diff options
| author | Siddhesh Poyarekar <siddhesh@sourceware.org> | 2017-12-13 18:50:27 +0530 |
|---|---|---|
| committer | Wilco Dijkstra <wdijkstr@arm.com> | 2019-09-06 17:13:02 +0100 |
| commit | 4e75091d6ce3f7ac8b1750ca6135bc37d6707caf (patch) | |
| tree | 82fdb5df1150aeeb72cff2ad6b354d719f854e90 | |
| parent | 8569357e11aa7b8e912142727eac1d106c785433 (diff) | |
| download | glibc-4e75091d6ce3f7ac8b1750ca6135bc37d6707caf.tar glibc-4e75091d6ce3f7ac8b1750ca6135bc37d6707caf.tar.gz glibc-4e75091d6ce3f7ac8b1750ca6135bc37d6707caf.tar.bz2 glibc-4e75091d6ce3f7ac8b1750ca6135bc37d6707caf.zip | |
aarch64: Improve strcmp unaligned performance
Replace the simple byte-wise compare in the misaligned case with a
dword compare with page boundary checks in place. For simplicity I've
chosen a 4K page boundary so that we don't have to query the actual
page size on the system.
This results in up to 3x improvement in performance in the unaligned
case on falkor and about 2.5x improvement on mustang as measured using
bench-strcmp.
* sysdeps/aarch64/strcmp.S (misaligned8): Compare dword at a
time whenever possible.
(cherry picked from commit 2bce01ebbaf8db52ba4a5635eb5744f989cdbf69)
| -rw-r--r-- | ChangeLog | 5 | ||||
| -rw-r--r-- | sysdeps/aarch64/strcmp.S | 31 |
2 files changed, 34 insertions, 2 deletions
@@ -1,5 +1,10 @@ 2019-09-06 Siddhesh Poyarekar <siddhesh@sourceware.org> + * sysdeps/aarch64/strcmp.S (misaligned8): Compare dword at a + time whenever possible. + +2019-09-06 Siddhesh Poyarekar <siddhesh@sourceware.org> + * sysdeps/aarch64/memcmp.S (more16): Fix loop16 branch target. * sysdeps/aarch64/memcmp.S: Widen comparison to 16 bytes at a diff --git a/sysdeps/aarch64/strcmp.S b/sysdeps/aarch64/strcmp.S index e99d6625b7..c260e1d8ac 100644 --- a/sysdeps/aarch64/strcmp.S +++ b/sysdeps/aarch64/strcmp.S @@ -72,6 +72,7 @@ L(start_realigned): cbz syndrome, L(loop_aligned) /* End of performance-critical section -- one 64B cache line. */ +L(end): #ifndef __AARCH64EB__ rev syndrome, syndrome rev data1, data1 @@ -145,12 +146,38 @@ L(mutual_align): b L(start_realigned) L(misaligned8): - /* We can do better than this. */ + /* Align SRC1 to 8 bytes and then compare 8 bytes at a time, always + checking to make sure that we don't access beyond page boundary in + SRC2. */ + tst src1, #7 + b.eq L(loop_misaligned) +L(do_misaligned): ldrb data1w, [src1], #1 ldrb data2w, [src2], #1 cmp data1w, #1 ccmp data1w, data2w, #0, cs /* NZCV = 0b0000. */ - b.eq L(misaligned8) + b.ne L(done) + tst src1, #7 + b.ne L(misaligned8) + +L(loop_misaligned): + /* Test if we are within the last dword of the end of a 4K page. If + yes then jump back to the misaligned loop to copy a byte at a time. */ + and tmp1, src2, #0xff8 + eor tmp1, tmp1, #0xff8 + cbz tmp1, L(do_misaligned) + ldr data1, [src1], #8 + ldr data2, [src2], #8 + + sub tmp1, data1, zeroones + orr tmp2, data1, #REP8_7f + eor diff, data1, data2 /* Non-zero if differences found. */ + bic has_nul, tmp1, tmp2 /* Non-zero if NUL terminator. */ + orr syndrome, diff, has_nul + cbz syndrome, L(loop_misaligned) + b L(end) + +L(done): sub result, data1, data2 RET END(strcmp) |