* [PATCH 1/5] alpha: put mcount prologue at correct location
@ 2012-06-06 21:52 Richard Henderson
2012-06-06 21:52 ` [PATCH 3/5] alpha: Fix ev4 build with ev6 compiler Richard Henderson
` (3 more replies)
0 siblings, 4 replies; 5+ messages in thread
From: Richard Henderson @ 2012-06-06 21:52 UTC (permalink / raw)
To: libc-ports
---
ChangeLog.alpha | 4 ++++
sysdeps/alpha/_mcount.S | 3 +--
2 files changed, 5 insertions(+), 2 deletions(-)
diff --git a/ChangeLog.alpha b/ChangeLog.alpha
index 705a020..2c107ad 100644
--- a/ChangeLog.alpha
+++ b/ChangeLog.alpha
@@ -1,3 +1,7 @@
+2012-06-06 Richard Henderson <rth@twiddle.net>
+
+ * sysdeps/alpha/_mcount.S: Move .prologue after stack alloc.
+
2012-06-05 Richard Henderson <rth@twiddle.net>
* sysdeps/unix/sysv/linux/alpha/rt_sigaction.S: Use .cfi_signal_frame
diff --git a/sysdeps/alpha/_mcount.S b/sysdeps/alpha/_mcount.S
index 9d7c15b..67c579e 100644
--- a/sysdeps/alpha/_mcount.S
+++ b/sysdeps/alpha/_mcount.S
@@ -38,9 +38,8 @@
.set noreorder
LEAF(_mcount, 0xb0)
- .prologue 0
-
subq sp, 0xb0, sp
+ .prologue 0
stq a0, 0x00(sp)
mov ra, a0 # a0 = caller-pc
stq a1, 0x08(sp)
--
1.7.7.6
^ permalink raw reply [flat|nested] 5+ messages in thread
* [PATCH 2/5] alpha: Use cfi_startproc instead of dual procedure descriptors
2012-06-06 21:52 [PATCH 1/5] alpha: put mcount prologue at correct location Richard Henderson
` (2 preceding siblings ...)
2012-06-06 21:52 ` [PATCH 4/5] alpha: Fix [BZ #13718] Richard Henderson
@ 2012-06-06 21:52 ` Richard Henderson
3 siblings, 0 replies; 5+ messages in thread
From: Richard Henderson @ 2012-06-06 21:52 UTC (permalink / raw)
To: libc-ports
---
ChangeLog.alpha | 9 +++++++
sysdeps/alpha/alphaev6/stxcpy.S | 33 +++++++-----------------
sysdeps/alpha/alphaev6/stxncpy.S | 49 ++++++++++++-------------------------
sysdeps/alpha/bzero.S | 21 ++++-----------
sysdeps/alpha/memset.S | 25 ++++++-------------
sysdeps/alpha/stxcpy.S | 30 +++++++----------------
sysdeps/alpha/stxncpy.S | 31 +++++++-----------------
sysdeps/unix/alpha/sysdep.h | 6 ++++
8 files changed, 73 insertions(+), 131 deletions(-)
diff --git a/ChangeLog.alpha b/ChangeLog.alpha
index 2c107ad..ceee910 100644
--- a/ChangeLog.alpha
+++ b/ChangeLog.alpha
@@ -1,5 +1,14 @@
2012-06-06 Richard Henderson <rth@twiddle.net>
+ * sysdeps/alpha/alphaev6/stxcpy.S: Use cfi markup instead of
+ dual ecoff procedure descriptors.
+ * sysdeps/alpha/alphaev6/stxncpy.S: Likewise.
+ * sysdeps/alpha/bzero.S: Likewise.
+ * sysdeps/alpha/memset.S: Likewise.
+ * sysdeps/alpha/stxcpy.S: Likewise.
+ * sysdeps/alpha/stxncpy.S: Likewise.
+ * sysdeps/unix/alpha/sysdep.h (USEPV_PROF): New.
+
* sysdeps/alpha/_mcount.S: Move .prologue after stack alloc.
2012-06-05 Richard Henderson <rth@twiddle.net>
diff --git a/sysdeps/alpha/alphaev6/stxcpy.S b/sysdeps/alpha/alphaev6/stxcpy.S
index 37ad737..bdc8e72 100644
--- a/sysdeps/alpha/alphaev6/stxcpy.S
+++ b/sysdeps/alpha/alphaev6/stxcpy.S
@@ -1,4 +1,4 @@
-/* Copyright (C) 2000 Free Software Foundation, Inc.
+/* Copyright (C) 2000-2012 Free Software Foundation, Inc.
Contributed by Richard Henderson (rth@tamu.edu)
EV6 optimized by Rick Gorton <rick.gorton@alpha-processor.com>.
This file is part of the GNU C Library.
@@ -41,24 +41,20 @@
.arch ev6
.set noat
.set noreorder
- .text
-
-/* There is a problem with either gdb (as of 4.16) or gas (as of 2.7) that
- doesn't like putting the entry point for a procedure somewhere in the
- middle of the procedure descriptor. Work around this by putting the
- aligned copy in its own procedure descriptor */
+ .text
+ .type __stxcpy, @function
+ .globl __stxcpy
+ .usepv __stxcpy, no
- .ent stxcpy_aligned
- .align 4
-stxcpy_aligned:
- .frame sp, 0, t9
- .prologue 0
+ cfi_startproc
+ cfi_return_column (t9)
/* On entry to this basic block:
t0 == the first destination word for masking back in
t1 == the first source word. */
-
+ .align 4
+stxcpy_aligned:
/* Create the 1st output word and detect 0's in the 1st input word. */
lda t2, -1 # E : build a mask against false zero
mskqh t2, a1, t2 # U : detection in the src word (stall)
@@ -115,15 +111,8 @@ $a_eos:
nop
nop
- .end stxcpy_aligned
-
.align 4
- .ent __stxcpy
- .globl __stxcpy
__stxcpy:
- .frame sp, 0, t9
- .prologue 0
-
/* Are source and destination co-aligned? */
xor a0, a1, t0 # E :
unop # E :
@@ -321,7 +310,5 @@ $unaligned:
or t0, t1, t1 # e1 : and put it there
stq_u t1, 0(a0) # .. e0 : (stall)
ret (t9) # e1 :
- nop
-
- .end __stxcpy
+ cfi_endproc
diff --git a/sysdeps/alpha/alphaev6/stxncpy.S b/sysdeps/alpha/alphaev6/stxncpy.S
index 31f8570..d134eb8 100644
--- a/sysdeps/alpha/alphaev6/stxncpy.S
+++ b/sysdeps/alpha/alphaev6/stxncpy.S
@@ -1,4 +1,4 @@
-/* Copyright (C) 2000, 2002 Free Software Foundation, Inc.
+/* Copyright (C) 2000-2012 Free Software Foundation, Inc.
Contributed by Richard Henderson (rth@tamu.edu)
EV6 optimized by Rick Gorton <rick.gorton@alpha-processor.com>.
This file is part of the GNU C Library.
@@ -49,22 +49,19 @@
.set noat
.set noreorder
-/* There is a problem with either gdb (as of 4.16) or gas (as of 2.7) that
- doesn't like putting the entry point for a procedure somewhere in the
- middle of the procedure descriptor. Work around this by putting the
- aligned copy in its own procedure descriptor */
+ .text
+ .type __stxncpy, @function
+ .globl __stxncpy
+ .usepv __stxncpy, no
-
- .ent stxncpy_aligned
- .align 4
-stxncpy_aligned:
- .frame sp, 0, t9, 0
- .prologue 0
+ cfi_startproc
+ cfi_return_column (t9)
/* On entry to this basic block:
t0 == the first destination word for masking back in
t1 == the first source word. */
-
+ .align 4
+stxncpy_aligned:
/* Create the 1st output word and detect 0's in the 1st input word. */
lda t2, -1 # E : build a mask against false zero
mskqh t2, a1, t2 # U : detection in the src word (stall)
@@ -111,7 +108,6 @@ $a_loop:
On entry to this basic block we have:
t0 == the source word containing the null
t7 == the cmpbge mask that found it. */
-
$a_eos:
negq t7, t8 # E : find low bit set
and t7, t8, t8 # E : (stall)
@@ -144,15 +140,8 @@ $a_eoc:
nop
nop
- .end stxncpy_aligned
-
.align 4
- .ent __stxncpy
- .globl __stxncpy
__stxncpy:
- .frame sp, 0, t9, 0
- .prologue 0
-
/* Are source and destination co-aligned? */
xor a0, a1, t1 # E :
and a0, 7, t0 # E : find dest misalignment
@@ -166,16 +155,14 @@ __stxncpy:
sll t10, t2, t10 # U : t10 = bitmask of last count byte
bne t1, $unaligned # U :
+
/* We are co-aligned; take care of a partial first word. */
ldq_u t1, 0(a1) # L : load first src word
addq a1, 8, a1 # E :
-
- beq t0, stxncpy_aligned # U : avoid loading dest word if not needed
+ beq t0, stxncpy_aligned # U : avoid loading dest word if not needed
ldq_u t0, 0(a0) # L :
- nop
- nop
- br stxncpy_aligned # .. e1 :
+ br stxncpy_aligned # U :
nop
nop
nop
@@ -231,7 +218,7 @@ $u_head:
extqh t2, a1, t0 # U : position lo-bits of hi word (stall)
cmpbge zero, t2, t7 # E :
- nop
+ nop
bne t7, $u_eos # U :
/* Unaligned copy main loop. In order to avoid reading too much,
@@ -314,7 +301,7 @@ $u_final:
1: stq_u t0, 0(a0) # L :
ret (t9) # L0 : Latency=3
- /* Got to end-of-count before end of string.
+ /* Got to end-of-count before end of string.
On entry to this basic block:
t1 == the shifted high-order bits from the previous source word */
$u_eoc:
@@ -325,7 +312,7 @@ $u_eoc:
ldq_u t2, 8(a1) # L : load final src word
nop
- extqh t2, a1, t0 # U : extract low bits for last word (stall)
+ extqh t2, a1, t0 # U : extract low bits for last word (stall)
or t1, t0, t1 # E : (stall)
1: cmpbge zero, t1, t7 # E :
@@ -394,9 +381,5 @@ $unaligned:
stq_u t0, 0(a0) # L : (stall)
ret (t9) # L0 : Latency=3
- nop
- nop
- nop
-
- .end __stxncpy
+ cfi_endproc
diff --git a/sysdeps/alpha/bzero.S b/sysdeps/alpha/bzero.S
index 091024b..a691ff6 100644
--- a/sysdeps/alpha/bzero.S
+++ b/sysdeps/alpha/bzero.S
@@ -36,23 +36,19 @@
.set noreorder
.text
+ .type __bzero, @function
+ .globl __bzero
+ .usepv __bzero, USEPV_PROF
-/* There is a problem with either gdb (as of 4.16) or gas (as of 2.7) that
- doesn't like putting the entry point for a procedure somewhere in the
- middle of the procedure descriptor. Work around this by putting the main
- loop in its own procedure descriptor. */
+ cfi_startproc
/* On entry to this basic block:
t3 == loop counter
t4 == bytes in partial final word
a0 == possibly misaligned destination pointer */
- .ent bzero_loop
.align 3
bzero_loop:
- .frame sp, 0, ra, 0
- .prologue 0
-
beq t3, $tail #
blbc t3, 0f # skip single store if count even
@@ -75,16 +71,11 @@ $tail: bne t4, 1f # is there a tail to do?
stq_u t0, 0(a0) #
ret #
- .end bzero_loop
-
-ENTRY(__bzero)
+__bzero:
#ifdef PROF
ldgp gp, 0(pv)
lda AT, _mcount
jsr AT, (AT), _mcount
- .prologue 1
-#else
- .prologue 0
#endif
mov a0, v0 # e0 : move return value in place
@@ -115,5 +106,5 @@ $oneq:
$done: ret
- END(__bzero)
+ cfi_endproc
weak_alias (__bzero, bzero)
diff --git a/sysdeps/alpha/memset.S b/sysdeps/alpha/memset.S
index 4ac7092..db92771 100644
--- a/sysdeps/alpha/memset.S
+++ b/sysdeps/alpha/memset.S
@@ -1,4 +1,4 @@
-/* Copyright (C) 1996, 1997, 2003 Free Software Foundation, Inc.
+/* Copyright (C) 1996-2012 Free Software Foundation, Inc.
Contributed by Richard Henderson (rth@tamu.edu)
This file is part of the GNU C Library.
@@ -37,11 +37,11 @@
.set noreorder
.text
+ .type memset, @function
+ .globl memset
+ .usepv memset, USEPV_PROF
-/* There is a problem with either gdb (as of 4.16) or gas (as of 2.7) that
- doesn't like putting the entry point for a procedure somewhere in the
- middle of the procedure descriptor. Work around this by putting the main
- loop in its own procedure descriptor. */
+ cfi_startproc
/* On entry to this basic block:
t3 == loop counter
@@ -49,12 +49,8 @@
a0 == possibly misaligned destination pointer
a1 == replicated source character */
- .ent memset_loop
.align 3
memset_loop:
- .frame sp, 0, ra, 0
- .prologue 0
-
beq t3, $tail
blbc t3, 0f # skip single store if count even
@@ -80,19 +76,14 @@ $tail: bne t4, 1f # is there a tail to do?
stq_u t0, 0(a0) # e0 :
ret # .. e1 :
- .end memset_loop
-
-ENTRY(memset)
+memset:
#ifdef PROF
ldgp gp, 0(pv)
lda AT, _mcount
jsr AT, (AT), _mcount
- .prologue 1
-#else
- .prologue 0
#endif
- zapnot a1, 1, a1 # e0 : zero extend input character
+ and a1, 0xff, a1 # e0 : zero extend input character
mov a0, v0 # .. e1 : move return value in place
sll a1, 8, t0 # e0 : begin replicating the char
beq a2, $done # .. e1 : early exit for zero-length store
@@ -132,5 +123,5 @@ $oneq:
$done: ret
- END(memset)
+ cfi_endproc
libc_hidden_builtin_def (memset)
diff --git a/sysdeps/alpha/stxcpy.S b/sysdeps/alpha/stxcpy.S
index 46f19a9..c55d005 100644
--- a/sysdeps/alpha/stxcpy.S
+++ b/sysdeps/alpha/stxcpy.S
@@ -1,4 +1,4 @@
-/* Copyright (C) 1996, 1997 Free Software Foundation, Inc.
+/* Copyright (C) 1996-2012 Free Software Foundation, Inc.
Contributed by Richard Henderson (rth@tamu.edu)
This file is part of the GNU C Library.
@@ -43,22 +43,18 @@
.set noreorder
.text
+ .type __stxcpy, @function
+ .globl __stxcpy
+ .usepv __stxcpy, no
-/* There is a problem with either gdb (as of 4.16) or gas (as of 2.7) that
- doesn't like putting the entry point for a procedure somewhere in the
- middle of the procedure descriptor. Work around this by putting the
- aligned copy in its own procedure descriptor */
-
- .ent stxcpy_aligned
- .align 3
-stxcpy_aligned:
- .frame sp, 0, t9
- .prologue 0
+ cfi_startproc
+ cfi_return_column (t9)
/* On entry to this basic block:
t0 == the first destination word for masking back in
t1 == the first source word. */
-
+ .align 3
+stxcpy_aligned:
/* Create the 1st output word and detect 0's in the 1st input word. */
lda t2, -1 # e1 : build a mask against false zero
mskqh t2, a1, t2 # e0 : detection in the src word
@@ -72,7 +68,6 @@ stxcpy_aligned:
/* On entry to this basic block:
t0 == the first destination word for masking back in
t1 == a source word not containing a null. */
-
$a_loop:
stq_u t1, 0(a0) # e0 :
addq a0, 8, a0 # .. e1 :
@@ -106,15 +101,8 @@ $a_eos:
1: stq_u t1, 0(a0) # e0 :
ret (t9) # .. e1 :
- .end stxcpy_aligned
-
.align 3
- .ent __stxcpy
- .globl __stxcpy
__stxcpy:
- .frame sp, 0, t9
- .prologue 0
-
/* Are source and destination co-aligned? */
xor a0, a1, t0 # e0 :
unop # :
@@ -303,4 +291,4 @@ $unaligned:
stq_u t1, 0(a0) # .. e0 :
ret (t9)
- .end __stxcpy
+ cfi_endproc
diff --git a/sysdeps/alpha/stxncpy.S b/sysdeps/alpha/stxncpy.S
index 5b81ac9..f8b494a 100644
--- a/sysdeps/alpha/stxncpy.S
+++ b/sysdeps/alpha/stxncpy.S
@@ -1,4 +1,4 @@
-/* Copyright (C) 1996, 1997, 2002 Free Software Foundation, Inc.
+/* Copyright (C) 1996-2012 Free Software Foundation, Inc.
Contributed by Richard Henderson (rth@tamu.edu)
This file is part of the GNU C Library.
@@ -52,22 +52,18 @@
.set noreorder
.text
+ .type __stxncpy, @function
+ .globl __stxncpy
+ .usepv __stxncpy, no
-/* There is a problem with either gdb (as of 4.16) or gas (as of 2.7) that
- doesn't like putting the entry point for a procedure somewhere in the
- middle of the procedure descriptor. Work around this by putting the
- aligned copy in its own procedure descriptor */
-
- .ent stxncpy_aligned
- .align 3
-stxncpy_aligned:
- .frame sp, 0, t9, 0
- .prologue 0
+ cfi_startproc
+ cfi_return_column (t9)
/* On entry to this basic block:
t0 == the first destination word for masking back in
t1 == the first source word. */
-
+ .align 3
+stxncpy_aligned:
/* Create the 1st output word and detect 0's in the 1st input word. */
lda t2, -1 # e1 : build a mask against false zero
mskqh t2, a1, t2 # e0 : detection in the src word
@@ -81,7 +77,6 @@ stxncpy_aligned:
/* On entry to this basic block:
t0 == a source word not containing a null. */
-
$a_loop:
stq_u t0, 0(a0) # e0 :
addq a0, 8, a0 # .. e1 :
@@ -98,7 +93,6 @@ $a_loop:
On entry to this basic block we have:
t0 == the source word containing the null
t7 == the cmpbge mask that found it. */
-
$a_eos:
negq t7, t8 # e0 : find low bit set
and t7, t8, t8 # e1 (stall)
@@ -126,15 +120,8 @@ $a_eoc:
or t10, t7, t7
br $a_eos
- .end stxncpy_aligned
-
.align 3
- .ent __stxncpy
- .globl __stxncpy
__stxncpy:
- .frame sp, 0, t9, 0
- .prologue 0
-
/* Are source and destination co-aligned? */
xor a0, a1, t1 # e0 :
and a0, 7, t0 # .. e1 : find dest misalignment
@@ -359,4 +346,4 @@ $unaligned:
stq_u t0, 0(a0) # e0 :
ret (t9) # .. e1 :
- .end __stxncpy
+ cfi_endproc
diff --git a/sysdeps/unix/alpha/sysdep.h b/sysdeps/unix/alpha/sysdep.h
index e17bf21..4ee0746 100644
--- a/sysdeps/unix/alpha/sysdep.h
+++ b/sysdeps/unix/alpha/sysdep.h
@@ -71,6 +71,12 @@
.prologue 1
#endif /* PROF */
+#ifdef PROF
+# define USEPV_PROF std
+#else
+# define USEPV_PROF no
+#endif
+
#if RTLD_PRIVATE_ERRNO
# define SYSCALL_ERROR_LABEL $syscall_error
# define SYSCALL_ERROR_HANDLER \
--
1.7.7.6
^ permalink raw reply [flat|nested] 5+ messages in thread
* [PATCH 3/5] alpha: Fix ev4 build with ev6 compiler
2012-06-06 21:52 [PATCH 1/5] alpha: put mcount prologue at correct location Richard Henderson
@ 2012-06-06 21:52 ` Richard Henderson
2012-06-06 21:52 ` [PATCH 5/5] alpha: Fix end-of-count checks in strncmp Richard Henderson
` (2 subsequent siblings)
3 siblings, 0 replies; 5+ messages in thread
From: Richard Henderson @ 2012-06-06 21:52 UTC (permalink / raw)
To: libc-ports
---
ChangeLog.alpha | 3 +++
sysdeps/alpha/fpu/e_sqrt.c | 3 ++-
2 files changed, 5 insertions(+), 1 deletions(-)
diff --git a/ChangeLog.alpha b/ChangeLog.alpha
index ceee910..0598d7e 100644
--- a/ChangeLog.alpha
+++ b/ChangeLog.alpha
@@ -1,5 +1,8 @@
2012-06-06 Richard Henderson <rth@twiddle.net>
+ * sysdeps/alpha/fpu/e_sqrt.c: Include <math_private.h> before
+ redefining __ieee758_sqrt.
+
* sysdeps/alpha/alphaev6/stxcpy.S: Use cfi markup instead of
dual ecoff procedure descriptors.
* sysdeps/alpha/alphaev6/stxncpy.S: Likewise.
diff --git a/sysdeps/alpha/fpu/e_sqrt.c b/sysdeps/alpha/fpu/e_sqrt.c
index 22b24b5..ad10dad 100644
--- a/sysdeps/alpha/fpu/e_sqrt.c
+++ b/sysdeps/alpha/fpu/e_sqrt.c
@@ -16,7 +16,8 @@
License along with the GNU C Library. If not, see
<http://www.gnu.org/licenses/>. */
-#include <features.h>
+#include <math.h>
+#include <math_private.h>
#if !defined(_IEEE_FP_INEXACT)
--
1.7.7.6
^ permalink raw reply [flat|nested] 5+ messages in thread
* [PATCH 5/5] alpha: Fix end-of-count checks in strncmp
2012-06-06 21:52 [PATCH 1/5] alpha: put mcount prologue at correct location Richard Henderson
2012-06-06 21:52 ` [PATCH 3/5] alpha: Fix ev4 build with ev6 compiler Richard Henderson
@ 2012-06-06 21:52 ` Richard Henderson
2012-06-06 21:52 ` [PATCH 4/5] alpha: Fix [BZ #13718] Richard Henderson
2012-06-06 21:52 ` [PATCH 2/5] alpha: Use cfi_startproc instead of dual procedure descriptors Richard Henderson
3 siblings, 0 replies; 5+ messages in thread
From: Richard Henderson @ 2012-06-06 21:52 UTC (permalink / raw)
To: libc-ports
This routine suffered a similar problem as stxncpy in needing to
bias a potentially very large unsigned number with wraparound.
This exposed a secondary problem where we didn't properly handle
end-of-count condition for the second string input resulting in
stratcliff failures.
---
ChangeLog.alpha | 3 +
sysdeps/alpha/strncmp.S | 251 ++++++++++++++++++++++++++---------------------
2 files changed, 144 insertions(+), 110 deletions(-)
diff --git a/ChangeLog.alpha b/ChangeLog.alpha
index d291df9..824083c 100644
--- a/ChangeLog.alpha
+++ b/ChangeLog.alpha
@@ -1,5 +1,8 @@
2012-06-06 Richard Henderson <rth@twiddle.net>
+ * sysdeps/alpha/strncmp.S: Bound count to LONG_MAX at startup.
+ Re-organize checks vs s2 end-of-count.
+
[BZ #13718]
* sysdeps/alpha/stxncmp.S: Bound count to LONG_MAX at startup.
* sysdeps/alpha/alphaev6/stxncmp.S: Likewise.
diff --git a/sysdeps/alpha/strncmp.S b/sysdeps/alpha/strncmp.S
index c9981e1..828f1b9 100644
--- a/sysdeps/alpha/strncmp.S
+++ b/sysdeps/alpha/strncmp.S
@@ -1,4 +1,4 @@
-/* Copyright (C) 1996, 1997, 2003 Free Software Foundation, Inc.
+/* Copyright (C) 1996-2012 Free Software Foundation, Inc.
Contributed by Richard Henderson (rth@tamu.edu)
This file is part of the GNU C Library.
@@ -23,6 +23,15 @@
.set noat
.set noreorder
+/* EV6 only predicts one branch per octaword. We'll use these to push
+ subsequent branches back to the next bundle. This will generally add
+ a fetch+decode cycle to older machines, so skip in that case. */
+#ifdef __alpha_fix__
+# define ev6_unop unop
+#else
+# define ev6_unop
+#endif
+
.text
ENTRY(strncmp)
@@ -35,128 +44,140 @@ ENTRY(strncmp)
.prologue 0
#endif
- xor a0, a1, t2 # e0 : are s1 and s2 co-aligned?
- beq a2, $zerolength # .. e1 :
- ldq_u t0, 0(a0) # e0 : give cache time to catch up
- ldq_u t1, 0(a1) # .. e1 :
- and t2, 7, t2 # e0 :
- and a0, 7, t4 # .. e1 : find s1 misalignment
- lda t3, -1 # e0 :
- addq a2, t4, a2 # .. e1 : bias count by s1 misalignment
- and a2, 7, t10 # e1 : ofs of last byte in last word
- srl a2, 3, a2 # .. e0 : remaining full words in count
- and a1, 7, t5 # e0 : find s2 misalignment
- bne t2, $unaligned # .. e1 :
+ xor a0, a1, t2 # are s1 and s2 co-aligned?
+ beq a2, $zerolength
+ ldq_u t0, 0(a0) # load asap to give cache time to catch up
+ ldq_u t1, 0(a1)
+ lda t3, -1
+ and t2, 7, t2
+ srl t3, 1, t6
+ and a0, 7, t4 # find s1 misalignment
+ and a1, 7, t5 # find s2 misalignment
+ cmovlt a2, t6, a2 # bound neg count to LONG_MAX
+ addq a1, a2, a3 # s2+count
+ addq a2, t4, a2 # bias count by s1 misalignment
+ and a2, 7, t10 # ofs of last byte in s1 last word
+ srl a2, 3, a2 # remaining full words in s1 count
+ bne t2, $unaligned
/* On entry to this basic block:
t0 == the first word of s1.
t1 == the first word of s2.
t3 == -1. */
-
$aligned:
- mskqh t3, a1, t3 # e0 : mask off leading garbage
- nop # .. e1 :
- ornot t1, t3, t1 # e0 :
- ornot t0, t3, t0 # .. e1 :
- cmpbge zero, t1, t7 # e0 : bits set iff null found
- beq a2, $eoc # .. e1 : check end of count
- unop # e0 :
- bne t7, $eos # .. e1 :
- unop # e0 :
- beq t10, $ant_loop # .. e1 :
+ mskqh t3, a1, t8 # mask off leading garbage
+ ornot t1, t8, t1
+ ornot t0, t8, t0
+ cmpbge zero, t1, t7 # bits set iff null found
+ beq a2, $eoc # check end of count
+ bne t7, $eos
+ beq t10, $ant_loop
/* Aligned compare main loop.
On entry to this basic block:
t0 == an s1 word.
t1 == an s2 word not containing a null. */
+ .align 4
$a_loop:
xor t0, t1, t2 # e0 :
bne t2, $wordcmp # .. e1 (zdb)
ldq_u t1, 8(a1) # e0 :
ldq_u t0, 8(a0) # .. e1 :
+
subq a2, 1, a2 # e0 :
addq a1, 8, a1 # .. e1 :
addq a0, 8, a0 # e0 :
beq a2, $eoc # .. e1 :
+
cmpbge zero, t1, t7 # e0 :
beq t7, $a_loop # .. e1 :
- unop # e0 :
- br $eos # .. e1 :
+
+ br $eos
/* Alternate aligned compare loop, for when there's no trailing
bytes on the count. We have to avoid reading too much data. */
+ .align 4
$ant_loop:
xor t0, t1, t2 # e0 :
+ ev6_unop
+ ev6_unop
bne t2, $wordcmp # .. e1 (zdb)
+
subq a2, 1, a2 # e0 :
beq a2, $zerolength # .. e1 :
ldq_u t1, 8(a1) # e0 :
ldq_u t0, 8(a0) # .. e1 :
+
addq a1, 8, a1 # e0 :
addq a0, 8, a0 # .. e1 :
cmpbge zero, t1, t7 # e0 :
beq t7, $ant_loop # .. e1 :
- unop # e0 :
- br $eos # .. e1 :
+
+ br $eos
/* The two strings are not co-aligned. Align s1 and cope. */
+ /* On entry to this basic block:
+ t0 == the first word of s1.
+ t1 == the first word of s2.
+ t3 == -1.
+ t4 == misalignment of s1.
+ t5 == misalignment of s2.
+ t10 == misalignment of s1 end. */
+ .align 4
$unaligned:
- subq a1, t4, a1 # e0 :
- unop # :
-
- /* If s2 misalignment is larger than s2 misalignment, we need
+ /* If s1 misalignment is larger than s2 misalignment, we need
extra startup checks to avoid SEGV. */
+ subq a1, t4, a1 # adjust s2 for s1 misalignment
+ cmpult t4, t5, t9
+ subq a3, 1, a3 # last byte of s2
+ bic a1, 7, t8
+ mskqh t3, t5, t7 # mask garbage in s2
+ subq a3, t8, a3
+ ornot t1, t7, t7
+ srl a3, 3, a3 # remaining full words in s2 count
+ beq t9, $u_head
+
+ /* Failing that, we need to look for both eos and eoc within the
+ first word of s2. If we find either, we can continue by
+ pretending that the next word of s2 is all zeros. */
+ lda t2, 0 # next = zero
+ cmpeq a3, 0, t8 # eoc in the first word of s2?
+ cmpbge zero, t7, t7 # eos in the first word of s2?
+ or t7, t8, t8
+ bne t8, $u_head_nl
- cmplt t4, t5, t8 # .. e1 :
- beq t8, $u_head # e1 :
-
- mskqh t3, t5, t3 # e0 :
- ornot t1, t3, t3 # e0 :
- cmpbge zero, t3, t7 # e1 : is there a zero?
- beq t7, $u_head # e1 :
-
- /* We've found a zero in the first partial word of s2. Align
- our current s1 and s2 words and compare what we've got. */
-
- extql t1, t5, t1 # e0 :
- lda t3, -1 # .. e1 :
- insql t1, a0, t1 # e0 :
- mskqh t3, a0, t3 # e0 :
- ornot t1, t3, t1 # e0 :
- ornot t0, t3, t0 # .. e1 :
- cmpbge zero, t1, t7 # e0 : find that zero again
- beq a2, $eoc # .. e1 : and finish up
- br $eos # e1 :
-
- .align 3
-$u_head:
/* We know just enough now to be able to assemble the first
full word of s2. We can still find a zero at the end of it.
On entry to this basic block:
t0 == first word of s1
- t1 == first partial word of s2. */
-
- ldq_u t2, 8(a1) # e0 : load second partial s2 word
- lda t3, -1 # .. e1 : create leading garbage mask
- extql t1, a1, t1 # e0 : create first s2 word
- mskqh t3, a0, t3 # e0 :
- extqh t2, a1, t4 # e0 :
- ornot t0, t3, t0 # .. e1 : kill s1 garbage
- or t1, t4, t1 # e0 : s2 word now complete
- ornot t1, t3, t1 # e1 : kill s2 garbage
- cmpbge zero, t0, t7 # e0 : find zero in first s1 word
- beq a2, $eoc # .. e1 :
- lda t3, -1 # e0 :
- bne t7, $eos # .. e1 :
- subq a2, 1, a2 # e0 :
- xor t0, t1, t4 # .. e1 : compare aligned words
- mskql t3, a1, t3 # e0 : mask out s2[1] bits we have seen
- bne t4, $wordcmp # .. e1 :
- or t2, t3, t3 # e0 :
- cmpbge zero, t3, t7 # e1 : find zero in high bits of s2[1]
- bne t7, $u_final # e1 :
+ t1 == first partial word of s2.
+ t3 == -1.
+ t10 == ofs of last byte in s1 last word.
+ t11 == ofs of last byte in s2 last word. */
+$u_head:
+ ldq_u t2, 8(a1) # load second partial s2 word
+ subq a3, 1, a3
+$u_head_nl:
+ extql t1, a1, t1 # create first s2 word
+ mskqh t3, a0, t8
+ extqh t2, a1, t4
+ ornot t0, t8, t0 # kill s1 garbage
+ or t1, t4, t1 # s2 word now complete
+ cmpbge zero, t0, t7 # find eos in first s1 word
+ ornot t1, t8, t1 # kill s2 garbage
+ beq a2, $eoc
+ subq a2, 1, a2
+ bne t7, $eos
+ mskql t3, a1, t8 # mask out s2[1] bits we have seen
+ xor t0, t1, t4 # compare aligned words
+ or t2, t8, t8
+ bne t4, $wordcmp
+ cmpbge zero, t8, t7 # eos in high bits of s2[1]?
+ cmpeq a3, 0, t8 # eoc in s2[1]?
+ or t7, t8, t7
+ bne t7, $u_final
/* Unaligned copy main loop. In order to avoid reading too much,
the loop is structured to detect zeros in aligned words from s2.
@@ -166,43 +187,54 @@ $u_head:
to run as fast as possible.
On entry to this basic block:
- t2 == the unshifted low-bits from the next s2 word. */
-
- .align 3
+ t2 == the unshifted low-bits from the next s2 word.
+ t10 == ofs of last byte in s1 last word.
+ t11 == ofs of last byte in s2 last word. */
+ .align 4
$u_loop:
extql t2, a1, t3 # e0 :
ldq_u t2, 16(a1) # .. e1 : load next s2 high bits
ldq_u t0, 8(a0) # e0 : load next s1 word
addq a1, 8, a1 # .. e1 :
+
addq a0, 8, a0 # e0 :
- nop # .. e1 :
+ subq a3, 1, a3 # .. e1 :
extqh t2, a1, t1 # e0 :
- cmpbge zero, t0, t7 # .. e1 : find zero in current s1 word
+ cmpbge zero, t0, t7 # .. e1 : eos in current s1 word
+
or t1, t3, t1 # e0 :
- beq a2, $eoc # .. e1 : check for end of count
+ beq a2, $eoc # .. e1 : eoc in current s1 word
subq a2, 1, a2 # e0 :
+ cmpbge zero, t2, t4 # .. e1 : eos in s2[1]
+
+ xor t0, t1, t3 # e0 : compare the words
+ ev6_unop
+ ev6_unop
bne t7, $eos # .. e1 :
- xor t0, t1, t4 # e0 : compare the words
- bne t4, $wordcmp # .. e1 (zdb)
- cmpbge zero, t2, t4 # e0 : find zero in next low bits
+
+ cmpeq a3, 0, t5 # e0 : eoc in s2[1]
+ ev6_unop
+ ev6_unop
+ bne t3, $wordcmp # .. e1 :
+
+ or t4, t5, t4 # e0 : eos or eoc in s2[1].
beq t4, $u_loop # .. e1 (zdb)
/* We've found a zero in the low bits of the last s2 word. Get
the next s1 word and align them. */
+ .align 3
$u_final:
- ldq_u t0, 8(a0) # e1 :
- extql t2, a1, t1 # .. e0 :
- cmpbge zero, t1, t7 # e0 :
- bne a2, $eos # .. e1 :
+ ldq_u t0, 8(a0)
+ extql t2, a1, t1
+ cmpbge zero, t1, t7
+ bne a2, $eos
/* We've hit end of count. Zero everything after the count
and compare whats left. */
-
.align 3
$eoc:
mskql t0, t10, t0
mskql t1, t10, t1
- unop
cmpbge zero, t1, t7
/* We've found a zero somewhere in a word we just read.
@@ -210,32 +242,31 @@ $eoc:
t0 == s1 word
t1 == s2 word
t7 == cmpbge mask containing the zero. */
-
+ .align 3
$eos:
- negq t7, t6 # e0 : create bytemask of valid data
- and t6, t7, t8 # e1 :
- subq t8, 1, t6 # e0 :
- or t6, t8, t7 # e1 :
- zapnot t0, t7, t0 # e0 : kill the garbage
- zapnot t1, t7, t1 # .. e1 :
- xor t0, t1, v0 # e0 : and compare
- beq v0, $done # .. e1 :
+ negq t7, t6 # create bytemask of valid data
+ and t6, t7, t8
+ subq t8, 1, t6
+ or t6, t8, t7
+ zapnot t0, t7, t0 # kill the garbage
+ zapnot t1, t7, t1
+ xor t0, t1, v0 # ... and compare
+ beq v0, $done
/* Here we have two differing co-aligned words in t0 & t1.
Bytewise compare them and return (t0 > t1 ? 1 : -1). */
.align 3
$wordcmp:
- cmpbge t0, t1, t2 # e0 : comparison yields bit mask of ge
- cmpbge t1, t0, t3 # .. e1 :
- xor t2, t3, t0 # e0 : bits set iff t0/t1 bytes differ
- negq t0, t1 # e1 : clear all but least bit
- and t0, t1, t0 # e0 :
- lda v0, -1 # .. e1 :
- and t0, t2, t1 # e0 : was bit set in t0 > t1?
- cmovne t1, 1, v0 # .. e1 (zdb)
-
+ cmpbge t0, t1, t2 # comparison yields bit mask of ge
+ cmpbge t1, t0, t3
+ xor t2, t3, t0 # bits set iff t0/t1 bytes differ
+ negq t0, t1 # clear all but least bit
+ and t0, t1, t0
+ lda v0, -1
+ and t0, t2, t1 # was bit set in t0 > t1?
+ cmovne t1, 1, v0
$done:
- ret # e1 :
+ ret
.align 3
$zerolength:
--
1.7.7.6
^ permalink raw reply [flat|nested] 5+ messages in thread
* [PATCH 4/5] alpha: Fix [BZ #13718]
2012-06-06 21:52 [PATCH 1/5] alpha: put mcount prologue at correct location Richard Henderson
2012-06-06 21:52 ` [PATCH 3/5] alpha: Fix ev4 build with ev6 compiler Richard Henderson
2012-06-06 21:52 ` [PATCH 5/5] alpha: Fix end-of-count checks in strncmp Richard Henderson
@ 2012-06-06 21:52 ` Richard Henderson
2012-06-06 21:52 ` [PATCH 2/5] alpha: Use cfi_startproc instead of dual procedure descriptors Richard Henderson
3 siblings, 0 replies; 5+ messages in thread
From: Richard Henderson @ 2012-06-06 21:52 UTC (permalink / raw)
To: libc-ports
The routines expect to be able to bias the count by a small number.
If the count is near -1ull, the count will overflow. Since we cannot
use the whole 64-bit address space, bound the count to LONG_MAX.
---
ChangeLog.alpha | 4 ++++
sysdeps/alpha/alphaev6/stxncpy.S | 19 +++++++++++++------
sysdeps/alpha/stxncpy.S | 23 +++++++++++++----------
3 files changed, 30 insertions(+), 16 deletions(-)
diff --git a/ChangeLog.alpha b/ChangeLog.alpha
index 0598d7e..d291df9 100644
--- a/ChangeLog.alpha
+++ b/ChangeLog.alpha
@@ -1,5 +1,9 @@
2012-06-06 Richard Henderson <rth@twiddle.net>
+ [BZ #13718]
+ * sysdeps/alpha/stxncmp.S: Bound count to LONG_MAX at startup.
+ * sysdeps/alpha/alphaev6/stxncmp.S: Likewise.
+
* sysdeps/alpha/fpu/e_sqrt.c: Include <math_private.h> before
redefining __ieee758_sqrt.
diff --git a/sysdeps/alpha/alphaev6/stxncpy.S b/sysdeps/alpha/alphaev6/stxncpy.S
index d134eb8..28495df 100644
--- a/sysdeps/alpha/alphaev6/stxncpy.S
+++ b/sysdeps/alpha/alphaev6/stxncpy.S
@@ -143,18 +143,25 @@ $a_eoc:
.align 4
__stxncpy:
/* Are source and destination co-aligned? */
+ lda t2, -1 # E :
xor a0, a1, t1 # E :
and a0, 7, t0 # E : find dest misalignment
- and t1, 7, t1 # E : (stall)
- addq a2, t0, a2 # E : bias count by dest misalignment (stall)
+ nop # E :
- subq a2, 1, a2 # E :
+ srl t2, 1, t2 # U :
+ and t1, 7, t1 # E :
+ cmovlt a2, t2, a2 # E : bound count to LONG_MAX (stall)
+ nop # E :
+
+ addq a2, t0, a2 # E : bias count by dest misalignment
+ subq a2, 1, a2 # E : (stall)
and a2, 7, t2 # E : (stall)
- srl a2, 3, a2 # U : a2 = loop counter = (count - 1)/8 (stall)
- addq zero, 1, t10 # E :
+ lda t10, 1 # E :
+ srl a2, 3, a2 # U : a2 = loop counter = (count - 1)/8
sll t10, t2, t10 # U : t10 = bitmask of last count byte
- bne t1, $unaligned # U :
+ nop # E :
+ bne t1, $unaligned # U : (stall)
/* We are co-aligned; take care of a partial first word. */
ldq_u t1, 0(a1) # L : load first src word
diff --git a/sysdeps/alpha/stxncpy.S b/sysdeps/alpha/stxncpy.S
index f8b494a..d2cb9c3 100644
--- a/sysdeps/alpha/stxncpy.S
+++ b/sysdeps/alpha/stxncpy.S
@@ -123,16 +123,19 @@ $a_eoc:
.align 3
__stxncpy:
/* Are source and destination co-aligned? */
- xor a0, a1, t1 # e0 :
- and a0, 7, t0 # .. e1 : find dest misalignment
- and t1, 7, t1 # e0 :
- addq a2, t0, a2 # .. e1 : bias count by dest misalignment
- subq a2, 1, a2 # e0 :
- and a2, 7, t2 # e1 :
- srl a2, 3, a2 # e0 : a2 = loop counter = (count - 1)/8
- addq zero, 1, t10 # .. e1 :
- sll t10, t2, t10 # e0 : t10 = bitmask of last count byte
- bne t1, $unaligned # .. e1 :
+ lda t2, -1
+ xor a0, a1, t1
+ srl t2, 1, t2
+ and a0, 7, t0 # find dest misalignment
+ cmovlt a2, t2, a2 # bound neg count to LONG_MAX
+ and t1, 7, t1
+ addq a2, t0, a2 # bias count by dest misalignment
+ subq a2, 1, a2
+ and a2, 7, t2
+ srl a2, 3, a2 # a2 = loop counter = (count - 1)/8
+ addq zero, 1, t10
+ sll t10, t2, t10 # t10 = bitmask of last count byte
+ bne t1, $unaligned
/* We are co-aligned; take care of a partial first word. */
--
1.7.7.6
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2012-06-06 21:52 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2012-06-06 21:52 [PATCH 1/5] alpha: put mcount prologue at correct location Richard Henderson
2012-06-06 21:52 ` [PATCH 3/5] alpha: Fix ev4 build with ev6 compiler Richard Henderson
2012-06-06 21:52 ` [PATCH 5/5] alpha: Fix end-of-count checks in strncmp Richard Henderson
2012-06-06 21:52 ` [PATCH 4/5] alpha: Fix [BZ #13718] Richard Henderson
2012-06-06 21:52 ` [PATCH 2/5] alpha: Use cfi_startproc instead of dual procedure descriptors Richard Henderson
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).