diff options
| author | Alan Modra <amodra@gmail.com> | 2013-08-17 18:47:59 +0930 |
|---|---|---|
| committer | Alan Modra <amodra@gmail.com> | 2013-10-04 10:41:35 +0930 |
| commit | 3be87c77d24c4456ccca4034363b6d1814cd0c84 (patch) | |
| tree | c81589a21d74b94d5ae20235bc6d81715eca858c /sysdeps/powerpc/powerpc64/power4 | |
| parent | 759cfef3ac4c07dba1ece0bbc1207e099348816d (diff) | |
| download | glibc-3be87c77d24c4456ccca4034363b6d1814cd0c84.zip glibc-3be87c77d24c4456ccca4034363b6d1814cd0c84.tar.gz glibc-3be87c77d24c4456ccca4034363b6d1814cd0c84.tar.bz2 | |
PowerPC LE memset
http://sourceware.org/ml/libc-alpha/2013-08/msg00104.html
One of the things I noticed when looking at power7 timing is that rlwimi
is cracked and the two resulting insns have a register dependency.
That makes it a little slower than the equivalent rldimi.
* sysdeps/powerpc/powerpc64/memset.S: Replace rlwimi with
insrdi. Formatting.
* sysdeps/powerpc/powerpc64/power4/memset.S: Likewise.
* sysdeps/powerpc/powerpc64/power6/memset.S: Likewise.
* sysdeps/powerpc/powerpc64/power7/memset.S: Likewise.
* sysdeps/powerpc/powerpc32/power4/memset.S: Likewise.
* sysdeps/powerpc/powerpc32/power6/memset.S: Likewise.
* sysdeps/powerpc/powerpc32/power7/memset.S: Likewise.
Diffstat (limited to 'sysdeps/powerpc/powerpc64/power4')
| -rw-r--r-- | sysdeps/powerpc/powerpc64/power4/memset.S | 30 |
1 files changed, 15 insertions, 15 deletions
diff --git a/sysdeps/powerpc/powerpc64/power4/memset.S b/sysdeps/powerpc/powerpc64/power4/memset.S index dbecee8..ad0d381 100644 --- a/sysdeps/powerpc/powerpc64/power4/memset.S +++ b/sysdeps/powerpc/powerpc64/power4/memset.S @@ -50,14 +50,14 @@ L(_memset): /* Align to doubleword boundary. */ cmpldi cr5, rLEN, 31 - rlwimi rCHR, rCHR, 8, 16, 23 /* Replicate byte to halfword. */ + insrdi rCHR, rCHR, 8, 48 /* Replicate byte to halfword. */ beq+ L(aligned2) mtcrf 0x01, rMEMP0 subfic rALIGN, rALIGN, 8 cror 28,30,31 /* Detect odd word aligned. */ add rMEMP, rMEMP, rALIGN sub rLEN, rLEN, rALIGN - rlwimi rCHR, rCHR, 16, 0, 15 /* Replicate halfword to word. */ + insrdi rCHR, rCHR, 16, 32 /* Replicate halfword to word. */ bt 29, L(g4) /* Process the even word of doubleword. */ bf+ 31, L(g2) @@ -79,14 +79,14 @@ L(g0): /* Handle the case of size < 31. */ L(aligned2): - rlwimi rCHR, rCHR, 16, 0, 15 /* Replicate halfword to word. */ + insrdi rCHR, rCHR, 16, 32 /* Replicate halfword to word. */ L(aligned): mtcrf 0x01, rLEN ble cr5, L(medium) /* Align to 32-byte boundary. */ andi. rALIGN, rMEMP, 0x18 subfic rALIGN, rALIGN, 0x20 - insrdi rCHR,rCHR,32,0 /* Replicate word to double word. */ + insrdi rCHR, rCHR, 32, 0 /* Replicate word to double word. */ beq L(caligned) mtcrf 0x01, rALIGN add rMEMP, rMEMP, rALIGN @@ -146,24 +146,24 @@ L(zloopstart): L(getCacheAligned): cmpldi cr1,rLEN,32 andi. rTMP,rMEMP,127 - blt cr1,L(handletail32) - beq L(cacheAligned) + blt cr1,L(handletail32) + beq L(cacheAligned) addi rMEMP,rMEMP,32 addi rLEN,rLEN,-32 - std rCHR,-32(rMEMP) - std rCHR,-24(rMEMP) - std rCHR,-16(rMEMP) - std rCHR,-8(rMEMP) - b L(getCacheAligned) + std rCHR,-32(rMEMP) + std rCHR,-24(rMEMP) + std rCHR,-16(rMEMP) + std rCHR,-8(rMEMP) + b L(getCacheAligned) /* Now we are aligned to the cache line and can use dcbz. */ L(cacheAligned): cmpld cr1,rLEN,rCLS - blt cr1,L(handletail32) + blt cr1,L(handletail32) dcbz 0,rMEMP subf rLEN,rCLS,rLEN - add rMEMP,rMEMP,rCLS - b L(cacheAligned) + add rMEMP,rMEMP,rCLS + b L(cacheAligned) /* We are here because the cache line size was set and was not 32-bytes and the remainder (rLEN) is less than the actual cache line size. @@ -200,7 +200,7 @@ L(le4): /* Memset of 0-31 bytes. */ .align 5 L(medium): - insrdi rCHR,rCHR,32,0 /* Replicate word to double word. */ + insrdi rCHR, rCHR, 32, 0 /* Replicate word to double word. */ cmpldi cr1, rLEN, 16 L(medium_tail2): add rMEMP, rMEMP, rLEN |