arm-linux nrv2b works: conditon code bug, cache flush, cache lookahead

src/stub/src/arch/arm/v4a/nrv2b_d8.S

@@ -69,6 +69,20 @@ eof_n2b:
         SUB2(src,r1)  // 0 if actual src length equals expected length
         SUB2(dst,r3)  // actual dst length
         str dst,[r4]
+
+#if defined(LINUX_ARM_CACHEFLUSH)  /*{*/
+/* linux/include/asm-arm/unistd.h */
+#define __NR_SYSCALL_BASE       0x900000
+#define __ARM_NR_BASE           (__NR_SYSCALL_BASE+0x0f0000)
+#define __ARM_NR_cacheflush     (__ARM_NR_BASE+2)
+        mov r4,r0  // save result value
+        mov r0,r3  // orig_dst
+        add r1,r3,dst  // orig_dst + dst_len
+        mov r2,#0
+        swi __ARM_NR_cacheflush  // decompressed region
+        mov r0,r4  // result value
+#endif  /*}*/
+
         POP {r4,r5, pc}  // return
 
 get8_n2b:  // In: Carry set [from adding 0x80000000 (1<<31) to itself]
@@ -88,7 +102,7 @@ getoff_n2b:  // ss11 len= [2..)
         getnextb(len)
         jnextb0 getoff_n2b
 
-        sub tmp,len,#3  // set Carry
+        subs tmp,len,#3  // set Carry
         mov len,#0  // Carry unaffected
         blo offprev_n2b  // ss11 returned 2
         LDRB3(off,[src],#1)  // low 8 bits
@@ -107,7 +121,10 @@ plus1_n2b:
 /* 'cmn': add the inputs, set condition codes, discard the sum */
         cmn off,#0xd<<8  // within M2_MAX_OFFSET
         addcc len,len,#1  // too far away, so minimum match length is 3
-        ldrb tmp,[dst]  // force cacheline allocate
+
+        add tmp,dst,len
+        ldrb tmp,[tmp,#-1]  // force cacheline allocate
+
 copy_n2b:
         ldrb tmp,[dst,off]
         STRB3(tmp,[dst],#1)

src/stub/src/arch/arm/v4a/nrv2e_d8.S

@@ -179,7 +179,8 @@ near_n2e:
         add tmp,dst,cnt
         cmp tmp,dstlim; bhi bad_dst_n2e  // too much output
 #endif  /*}*/
-        ldrb tmp,[dst]  // force cacheline allocate
+        add tmp,dst,len
+        ldrb tmp,[tmp,#-1]  // force cacheline allocate
 copy_n2e:
         ldrb tmp,[dst,off]
         CHECK_BYTE