bromley (02/24/83)
In generating code to extract values from fields, the compiler uses only the
extv/extvz instructions. The optimizer changes all extractions of
byte/word fields on byte/word boundaries to cvt{b,w}l or movz{b,w}l depending
on the field being signed or unsigned. However, it does this incorrectly
for word size fields on non-longword boundaries. The specific transformations
it performs are:
extv $n*16, $16, A, B ==> cvtwl n+A,B
extvz $n*16, $16, A, B ==> movzwl n+A,B
The actual address of the field is 2n+A, as we are dealing with a word sized
field and not a byte sized one. The correct transformations are then:
extv $n*16, $16, A, B ==> cvtwl 2n+A,B
extvz $n*16, $16, A, B ==> movzwl 2n+A,B
The code that implements this is in c21.c and is easily fixed. The changes
are:
263c263
< ** extv $n*16,$16,A,B > cvtwl n+A,B
---
> ** extv $n*16,$16,A,B > cvtwl 2n+A,B
265c265
< ** extzv $n*16,$16,A,B > movzwl n+A,B
---
> ** extzv $n*16,$16,A,B > movzwl 2n+A,B
310c310,312
< sprintf(regs[RT1], "%d%s%s", coff, regs[RT3][0]=='(' ? "":"+",
---
> sprintf(regs[RT1], "%d%s%s",
> (flen == 8 ? coff : 2*coff),
> (regs[RT3][0] == '(' ? "" : "+"),
Mark Bromley
cornell