X-Git-Url: http://wien.tomnetworks.com/gitweb/?a=blobdiff_plain;f=asma%2Fasma.s;h=d1202ced60a193c1466f7ae3035f0d99abff2a5a;hb=7ca578dcc2d598744196f55a89ff1e0d6dfa1fd5;hp=4881e0b0924b2bc6f955b1fd1cebbce808fa6329;hpb=e195ada341d6b5f6aa5f8969430fee9a6a4780b9;p=uebersetzerbau-ss10.git

diff --git a/asma/asma.s b/asma/asma.s
index 4881e0b..d1202ce 100644
--- a/asma/asma.s
+++ b/asma/asma.s
@@ -1,51 +1,57 @@
 	.file	"asma.c"
+
+.section .rodata
+.align 128
+const9a:
+	.rept 16
+	.byte 0x9a
+	.endr
+const20:
+	.rept 16
+	.byte 0x20
+	.endr
+const3f:
+	.rept 16
+	.byte 0x3f
+	.endr
+
 	.text
 .globl asma
 	.type	asma, @function
 asma:
 .LFB2:
-	/*init %xmm1 mit "'Z' + 1 + min_t - 'A'" fuer jedes byte
+	/*init %xmm8 mit "'Z' + 1 + min_t - 'A'" fuer jedes byte
 	* 'Z' + 1 + min_t - 'A' = 90 + 1 - 128 - 65 = -102
 	* 102 = 01100110
 	* ~102 = 10011001
 	* (~102)+1 = 10011010 = 0x9A */
-	mov $0x9a9a9a9a9a9a9a9a, %rbx
-	movq %rbx, %xmm7
-	movddup %xmm7, %xmm1
-
-	//init %xmm2 mit "'a'-'A'= 97-65 = 32 = 0x20
-	mov $0x2020202020202020, %rbx
-	movq %rbx, %xmm7
-	movddup %xmm7, %xmm2
+	movdqa const9a, %xmm8
 
 	//speicheradresse des pointers zeigt auf 16*8 feld = 128bit
-	movdqu (%rdi), %xmm4
-	movdqu (%rdi), %xmm5
-
-	/*addiere in %xmm4 "min_t-'A'"
-	* = -128 - 65 = 63 = 0x3f */
-	mov $0x3f3f3f3f3f3f3f3f, %rbx
-	movq %rbx, %xmm7
-	movddup %xmm7, %xmm3
+	movdqu (%rdi), %xmm11
 
-	//c+min_t-'A'
-	paddb %xmm3, %xmm4
+	/*addiere in %xmm11 "min_t-'A'"
+	* = -128 - 65 = 63 = 0x3f
+	* c+min_t-'A' */
+	paddb const3f, %xmm11
 
 	/*"Packed COMpare Greater Than (Byte)"
 	* X = 'Z' + 1 + min_t - 'A' > c + min_t - 'A' ? 0xff : 0
 	* achtung beim intuitiven lesen des befehles. 'kleiner' 
 	* ist mit 'groesser' vertauscht und vice versa */
-	pcmpgtb %xmm4, %xmm1
+	pcmpgtb %xmm11, %xmm8
 
-	//Y = min(X, 'a' - 'A')
-	pminub %xmm2, %xmm1
-
-	//c += Y
-	paddb %xmm1, %xmm5
+	/*Y = min(X, 'a' - 'A')
+	* mit "'a'-'A'= 97-65 = 32 = 0x20 */
+	pminub const20, %xmm8
 
 	//retuniere an die richtige speicheradresse
 	mov %rdi, %rax
-	movapd %xmm5, (%rax)
+	//c += Y
+	psubb const3f, %xmm11
+	paddb %xmm11, %xmm8
+	movdqu %xmm8, (%rax)
+
 	ret
 .LFE2:
 	.size	asma, .-asma