mirror/upx
1
0
Fork 0
mirror of https://github.com/upx/upx synced 2025-10-05 19:20:23 +08:00
Code Issues Releases Wiki Activity

Fix THUMB mode return to ARM mode on ARMv4T. Spend 18 bytes to inline

Browse Source 
most of get1b via conditional subroutine call idiom.

committer: jreiser <jreiser> 1143302120 +0000
This commit is contained in:
John Reiser 2006-03-25 15:55:20 +00:00
parent a6e7afbc91
commit a7bff34784
1 changed files with 12 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
src/stub/arm_nrv2e_d8.S
View File

@ -62,14 +62,8 @@
#define CHECK_BYTE /*empty*/ #define CHECK_BYTE /*empty*/
#endif /*}*/ #endif /*}*/
/* Putting get1_n2e in a register [r6:wrnk] inhibits branch prediction, /* "mov lr,pc; bxx ..." implements conditional subroutine call */
and saves only 14 bytes (9 calls, but 2 Thumb instr to setup). #define GETBIT add bits,bits; mov lr,pc; beq get1_n2e
'bl' takes 4 bytes and 2 cycles. It is tempting to inline
"add bits,bits; beq <reload>" instead, but branching back costs
9*4 bytes with 4-byte alignment (adr tmp,<aligned_back>; b fetch8), or
9*6 bytes without alignment (bl fetch8; b <back>).
*/
#define GETBIT bl get1_n2e
#define getnextb(reg) GETBIT; adc reg,reg #define getnextb(reg) GETBIT; adc reg,reg
#define jnextb0 GETBIT; bcc #define jnextb0 GETBIT; bcc
@ -109,7 +103,16 @@ eof_n2e:
sub src,srclim @ 0 if actual src length equals expected length sub src,srclim @ 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]
pop {r4,r5,r6,r7, pc} @ return pop {r4,r5,r6,r7 /*,pc*/}
pop {r1}; bx r1 @ "pop {,pc}" fails return to ARM mode on ARMv4T
get1_n2e: @ In: Carry set [from adding 0x80000000 (1<<31) to itself]
ldrb bits,[src] @ zero-extend next byte
adc bits,bits @ double and insert CarryIn as low bit
CHECK_SRC
add src,#1
lsl bits,#24 @ move to top byte, and set CarryOut from old bit 8
mov pc,lr @ return, stay in current (THUMB) mode
lit_n2e: lit_n2e:
CHECK_SRC; ldrb tmp,[src]; add src,#1 CHECK_SRC; ldrb tmp,[src]; add src,#1
@ -171,17 +174,6 @@ copy_n2e:
b top_n2e b top_n2e
.size ucl_nrv2e_decompress_8, .-ucl_nrv2e_decompress_8 .size ucl_nrv2e_decompress_8, .-ucl_nrv2e_decompress_8
get1_n2e: .type get1_n2e, %function
add bits,bits; bne get1r_n2e @ CarryOut has data bit
ldrb bits,[src] @ zero-extend next byte
adc bits,bits @ double and insert CarryIn as low bit
CHECK_SRC
add src,#1
lsl bits,#24 @ move to top byte, and set CarryOut from old bit 8
get1r_n2e:
bx lr
.size get1_n2e, .-get1_n2e
/* /*
vi:ts=8:et:nowrap vi:ts=8:et:nowrap
*/ */