diff options
author | Will Schmidt <will_schmidt@vnet.ibm.com> | 2012-08-21 14:20:55 -0500 |
---|---|---|
committer | Ryan S. Arnold <rsa@linux.vnet.ibm.com> | 2012-08-21 14:20:55 -0500 |
commit | 14a50c9d238a8d79d48193482ed9a2be52d1dc3d (patch) | |
tree | 046c03b3bf03143c65e3f4eab969391df13e4dd1 | |
parent | 696da85994af013ae1e3e8ad3bb016793f9b2c5d (diff) | |
download | glibc-14a50c9d238a8d79d48193482ed9a2be52d1dc3d.tar.gz glibc-14a50c9d238a8d79d48193482ed9a2be52d1dc3d.tar.xz glibc-14a50c9d238a8d79d48193482ed9a2be52d1dc3d.zip |
[Powerpc] Tune/optimize powerpc{32,64}/power7/memchr.S.
Assorted tweaking, twisting and tuning to squeeze a few additional cycles out of the memchr code. Changes include bypassing the shift pairs (sld,srd) when they are not required, and unrolling the small_loop that handles short and trailing strings. Per scrollpipe data measuring aligned strings for 64-bit, these changes save between five and eight cycles (9-13% overall) for short strings (<32), Longer aligned strings see slight improvement of 1-3% due to bypassing the shifts and the instruction rearranging.
-rw-r--r-- | ChangeLog | 6 | ||||
-rw-r--r-- | sysdeps/powerpc/powerpc32/power7/memchr.S | 78 | ||||
-rw-r--r-- | sysdeps/powerpc/powerpc64/power7/memchr.S | 85 |
3 files changed, 125 insertions, 44 deletions
diff --git a/ChangeLog b/ChangeLog index 0c34ac30cf..12e5b4c2df 100644 --- a/ChangeLog +++ b/ChangeLog @@ -1,3 +1,9 @@ +2012-08-21 Will Schmidt <will_schmidt@vnet.ibm.com> + + * sysdeps/powerpc/powerpc64/power7/memchr.S: Unrolled short loop and + slight instruction rearrangements per scrollpipe analysis. + * sysdeps/powerpc/powerpc64/power7/memchr.S: Likewise. + 2012-08-20 Roland McGrath <roland@hack.frob.com> * manual/syslog.texi (syslog; vsyslog, closelog): diff --git a/sysdeps/powerpc/powerpc32/power7/memchr.S b/sysdeps/powerpc/powerpc32/power7/memchr.S index 85cc2999ae..ca2bc49579 100644 --- a/sysdeps/powerpc/powerpc32/power7/memchr.S +++ b/sysdeps/powerpc/powerpc32/power7/memchr.S @@ -1,5 +1,5 @@ /* Optimized memchr implementation for PowerPC32/POWER7 using cmpb insn. - Copyright (C) 2010 Free Software Foundation, Inc. + Copyright (C) 2010-2012 Free Software Foundation, Inc. Contributed by Luis Machado <luisgpm@br.ibm.com>. This file is part of the GNU C Library. @@ -33,19 +33,23 @@ ENTRY (BP_SYM (__memchr)) cmplwi r5,16 ble L(small_range) - cmplw cr7,r3,r7 /* Is the address equal or less than r3? If - it's equal or less, it means size is either 0 - or a negative number. */ + cmplw cr7,r3,r7 /* Compare the starting address (r3) with the + ending address (r7). If (r3 >= r7), the size + passed in is zero or negative. */ ble cr7,L(proceed) - li r7,-1 /* Make r11 the biggest if r4 <= 0. */ + li r7,-1 /* Artificially set our ending address (r7) + such that we will exit early. */ L(proceed): rlwinm r6,r3,3,27,28 /* Calculate padding. */ + cmpli cr6,r6,0 /* cr6 == Do we have padding? */ lwz r12,0(r8) /* Load word from memory. */ cmpb r10,r12,r4 /* Check for BYTE's in WORD1. */ + beq cr6,L(proceed_no_padding) slw r10,r10,r6 srw r10,r10,r6 - cmplwi cr7,r10,0 /* If r10 == 0, no BYTE's have been found. */ +L(proceed_no_padding): + cmplwi cr7,r10,0 /* If r10 == 0, no BYTEs have been found. */ bne cr7,L(done) /* Are we done already? */ @@ -73,7 +77,7 @@ L(proceed): L(loop_setup): sub r5,r7,r9 srwi r6,r5,3 /* Number of loop iterations. */ - mtctr r6 /* Setup the counter. */ + mtctr r6 /* Setup the counter. */ b L(loop) /* Main loop to look for BYTE backwards in the string. Since it's a small loop (< 8 instructions), align it to 32-bytes. */ @@ -82,7 +86,6 @@ L(loop): /* Load two words, compare and merge in a single register for speed. This is an attempt to speed up the byte-checking process for bigger strings. */ - lwz r12,4(r8) lwzu r11,8(r8) cmpb r10,r12,r4 @@ -91,11 +94,11 @@ L(loop): cmplwi cr7,r5,0 bne cr7,L(found) bdnz L(loop) + /* We're here because the counter reached 0, and that means we - didn't have any matches for BYTE in the whole range. Just - return the original range. */ - addi r9,r8,4 - cmplw cr6,r9,r7 + didn't have any matches for BYTE in the whole range. */ + subi r11,r7,4 + cmplw cr6,r8,r11 blt cr6,L(loop_small) b L(null) @@ -118,8 +121,7 @@ L(found): /* r10 has the output of the cmpb instruction, that is, it contains 0xff in the same position as BYTE in the original word from the string. Use that to calculate the pointer. - We need to make sure BYTE is *before* the end of the - range. */ + We need to make sure BYTE is *before* the end of the range. */ L(done): cntlzw r0,r10 /* Count leading zeroes before the match. */ srwi r0,r0,3 /* Convert leading zeroes to bytes. */ @@ -137,13 +139,15 @@ L(null): .align 4 L(small_range): cmplwi r5,0 - beq L(null) - rlwinm r6,r3,3,27,28 /* Calculate padding. */ + beq L(null) /* This branch is for the cmplwi r5,0 above */ lwz r12,0(r8) /* Load word from memory. */ + cmplwi cr6,r6,0 /* cr6 == Do we have padding? */ cmpb r10,r12,r4 /* Check for BYTE in DWORD1. */ + beq cr6,L(small_no_padding) slw r10,r10,r6 srw r10,r10,r6 +L(small_no_padding): cmplwi cr7,r10,0 bne cr7,L(done) @@ -151,18 +155,52 @@ L(small_range): addi r9,r8,4 cmplw r9,r7 bge L(null) - b L(loop_small) - .p2align 5 -L(loop_small): +L(loop_small): /* loop_small has been unrolled. */ lwzu r12,4(r8) cmpb r10,r12,r4 addi r9,r8,4 cmplwi cr6,r10,0 + cmplw r9,r7 bne cr6,L(done) + bge L(null) + + lwzu r12,4(r8) + cmpb r10,r12,r4 + addi r9,r8,4 + cmplwi cr6,r10,0 cmplw r9,r7 + bne cr6,L(done) bge L(null) - b L(loop_small) + + lwzu r12,4(r8) + cmpb r10,r12,r4 + addi r9,r8,4 + cmplwi cr6,r10,0 + cmplw r9,r7 + bne cr6,L(done) + bge L(null) + + lwzu r12,4(r8) + cmpb r10,r12,r4 + addi r9,r8,4 + cmplwi cr6,r10,0 + cmplw r9,r7 + bne cr6,L(done) + bge L(null) + + /* For most cases we will never get here. Under some combinations of + padding + length there is a leftover word that still needs to be + checked. */ + lwzu r12,4(r8) + cmpb r10,r12,r4 + addi r9,r8,4 + cmplwi cr6,r10,0 + bne cr6,L(done) + + /* save a branch and exit directly */ + li r3,0 + blr END (BP_SYM (__memchr)) weak_alias (BP_SYM (__memchr), BP_SYM(memchr)) diff --git a/sysdeps/powerpc/powerpc64/power7/memchr.S b/sysdeps/powerpc/powerpc64/power7/memchr.S index f0dbf60d0b..c125a5b354 100644 --- a/sysdeps/powerpc/powerpc64/power7/memchr.S +++ b/sysdeps/powerpc/powerpc64/power7/memchr.S @@ -1,5 +1,5 @@ /* Optimized memchr implementation for PowerPC64/POWER7 using cmpb insn. - Copyright (C) 2010 Free Software Foundation, Inc. + Copyright (C) 2010-2012 Free Software Foundation, Inc. Contributed by Luis Machado <luisgpm@br.ibm.com>. This file is part of the GNU C Library. @@ -29,27 +29,32 @@ ENTRY (BP_SYM (__memchr)) clrrdi r8,r3,3 rlwimi r4,r4,8,16,23 rlwimi r4,r4,16,0,15 - insrdi r4,r4,32,0 add r7,r3,r5 /* Calculate the last acceptable address. */ cmpldi r5,32 + insrdi r4,r4,32,0 ble L(small_range) - cmpld cr7,r3,r7 /* Is the address equal or less than r3? If - it's equal or less, it means size is either 0 - or a negative number. */ + cmpld cr7,r3,r7 /* Compare the starting address (r3) with the + ending address (r7). If (r3 >= r7), + the size passed in was zero or negative. */ ble cr7,L(proceed) - li r7,-1 /* Make r11 the biggest if r4 <= 0. */ + li r7,-1 /* Artificially set our ending address (r7) + such that we will exit early. */ + L(proceed): rlwinm r6,r3,3,26,28 /* Calculate padding. */ + cmpldi cr6,r6,0 /* cr6 == Do we have padding? */ ld r12,0(r8) /* Load doubleword from memory. */ - cmpb r10,r12,r4 /* Check for BYTE's in DWORD1. */ + cmpb r10,r12,r4 /* Check for BYTEs in DWORD1. */ + beq cr6,L(proceed_no_padding) sld r10,r10,r6 srd r10,r10,r6 - cmpldi cr7,r10,0 /* If r10 == 0, no BYTE's's have been found. */ +L(proceed_no_padding): + cmpldi cr7,r10,0 /* Does r10 indicate we got a hit? */ bne cr7,L(done) - /* Are we done already? */ + /* See if we are at the last acceptable address yet. */ addi r9,r8,8 cmpld cr6,r9,r7 bge cr6,L(null) @@ -74,7 +79,7 @@ L(proceed): L(loop_setup): sub r5,r7,r9 srdi r6,r5,4 /* Number of loop iterations. */ - mtctr r6 /* Setup the counter. */ + mtctr r6 /* Setup the counter. */ b L(loop) /* Main loop to look for BYTE backwards in the string. Since it's a small loop (< 8 instructions), align it to 32-bytes. */ @@ -83,7 +88,6 @@ L(loop): /* Load two doublewords, compare and merge in a single register for speed. This is an attempt to speed up the byte-checking process for bigger strings. */ - ld r12,8(r8) ldu r11,16(r8) cmpb r10,r12,r4 @@ -92,11 +96,11 @@ L(loop): cmpldi cr7,r5,0 bne cr7,L(found) bdnz L(loop) + /* We're here because the counter reached 0, and that means we - didn't have any matches for BYTE in the whole range. Just return - the original range. */ - addi r9,r8,8 - cmpld cr6,r9,r7 + didn't have any matches for BYTE in the whole range. */ + subi r11,r7,8 + cmpld cr6,r8,r11 blt cr6,L(loop_small) b L(null) @@ -119,8 +123,7 @@ L(found): /* r10 has the output of the cmpb instruction, that is, it contains 0xff in the same position as BYTE in the original doubleword from the string. Use that to calculate the pointer. - We need to make sure BYTE is *before* the end of the - range. */ + We need to make sure BYTE is *before* the end of the range. */ L(done): cntlzd r0,r10 /* Count leading zeroes before the match. */ srdi r0,r0,3 /* Convert leading zeroes to bytes. */ @@ -138,13 +141,16 @@ L(null): .align 4 L(small_range): cmpldi r5,0 - beq L(null) - rlwinm r6,r3,3,26,28 /* Calculate padding. */ + beq L(null) /* This branch is for the cmpldi r5,0 above. */ ld r12,0(r8) /* Load word from memory. */ + cmpldi cr6,r6,0 /* cr6 == Do we have padding? */ cmpb r10,r12,r4 /* Check for BYTE in DWORD1. */ + /* If no padding, skip the shifts. */ + beq cr6,L(small_no_padding) sld r10,r10,r6 srd r10,r10,r6 +L(small_no_padding): cmpldi cr7,r10,0 bne cr7,L(done) @@ -152,18 +158,49 @@ L(small_range): addi r9,r8,8 cmpld r9,r7 bge L(null) - b L(loop_small) + /* If we're not done, drop through into loop_small. */ + +L(loop_small): /* loop_small has been unrolled. */ + ldu r12,8(r8) + cmpb r10,r12,r4 + addi r9,r8,8 + cmpldi cr6,r10,0 + cmpld r9,r7 + bne cr6,L(done) /* Found something. */ + bge L(null) /* Hit end of string (length). */ - .p2align 5 -L(loop_small): ldu r12,8(r8) cmpb r10,r12,r4 addi r9,r8,8 cmpldi cr6,r10,0 - bne cr6,L(done) cmpld r9,r7 + bne cr6,L(done) /* Found something. */ bge L(null) - b L(loop_small) + + ldu r12,8(r8) + subi r11,r7,8 + cmpb r10,r12,r4 + cmpldi cr6,r10,0 + ori r2,r2,0 /* Force a dispatch group. */ + bne cr6,L(done) + + cmpld r8,r11 /* At end of range? */ + bge L(null) + + /* For most cases we will never get here. Under some combinations of + padding + length there is a leftover double that still needs to be + checked. */ + ldu r12,8(r8) + cmpb r10,r12,r4 + addi r9,r8,8 + cmpldi cr6,r10,0 + cmpld r9,r7 + bne cr6,L(done) /* Found something. */ + + /* Save a branch and exit directly. */ + li r3,0 + blr + END (BP_SYM (__memchr)) weak_alias (BP_SYM (__memchr), BP_SYM(memchr)) |