mirror of
https://gitlab.com/freepascal.org/fpc/source.git
synced 2025-04-14 18:19:54 +02:00
* Did a clean rewrite of IndexByte and IndexWord, resulting in somewhat less instructions.
+ IndexWord for the case of unaligned buffer: implemented using aligned reads. + tindex.pp: Added testing correctness of IndexWord with unlimited length. git-svn-id: trunk@17317 -
This commit is contained in:
sergei
parent
9781c0d051
commit
30f7bff09d
@ -459,7 +459,6 @@ Procedure FillChar(var x;count:SizeInt;value:byte);assembler;nostackframe;
|
||||
{$endif FPC_SYSTEM_HAS_FILLCHAR}
|
||||
|
||||
{$ifndef FPC_SYSTEM_HAS_INDEXBYTE}
|
||||
{ based on libc/sysdeps/x86_64/memchr.S }
|
||||
{$define FPC_SYSTEM_HAS_INDEXBYTE}
|
||||
function IndexByte(Const buf;len:SizeInt;b:byte):SizeInt; assembler; nostackframe;
|
||||
{ win64: rcx buf, rdx len, r8b word
|
||||
@ -472,51 +471,45 @@ asm
|
||||
movq %rdi, %rcx
|
||||
movq %rsi, %rdx
|
||||
{$endif}
|
||||
mov %rcx, %rax { duplicate buf }
|
||||
mov %rcx, %r8
|
||||
punpcklbw %xmm1, %xmm1
|
||||
and $0xfffffffffffffff0, %rax
|
||||
and $-0x10, %rcx { highest aligned address before buf }
|
||||
test %rdx, %rdx
|
||||
punpcklbw %xmm1, %xmm1
|
||||
jz .L3 { exit if len=0 }
|
||||
orl $0xffffffff, %r8d
|
||||
movdqa (%rax), %xmm0 { Fetch first 16 bytes (up to 15 bytes before target) }
|
||||
jz .Lnotfound { exit if len=0 }
|
||||
add $16, %rcx { first aligned address after buf }
|
||||
pshufd $0, %xmm1, %xmm1
|
||||
sub %rax, %rcx { rcx=misalignment }
|
||||
pcmpeqb %xmm1, %xmm0
|
||||
add %rcx, %rdx { add misalignment to length }
|
||||
cmovb %r8, %rdx { if it overflows (happens when length=-1), set back to -1, }
|
||||
{ otherwise loop will terminate too early }
|
||||
mov %rcx, %r9 { and save it, will subtract back in the end }
|
||||
shl %cl, %r8d
|
||||
pmovmskb %xmm0, %ecx
|
||||
andl %r8d, %ecx { mask away matches before buffer start }
|
||||
movl $16, %r8d
|
||||
jnz .L1 { got a match within buffer -> we're done (almost) }
|
||||
cmpq %r8, %rdx
|
||||
jbe .L3
|
||||
movdqa -16(%rcx), %xmm0 { Fetch first 16 bytes (up to 15 bytes before target) }
|
||||
sub %r8, %rcx { rcx=number of valid bytes, r8=original ptr }
|
||||
|
||||
pcmpeqb %xmm1, %xmm0 { compare with pattern and get bitmask }
|
||||
pmovmskb %xmm0, %eax
|
||||
|
||||
shl %cl, %eax { shift valid bits into high word }
|
||||
and $0xffff0000, %eax { clear low word containing invalid bits }
|
||||
shr %cl, %eax { shift back }
|
||||
jmp .Lcontinue
|
||||
|
||||
.balign 16
|
||||
.L2:
|
||||
movdqa (%rax,%r8), %xmm0
|
||||
lea 16(%r8), %r8
|
||||
.Lloop:
|
||||
movdqa (%r8,%rcx), %xmm0 { r8 and rcx may have any values, }
|
||||
add $16, %rcx { but their sum is evenly divisible by 16. }
|
||||
pcmpeqb %xmm1, %xmm0
|
||||
pmovmskb %xmm0, %ecx
|
||||
test %ecx, %ecx
|
||||
jnz .L1
|
||||
cmp %r8, %rdx
|
||||
ja .L2
|
||||
pmovmskb %xmm0, %eax
|
||||
.Lcontinue:
|
||||
test %eax, %eax
|
||||
jnz .Lmatch
|
||||
cmp %rcx, %rdx
|
||||
ja .Lloop
|
||||
.Lnotfound:
|
||||
or $-1, %rax
|
||||
retq
|
||||
|
||||
.L3:
|
||||
or $-1, %rax
|
||||
jmp .Ldone
|
||||
|
||||
.L1:
|
||||
bsfl %ecx, %ecx { compute position of the first match }
|
||||
lea -16(%rcx,%r8), %rax
|
||||
cmp %rax, %rdx
|
||||
jbe .L3 { if it is after the specified length, ignore it }
|
||||
sub %r9, %rax
|
||||
.Ldone:
|
||||
.Lmatch:
|
||||
bsf %eax, %eax
|
||||
lea -16(%rcx,%rax), %rax
|
||||
cmp %rax, %rdx { check against the buffer length }
|
||||
jbe .Lnotfound
|
||||
end;
|
||||
{$endif FPC_SYSTEM_HAS_INDEXBYTE}
|
||||
|
||||
@ -533,77 +526,97 @@ asm
|
||||
movq %rdi, %rcx
|
||||
movq %rsi, %rdx
|
||||
{$endif}
|
||||
mov %rcx, %rax { duplicate buf }
|
||||
mov %rcx, %r8
|
||||
punpcklwd %xmm1, %xmm1
|
||||
and $0xfffffffffffffff0, %rax
|
||||
and $-0x10, %rcx
|
||||
test %rdx, %rdx
|
||||
pshufd $0, %xmm1, %xmm1
|
||||
jz .L3 { exit if len=0 }
|
||||
orl $0xffffffff, %r8d
|
||||
test $1, %cl { if buffer isn't aligned to word boundary, }
|
||||
jnz .Lunaligned { fallback to slower unaligned loop }
|
||||
jz .Lnotfound { exit if len=0 }
|
||||
add $16, %rcx
|
||||
movdqa -16(%rcx), %xmm0 { Fetch first 16 bytes (up to 14 bytes before target) }
|
||||
sub %r8, %rcx { rcx=number of valid bytes }
|
||||
|
||||
movdqa (%rax), %xmm0 { Fetch first 16 bytes (up to 14 bytes before target) }
|
||||
sub %rax, %rcx { rcx=misalignment }
|
||||
pcmpeqw %xmm1, %xmm0
|
||||
test $1, %r8b { if buffer isn't aligned to word boundary, }
|
||||
jnz .Lunaligned { use a different algorithm }
|
||||
|
||||
mov %rcx, %r9
|
||||
shr $1, %r9 { save misalignment in words }
|
||||
pcmpeqw %xmm1, %xmm0
|
||||
pmovmskb %xmm0, %eax
|
||||
|
||||
add %r9, %rdx { add misalignment to length }
|
||||
cmovb %r8, %rdx { if it overflows (happens when length=-1), set back to -1, }
|
||||
{ otherwise loop will terminate too early }
|
||||
shl %cl, %r8d
|
||||
pmovmskb %xmm0, %ecx
|
||||
andl %r8d, %ecx { mask away matches before buffer start }
|
||||
movl $8, %r8d
|
||||
jnz .L1 { got a match within buffer -> we're done (almost) }
|
||||
cmpq %r8, %rdx
|
||||
jbe .L3
|
||||
shl %cl, %eax
|
||||
and $0xffff0000, %eax
|
||||
shr %cl, %eax
|
||||
shr $1, %ecx { bytes->words }
|
||||
jmp .Lcontinue
|
||||
|
||||
.balign 16
|
||||
.L2:
|
||||
movdqa (%rax,%r8,2), %xmm0
|
||||
lea 8(%r8), %r8
|
||||
pcmpeqw %xmm1, %xmm0
|
||||
pmovmskb %xmm0, %ecx
|
||||
test %ecx, %ecx
|
||||
jnz .L1
|
||||
cmp %r8, %rdx
|
||||
ja .L2
|
||||
.Lloop:
|
||||
movdqa (%r8,%rcx,2), %xmm0
|
||||
add $8, %rcx
|
||||
pcmpeqw %xmm1, %xmm0
|
||||
pmovmskb %xmm0, %eax
|
||||
.Lcontinue:
|
||||
test %eax, %eax
|
||||
jnz .Lmatch
|
||||
cmp %rcx, %rdx
|
||||
ja .Lloop
|
||||
|
||||
.L3:
|
||||
.Lnotfound:
|
||||
or $-1, %rax
|
||||
jmp .Ldone
|
||||
|
||||
.L1:
|
||||
bsfl %ecx, %ecx { compute position of the first match }
|
||||
shr $1, %ecx { in words }
|
||||
lea -8(%rcx,%r8), %rax
|
||||
cmp %rax, %rdx
|
||||
jbe .L3 { if it is after the specified length, ignore it }
|
||||
sub %r9, %rax
|
||||
.Ldone:
|
||||
retq
|
||||
|
||||
{ TODO: aligned processing is still possible, but for now
|
||||
use the simplest form }
|
||||
.Lmatch:
|
||||
bsf %eax, %eax
|
||||
shr $1, %eax { in words }
|
||||
lea -8(%rcx,%rax), %rax
|
||||
cmp %rax, %rdx
|
||||
jbe .Lnotfound { if match is after the specified length, ignore it }
|
||||
retq
|
||||
|
||||
.Lunaligned:
|
||||
xor %r9, %r9
|
||||
xor %r8, %r8
|
||||
mov %rcx, %rax
|
||||
movdqa %xmm1, %xmm2 { (mis)align the pattern (in this particular case: }
|
||||
psllw $8, %xmm1 { swap bytes of each word of pattern) }
|
||||
psrlw $8, %xmm2
|
||||
por %xmm2, %xmm1
|
||||
|
||||
pcmpeqb %xmm1, %xmm0
|
||||
pmovmskb %xmm0, %eax
|
||||
|
||||
shl %cl, %eax
|
||||
and $0xffff0000, %eax
|
||||
shr %cl, %eax
|
||||
|
||||
add %rdx, %rdx { length words -> bytes }
|
||||
xor %r10d, %r10d { nothing to merge yet }
|
||||
jmp .Lcontinue_u
|
||||
|
||||
.balign 16
|
||||
.L2u:
|
||||
movdqu (%rax,%r8,2), %xmm0
|
||||
lea 8(%r8), %r8
|
||||
pcmpeqw %xmm1, %xmm0
|
||||
pmovmskb %xmm0, %ecx
|
||||
test %ecx, %ecx
|
||||
jnz .L1
|
||||
cmp %r8, %rdx
|
||||
ja .L2u
|
||||
.Lloop_u:
|
||||
movdqa (%r8,%rcx), %xmm0
|
||||
add $16, %rcx
|
||||
pcmpeqb %xmm1, %xmm0 { compare by bytes }
|
||||
shr $16, %r10d { bit 16 shifts into 0 }
|
||||
pmovmskb %xmm0, %eax
|
||||
.Lcontinue_u:
|
||||
shl $1, %eax { 15:0 -> 16:1 }
|
||||
or %r10d, %eax { merge bit 0 from previous round }
|
||||
mov %eax, %r10d
|
||||
shr $1, %eax { now AND together adjacent pairs of bits }
|
||||
and %r10d, %eax
|
||||
and $0x5555, %eax { also reset odd bits }
|
||||
jnz .Lmatch_u
|
||||
cmpq %rcx, %rdx
|
||||
ja .Lloop_u
|
||||
|
||||
.Lnotfound_u:
|
||||
or $-1, %rax
|
||||
retq
|
||||
.Lmatch_u:
|
||||
bsf %eax, %eax
|
||||
lea -16(%rcx,%rax), %rax
|
||||
cmp %rax, %rdx
|
||||
jbe .Lnotfound_u { if match is after the specified length, ignore it }
|
||||
sar $1, %eax { in words }
|
||||
retq
|
||||
end;
|
||||
{$endif FPC_SYSTEM_HAS_INDEXWORD}
|
||||
|
||||
|
||||
@ -1,5 +1,3 @@
|
||||
const
|
||||
err: boolean = false;
|
||||
|
||||
var
|
||||
a, b: array[0..515] of byte;
|
||||
@ -128,6 +126,13 @@ begin
|
||||
writeln('indexword error 7 for (',i,',',j,',',k,')');
|
||||
halt(7);
|
||||
end;
|
||||
{same for length=-1}
|
||||
if indexword(b[k+4],-1,0)<>index then
|
||||
begin
|
||||
writeln(indexword(b[k+4],-1,0),' <> ',index);
|
||||
writeln('indexword error 7a for (',i,',',j,',',k,')');
|
||||
halt(27);
|
||||
end;
|
||||
|
||||
if (i=0) then
|
||||
index:=0
|
||||
@ -140,6 +145,13 @@ begin
|
||||
writeln('indexword error 8 for (',i,',',j,',',k,')');
|
||||
halt(8);
|
||||
end;
|
||||
{same for length=-1}
|
||||
if indexword(b[k+4],-1,l)<>index then
|
||||
begin
|
||||
writeln(indexword(b[k+4],-1,l),' <> ',index);
|
||||
writeln('indexword error 8a for (',i,',',j,',',k,')');
|
||||
halt(28);
|
||||
end;
|
||||
|
||||
l:=unaligned(pword(@(b[k+4+((i shr 2) and not 1)-2]))^);
|
||||
if (i>=8) then
|
||||
@ -152,6 +164,14 @@ begin
|
||||
writeln('indexword error 9 for (',i,',',j,',',k,')');
|
||||
halt(9);
|
||||
end;
|
||||
if (i>1) and (index<>-1) then
|
||||
if indexword(b[k+4],-1,l)<>index then
|
||||
begin
|
||||
writeln(indexword(b[k+4],-1,l),' <> ',index);
|
||||
writeln('indexword error 9a for (',i,',',j,',',k,')');
|
||||
halt(29);
|
||||
end;
|
||||
|
||||
l:=unaligned(pword(@(b[k+4]))^);
|
||||
if (i<2) then
|
||||
index:=-1
|
||||
@ -162,6 +182,12 @@ begin
|
||||
writeln('indexword error 10 for (',i,',',j,',',k,')');
|
||||
halt(10);
|
||||
end;
|
||||
if i>1 then
|
||||
if indexword(b[k+4],-1,l)<>index then
|
||||
begin
|
||||
writeln('indexword error 10a for (',i,',',j,',',k,')');
|
||||
halt(30);
|
||||
end;
|
||||
|
||||
|
||||
if (unaligned(pdword(@b[k+4])^)=0) then
|
||||
|
||||
Loading…
Reference in New Issue
Block a user