tune for better branch prediction by avoiding more than 3 in a 16-byte block

amd_d_nrv2b.S amd_d_nrv2e.S l_lx_elf64amd.S committer: jreiser <jreiser> 1131994471 +0000
2005-11-14 18:54:31 +00:00
parent 2558243c83
commit beb4319d1e
3 changed files with 26 additions and 13 deletions
@@ -34,7 +34,7 @@ lit_n2e:
        incq %rsi; movb %dl,(%rdi)
        incq %rdi
 top_n2e:
-        movzbl (%rsi),%edx  # speculate: literal, or bottom 8 bits of offset
+        movb (%rsi),%dl  # speculate: literal, or bottom 8 bits of offset
        jnextb1y lit_n2e
        lea 1(lenq),off  # [len= 0] off= 1
        jmp getoff_n2e
@@ -47,7 +47,7 @@ getoff_n2e:
        jnextb0n off_n2e

        subl $ 3,off; jc offprev_n2e
-        shll $ 8,off
+        shll $ 8,off; movzbl %dl,%edx
        orl %edx,off; incq %rsi
        xorl $~0,off; jz eof
        sarl off  # Carry= original low bit