passes first tests! Only untested feature: large offset which forces

minimum match length to be >=3.
	arm_nrv2b_d32.S arm_nrv2e_d32.S

committer: jreiser <jreiser> 1140677145 +0000
This commit is contained in:
John Reiser
2006-02-23 06:45:45 +00:00
parent f3b8c9f34e
commit 20e4cbfd13
2 changed files with 113 additions and 100 deletions
+72 -58
View File
@@ -1,4 +1,4 @@
/* arm_nrv2b_d32.S -- ARM decompressor for NRV2E /* arm_nrv2b_d32.S -- ARM decompressor for NRV2B
This file is part of the UPX executable compressor. This file is part of the UPX executable compressor.
@@ -35,79 +35,93 @@
#define tmp r3 #define tmp r3
#define bits r4 #define bits r4
#define off r5 #define off r5
#define lr2 r6 #define g32 r6
#define g32 r7 #define wrnk r7 /* 0xd00 M2_MAX_OFFSET before "wrinkle" */
ucl_nrv2b_decompress_32: .globl ucl_nrv2b_decompress_32 /* r12 ("ip") is assumed to be a scratch register. */
@ ARM mode (char *src, int len_src, char *dst, int *plen_dst)
add r1,len,src @ r1= eof_src;
stmfd sp!,{r1,r2,r3, r4,r5,r6,r7,lr}
mov bits,#1<<31
blx hitch_n2b
hitch_n2b_r:
ldmfd sp!,{r4,r5,r6,r7,pc}
get32: @ ARM mode; In: Carry set (unchanged until final adc)
ldrb bits,[src],#1
ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #1*8
ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #2*8
ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #3*8
adc bits,bits,bits
bx lr
unhitch_n2b:
#define GETBIT \ #define GETBIT \
add bits,bits; beq 1f; 0: \ add bits,bits; beq 1f; 0: \
.subsection 1; \ .subsection 1; \
1: blx getb; b 0b; \ 1: blx g32; b 0b; \
.subsection 0 .subsection 0
#define getnextb(reg) GETBIT; adc reg,reg #define getnextb(reg) GETBIT; adc reg,reg
#define jnextb0 GETBIT; bcc #define jnextb0 GETBIT; bcc
#define jnextb1 GETBIT; bcs #define jnextb1 GETBIT; bcs
.code 16 @ THUMB mode ucl_nrv2b_decompress_32: .globl ucl_nrv2b_decompress_32 @ ARM mode
.type ucl_nrv2b_decompress_32, %function
/* error = (*)(char const *src, int len_src, char *dst, int *plen_dst) */
adr r12,1+thumb_nrv2b_d32; bx r12 @ enter THUMB mode
.code 16 @ THUMB mode
thumb_nrv2b_d32:
add r1,len,src @ r1= eof_src;
push {r1,r2,r3, r4,r5,r6,r7, lr}
mov bits,#1; neg off,bits @ off= -1 initial condition
lsl bits,#31 @ 1<<31: refill next time
mov wrnk,#0xd
lsl wrnk,#8 @ 0xd00
adr g32,get32 @ load pc-relative address
b top_n2b
eof_n2b: eof_n2b:
pop {r1,r3,r4} @ r1= eof_src; r3= orig_dst; r4= plen_dst pop {r1,r3,r4} @ r1= eof_src; r3= orig_dst; r4= plen_dst
sub src,r1 @ 0 if src length OK sub src,r1 @ 0 if actual src length equals expected length
sub dst,r3 @ actual dst length sub dst,r3 @ actual dst length
str dst,[r4] str dst,[r4]
sub g32,#get32 - hitch_n2b_r @ g32= &hitch_n2b_r pop {r4,r5,r6,r7, pc} @ return
bx g32
hitch_n2b:
mov g32,lr
add g32,#get32 - hitch_n2b_r @ g32= &get32
b top
lit_n2b: lit_n2b:
ldrb tmp,[src]; add src,#1 ldrb tmp,[src]; add src,#1
strb tmp,[dst]; add dst,#1 strb tmp,[dst]; add dst,#1
top_n2b: top_n2b:
jnextb1 lit_n2b jnextb1 lit_n2b
bl ss11 @ len= offset bl ss11 @ len= encoded offset [2..)
sub len,#3; bcs off_same sub tmp,len,#3 @ set Carry
ldrb off,[src]; add src,#1 @ low 8 bits mov len,#0 @ Carry unaffected
lsl len,#8 blo offprev_n2b @ ss11 returned 2
orr off,len lsl tmp,#8
mvn off,off; beq eof_n2b @ off= ~off ldrb off,[src]; add src,#1 @ low 8 bits
off_same: orr off,tmp
bl ss11 @ len= count -2 mvn off,off; beq eof_n2b @ off= ~off
mov tmp,#0xd; lsl tmp,#8 offprev_n2b: @ In: 0==len
cmn tmp,len getnextb(len); getnextb(len); bne plus1_n2b @ two bits; 1,2,3 ==> 2,3,4
mov tmp,#2 @ does not modify Carry bl ss11 @ len= encoded length [2..)
adc len,tmp @ len += 2+ (0xd00<=offset) add len,#2 @ [2..) ==> [4..);
ldrb tmp,[dst] @ force cacheline allocate plus1_n2b:
copy_n2v: add len,#1 @ 1,2,3 ==> 2,3,4; [4..) ==> [5..)
ldrb tmp,[dst,off] /* 'cmn': add the inputs, set condition codes, discard the sum */
strb tmp,[dst]; add dst,#1 cmn off,wrnk; bcs near_n2b @ within M2_MAX_OFFSET
sub len,#1; bne copy_n2b add len,#1 @ too far away, so minimum match length is 3
b top_n2b near_n2b:
ldrb tmp,[dst] @ force cacheline allocate
copy_n2b:
ldrb tmp,[dst,off]
strb tmp,[dst]; add dst,#1
sub len,#1; bne copy_n2b
b top_n2b
ss11: ss11:
mov len,#1 @ the msb mov len,#1 @ the msb
mov lr2,lr @ save return address mov r12,lr @ return address; alternate: "push {lr}"
ss11a: ss11a:
getnextb(len) getnextb(len)
jnextb0 ss11a jnextb0 ss11a
mov pc,lr2 @ return mov pc,r12 @ return; alternate: "pop {pc}"
.code 32 @ ARM mode for ease of Carry manipulation
get32: @ In: Carry set (unchanged until final adcs)
ldrb bits,[src],#1
ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #1*8
ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #2*8
ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #3*8
adcs bits,bits,bits @ Set Carry out
bx lr
.size ucl_nrv2b_decompress_32, .-ucl_nrv2b_decompress_32
/*
vi:ts=8:et:nowrap
*/
+41 -42
View File
@@ -40,22 +40,6 @@
#define cnt r1 /* overlaps 'len' while reading an offset */ #define cnt r1 /* overlaps 'len' while reading an offset */
ucl_nrv2e_decompress_32: .globl ucl_nrv2e_decompress_32 @ ARM mode
/* error = (*)(char const *src, int len_src, char *dst, int *plen_dst) */
add r1,len,src @ r1= eof_src;
stmfd sp!,{r1,r2,r3, r4,r5,r6,r7,lr}
blx hitch_n2e
hitch_n2e_r:
ldmfd sp!,{r4,r5,r6,r7,pc}
get32: @ ARM mode; In: Carry set (unchanged until final adcs)
ldrb bits,[src],#1
ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #1*8
ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #2*8
ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #3*8
adcs bits,bits,bits @ Set Carry out
bx lr
#define GETBIT \ #define GETBIT \
add bits,bits; beq 1f; 0: \ add bits,bits; beq 1f; 0: \
.subsection 1; \ .subsection 1; \
@@ -66,27 +50,31 @@ get32: @ ARM mode; In: Carry set (unchanged until final adcs)
#define jnextb0 GETBIT; bcc #define jnextb0 GETBIT; bcc
#define jnextb1 GETBIT; bcs #define jnextb1 GETBIT; bcs
.code 16 @ THUMB mode ucl_nrv2e_decompress_32: .globl ucl_nrv2e_decompress_32 @ ARM mode
eof_n2e: .type ucl_nrv2e_decompress_32, %function
pop {r1,r3,r4} @ r1= eof_src; r3= orig_dst; r4= plen_dst /* error = (*)(char const *src, int len_src, char *dst, int *plen_dst) */
sub src,r1 @ 0 if actual src length equals expected length adr r12,1+thumb_nrv2e_d32; bx r12 @ enter THUMB mode
sub dst,r3 @ actual dst length .code 16 @ THUMB mode
str dst,[r4] thumb_nrv2e_d32:
sub g32,#get32 - hitch_n2e_r @ g32= &hitch_n2e_r add r1,len,src @ r1= eof_src;
bx g32 push {r1,r2,r3, r4,r5,r6,r7, lr}
mov bits,#1; neg off,bits @ off= -1 initial condition
hitch_n2e:
mov g32,lr @ return address
add g32,#get32 - hitch_n2e_r @ g32= &get32
mov bits,#1; neg off,bits @ off= -1 initial condition
lsl bits,#31 @ 1<<31; refill next time lsl bits,#31 @ 1<<31; refill next time
mov wrnk,#5 mov wrnk,#5
lsl wrnk,#8 @ 0x500 lsl wrnk,#8 @ 0x500
b top_n2e adr g32,get32 @ load pc-relative address
b top_n2e
eof_n2e:
pop {r1,r3,r4} @ r1= eof_src; r3= orig_dst; r4= plen_dst
sub src,r1 @ 0 if actual src length equals expected length
sub dst,r3 @ actual dst length
str dst,[r4]
pop {r4,r5,r6,r7, pc} @ return
lit_n2e: lit_n2e:
ldrb tmp,[src]; add src,#1 ldrb tmp,[src]; add src,#1
strb tmp,[dst]; add dst,#1 strb tmp,[dst]; add dst,#1
top_n2e: top_n2e:
jnextb1 lit_n2e jnextb1 lit_n2e
mov cnt,#1; b getoff_n2e mov cnt,#1; b getoff_n2e
@@ -101,11 +89,11 @@ getoff_n2e:
sub tmp,cnt,#3 @ set Carry sub tmp,cnt,#3 @ set Carry
mov len,#0 @ Carry unaffected mov len,#0 @ Carry unaffected
blo offprev_n2e @ cnt was 2; tests Carry only blo offprev_n2e @ cnt was 2; tests Carry only
lsl off,tmp,#8 lsl tmp,#8
ldrb tmp,[src]; add src,#1 ldrb off,[src]; add src,#1 @ low 7+1 bits
orr off,tmp orr off,tmp
mvn off,off; beq eof_n2e @ off= ~off mvn off,off; beq eof_n2e @ off= ~off
asr off,#1; bcs lenlast_n2e asr off,#1; bcs lenlast_n2e
b lenmore_n2e b lenmore_n2e
offprev_n2e: offprev_n2e:
@@ -126,12 +114,23 @@ gotlen_n2e: @ 'cmn': add the inputs, set condition codes, discard the sum
cmn off,wrnk; bcs near_n2e @ within M2_MAX_OFFSET cmn off,wrnk; bcs near_n2e @ within M2_MAX_OFFSET
add len,#1 @ too far away, so minimum match length is 3 add len,#1 @ too far away, so minimum match length is 3
near_n2e: near_n2e:
ldrb tmp,[dst] @ force cacheline allocate ldrb tmp,[dst] @ force cacheline allocate
copy_n2e: copy_n2e:
ldrb tmp,[dst,off] ldrb tmp,[dst,off]
strb tmp,[dst]; add dst,#1 strb tmp,[dst]; add dst,#1
sub len,#1; bne copy_n2e sub len,#1; bne copy_n2e
b top_n2e b top_n2e
.code 32 @ ARM mode for ease of Carry manipulation
get32: @ In: Carry set (unchanged until final adcs)
ldrb bits,[src],#1
ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #1*8
ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #2*8
ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #3*8
adcs bits,bits,bits @ Set Carry out
bx lr
.size ucl_nrv2e_decompress_32, .-ucl_nrv2e_decompress_32
/* /*
vi:ts=8:et:nowrap vi:ts=8:et:nowrap