
varint-reading code (they are still in the history).
---
 upb_parse.c | 99 ++++++++++---------------------------------------------------
 1 file changed, 15 insertions(+), 84 deletions(-)

(limited to 'upb_parse.c')

diff --git a/upb_parse.c b/upb_parse.c
index 33ccb42..b13ca24 100644
--- a/upb_parse.c
+++ b/upb_parse.c
@@ -54,80 +54,6 @@ done:
   return UPB_STATUS_OK;
 }
 
-/* Alternate implementations of get_v_uint64_t -- one performs no branching,
- * another performs only a single branch.  TODO: test performance. */
-
-#if 0
-/* The no-branching version. */
-static upb_status_t get_v_uint64_t(void *restrict *buf,
-                                   uint64_t *restrict val)
-{
-  uint8_t *b = *buf;
-  /* Endian-specific! */
-  uint64_t b0 = *(uint64_t*)b;
-  uint16_t b10 = *(uint16_t*)(b+8);
-
-  /* Put the 10 continuation bits in the bottom 10 bits of cont. */
-  uint32_t cont = (b0 & 0x8080808080808080ULL) * 0x2040810204081ULL >> 56;
-  cont |= (*(b+8) & 0x80) << 1 | (*(b+9) & 0x80) << 2;
-  if(cont == 0x3FF) return UPB_ERROR_UNTERMINATED_VARINT;
-
-  int num_bytes = __builtin_ctzll(~cont) + 1;
-  *buf += num_bytes;
-
-  uint64_t more_than_8_mask = -(num_bytes > 8);
-  uint64_t low10_mask = (0x7F7F7F7F7F7F7F7FULL >> (8-num_bytes)*8);
-  low10_mask ^= (low10_mask ^ 0x7F7F7F7F7F7F7F7FULL) & more_than_8_mask;
-  uint64_t valb0 = b0 & low10_mask;
-  uint16_t valb10 = b10 & (0x7F7FU >> (10-num_bytes)*8) & more_than_8_mask;
-
-  *val = ((valb0 & 0xFF00000000000000) >> 7) |
-         ((valb0 & 0x00FF000000000000) >> 6) |
-         ((valb0 & 0x0000FF0000000000) >> 5) |
-         ((valb0 & 0x000000FF00000000) >> 4) |
-         ((valb0 & 0x00000000FF000000) >> 3) |
-         ((valb0 & 0x0000000000FF0000) >> 2) |
-         ((valb0 & 0x000000000000FF00) >> 1) |
-         ((valb0 & 0x00000000000000FF)) |
-         (((uint64_t)valb10 & 0xFF00) << 55) |
-         (((uint64_t)valb10 & 0xFF) << 56);
-  return UPB_STATUS_OK;
-}
-
-/* The single-branch version. */
-static upb_status_t get_v_uint64_t(void *restrict *buf,
-                                   uint64_t *restrict val)
-{
-  uint8_t *b = *buf;
-  /* Endian-specific! */
-  uint64_t b0 = *(uint64_t*)b;
-
-  /* Put the 10 continuation bits in the bottom 10 bits of cont. */
-  uint32_t cont = (b0 & 0x8080808080808080ULL) * 0x2040810204081ULL >> 56;
-  cont |= (*(b+8) & 0x80) << 1 | (*(b+9) & 0x80) << 2;
-
-  int num_bytes = __builtin_ffs(~cont);
-  uint32_t part0 = 0, part1 = 0, part2 = 0;
-
-  switch(num_bytes) {
-    default: return UPB_ERROR_UNTERMINATED_VARINT;
-    case 10: part2 |= (b[9] & 0x7F) <<  7;
-    case 9:  part2 |= (b[8] & 0x7F)      ;
-    case 8:  part1 |= (b[7] & 0x7F) << 21;
-    case 7:  part1 |= (b[6] & 0x7F) << 14;
-    case 6:  part1 |= (b[5] & 0x7F) <<  7;
-    case 5:  part1 |= (b[4] & 0x7F)      ;
-    case 4:  part0 |= (b[3] & 0x7F) << 21;
-    case 3:  part0 |= (b[2] & 0x7F) << 14;
-    case 2:  part0 |= (b[1] & 0x7F) <<  7;
-    case 1:  part0 |= (b[0] & 0x7F)      ;
-  }
-  *buf = b + num_bytes;
-  *val = (uint64_t)part0 | ((uint64_t)part1 << 28) | ((uint64_t)part2 << 56);
-  return UPB_STATUS_OK;
-}
-#endif
-
 static upb_status_t skip_v_uint64_t(void **buf)
 {
   uint8_t *ptr = *buf, b;
@@ -168,17 +94,21 @@ done:
   return UPB_STATUS_OK;
 }
 
+#define SHL(val, bits) ((uint32_t)val << bits)
 static upb_status_t get_f_uint32_t(void *restrict *buf,
                                    uint32_t *restrict val)
 {
   uint8_t *b = *buf;
-#define SHL(val, bits) ((uint32_t)val << bits)
+#if UPB_UNALIGNED_READS_OK
+  *val = *(uint32_t*)b;
+#else
   *val = SHL(b[0], 0) | SHL(b[1], 8) | SHL(b[2], 16) | SHL(b[3], 24);
-#undef SHL
+#endif
   b += sizeof(uint32_t);
   *buf = b;
   return UPB_STATUS_OK;
 }
+#undef SHL
 
 static upb_status_t skip_f_uint32_t(void **buf)
 {
@@ -189,14 +119,15 @@ static upb_status_t skip_f_uint32_t(void **buf)
 static upb_status_t get_f_uint64_t(void *restrict *buf,
                                    uint64_t *restrict val)
 {
-  uint8_t *b = *buf;
-  /* TODO: is this worth 32/64 specializing? */
-#define SHL(val, bits) ((uint64_t)val << bits)
-  *val = SHL(b[0], 0)  | SHL(b[1], 8)  | SHL(b[2], 16) | SHL(b[3], 24) |
-         SHL(b[4], 32) | SHL(b[5], 40) | SHL(b[6], 48) | SHL(b[7], 56);
-#undef SHL
-  b += sizeof(uint64_t);
-  *buf = b;
+#if UPB_UNALIGNED_READS_OK
+  *val = *(uint64_t*)*buf;
+  *buf = (char*)*buf + sizeof(uint64_t);
+#else
+  uint32_t lo32, hi32;
+  get_f_uint32_t(buf, &lo32);
+  get_f_uint32_t(buf, &hi32);
+  *val = lo32 | ((uint64_t)hi32 << 32);
+#endif
   return UPB_STATUS_OK;
 }
 
-- 
cgit v1.2.3