diff options
| author | Denis Vlasenko <vda.linux@googlemail.com> | 2008-04-15 08:27:24 +0000 |
|---|---|---|
| committer | Denis Vlasenko <vda.linux@googlemail.com> | 2008-04-15 08:27:24 +0000 |
| commit | df7958a9606a342e3c3ac5a40fc41f3a79669d62 (patch) | |
| tree | f22657788d4ca4bd427f7ff7247dfa353590ac9c /libc/string/x86_64/strcat.S | |
| parent | 534dfb536f19737f2642ee56dd67a97c5db6a74e (diff) | |
amd64 string ops: use alignment more carefully, and comment it.
By capping max padding to not be bigger than three next insns,
we avoid having ridiculously big NOPs like this one:
53:66 66 66 66 2e 0f 1f nopw %cs:0x0(%rax,%rax,1)
5a:84 00 00 00 00 00
which was bigger than next three insns combined!
Size changes:
text data bss dec hex filename
102 0 0 102 66 x86_64/memcpy.o
102 0 0 102 66 x86_64.old/memcpy.o
90 0 0 90 5a x86_64/mempcpy.o
102 0 0 102 66 x86_64.old/mempcpy.o
210 0 0 210 d2 x86_64/memset.o
242 0 0 242 f2 x86_64.old/memset.o
213 0 0 213 d5 x86_64/stpcpy.o
220 0 0 220 dc x86_64.old/stpcpy.o
428 0 0 428 1ac x86_64/strcat.o
444 0 0 444 1bc x86_64.old/strcat.o
417 0 0 417 1a1 x86_64/strchr.o
418 0 0 418 1a2 x86_64.old/strchr.o
33 0 0 33 21 x86_64/strcmp.o
33 0 0 33 21 x86_64.old/strcmp.o
213 0 0 213 d5 x86_64/strcpy.o
220 0 0 220 dc x86_64.old/strcpy.o
135 0 0 135 87 x86_64/strcspn.o
151 0 0 151 97 x86_64.old/strcspn.o
225 0 0 225 e1 x86_64/strlen.o
233 0 0 233 e9 x86_64.old/strlen.o
140 0 0 140 8c x86_64/strpbrk.o
156 0 0 156 9c x86_64.old/strpbrk.o
135 0 0 135 87 x86_64/strspn.o
151 0 0 151 97 x86_64.old/strspn.o
Also, a few files got their .text alignment relaxed from 16 to 8 bytes,
which reduces padding at link time.
Diffstat (limited to 'libc/string/x86_64/strcat.S')
| -rw-r--r-- | libc/string/x86_64/strcat.S | 19 |
1 files changed, 14 insertions, 5 deletions
diff --git a/libc/string/x86_64/strcat.S b/libc/string/x86_64/strcat.S index 9e2da50f2..23d068fea 100644 --- a/libc/string/x86_64/strcat.S +++ b/libc/string/x86_64/strcat.S @@ -45,7 +45,9 @@ ENTRY (BP_SYM (strcat)) /* Now the source is aligned. Scan for NUL byte. */ - .p2align 4 + + /* Next 3 insns are 10 bytes total, make sure we decode them in one go */ + .p2align 4,,10 4: /* First unroll. */ movq (%rax), %rcx /* get double word (= 8 bytes) in question */ @@ -103,8 +105,11 @@ ENTRY (BP_SYM (strcat)) the addition will not result in 0. */ jz 4b /* no NUL found => continue loop */ - .p2align 4 /* Align, it is a jump target. */ -3: subq $8,%rax /* correct pointer increment. */ + /* Align, it is a jump target. */ + /* Next 3 insns are 8 bytes total, make sure we decode them in one go */ + .p2align 3,,8 +3: + subq $8,%rax /* correct pointer increment. */ testb %cl, %cl /* is first byte NUL? */ jz 2f /* yes => return */ @@ -160,7 +165,9 @@ ENTRY (BP_SYM (strcat)) /* Now the sources is aligned. Unfortunatly we cannot force to have both source and destination aligned, so ignore the alignment of the destination. */ - .p2align 4 + + /* Next 3 insns are 10 bytes total, make sure we decode them in one go */ + .p2align 4,,10 22: /* 1st unroll. */ movq (%rsi), %rax /* Read double word (8 bytes). */ @@ -237,7 +244,9 @@ ENTRY (BP_SYM (strcat)) /* Do the last few bytes. %rax contains the value to write. The loop is unrolled twice. */ - .p2align 4 + + /* Next 3 insns are 6 bytes total, make sure we decode them in one go */ + .p2align 3,,6 23: movb %al, (%rdx) /* 1st byte. */ testb %al, %al /* Is it NUL. */ |