jory/upx
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

tune: avoid MOVSB for literals; inline default case of NEXTB macros

Browse Source 
committer: jreiser <jreiser> 1131923823 +0000
This commit is contained in:
John Reiser
2005-11-13 23:17:03 +00:00
parent db5a099097
commit 68f8358f06
1 changed files with 13 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/stub/amd_d_nrv2e.S
+13 -11
View File
@@ -60,34 +60,35 @@ decompress: # (uchar const *src, size_t lsrc, uchar *dst, size_t &ldst, uint me
movq src,%rsi # hardware src for movsb, lodsb movq src,%rsi # hardware src for movsb, lodsb
movq dst,%rdi # hardware dst for movsb movq dst,%rdi # hardware dst for movsb
subl bits,bits # empty; force refill xorl bits,bits # empty; force refill
subl len,len # create loop invariant xorl len,len # create loop invariant
orq $~0,disp # -1: initial displacement orq $~0,disp # -1: initial displacement
call setup_rdx call setup_rdx
ra_setup_rdx: ra_setup_rdx:
/* jump on next bit {0,1} with prediction {y==>likely, n==>unlikely} */ /* jump on next bit {0,1} with prediction {y==>likely, n==>unlikely} */
/* Prediction omitted for now. */ /* Prediction omitted for now. */
#define jnextb0n call *%rdx; jnc #define jnextb0n jnextb0y
#define jnextb0y call *%rdx; jnc #define jnextb0y addl bits,bits; jnz 0f; call refill; 0: jnc
#define jnextb1n call *%rdx; jc #define jnextb1n jnextb1y
#define jnextb1y call *%rdx; jc #define jnextb1y addl bits,bits; jnz 0f; call refill; 0: jc
/* rotate next bit into bottom bit of reg */ /* rotate next bit into bottom bit of reg */
#define getnextb(reg) call *%rdx; adcl reg,reg #define getnextb(reg) addl bits,bits; jnz 0f; call refill; 0: adcl reg,reg
ALIGN(1<<3) ALIGN(1<<3)
getbit: getbit:
addl bits,bits; jz refill # Carry= next bit addl bits,bits; jz refill # Carry= next bit
ret rep; ret
refill: refill:
movl (%rsi),bits; subq $-4,%rsi # next 32 bits; set Carry movl (%rsi),bits; subq $-4,%rsi # next 32 bits; set Carry
adcl bits,bits # LSB= 1 (CarryIn); CarryOut= next bit adcl bits,bits # LSB= 1 (CarryIn); CarryOut= next bit
ret rep; ret
ALIGN(1<<3) ALIGN(1<<3)
lit_n2e: lit_n2e:
movsb # *%rdi++ = *%rsi++; movb (%rsi),%al; addq $1,%rsi
movb %al,(%rdi); addq $1,%rdi
top_n2e: top_n2e:
jnextb1y lit_n2e jnextb1y lit_n2e
lea 1(lenq),off # [len= 0] off= 1 lea 1(lenq),off # [len= 0] off= 1
@@ -137,7 +138,8 @@ bot_n2e: # In: 0==len
ALIGN(1<<3) ALIGN(1<<3)
lit_n2b: lit_n2b:
movsb # *%rdi++ = %rsi++; movb (%rsi),%al; addq $1,%rsi
movb %al,(%rdi); addq $1,%rdi
top_n2b: top_n2b:
jnextb1y lit_n2b jnextb1y lit_n2b
lea 1(lenq),off # [len= 0] off= 1 lea 1(lenq),off # [len= 0] off= 1