http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50110
Bug #: 50110
Summary: Endian reversal when adding extzv instruction
Classification: Unclassified
Product: gcc
Version: 4.6.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
AssignedTo: [email protected]
ReportedBy: [email protected]
We are in the process of adding a bitfield extract instruction to the SPARC
Leon processor for a custom embedded product
I've added the following to the sparc.md file
(define_insn "extzvsi"
[(set (match_operand:SI 0 "register_operand" "=r")
(zero_extract:SI (match_operand:SI 1 "register_operand" "r")
(match_operand 2 "const_int_operand" "")
(match_operand 3 "const_int_operand" "")))]
""
"ext\t%1,%2,%3,%0"
)
and GCC will start producing assembly code that uses the new opcode
However on simple test programs the endian seems to be reversed
struct x {
unsigned int a:4;
unsigned int b:2;
unsigned int c:3;
unsigned int d:3;
unsigned int e:20;
};
unsigned int a(struct x * arg)
{ return arg->a; }
unsigned int b(struct x * arg)
{ return arg->b; }
unsigned int c(struct x * arg)
{ return arg->c; }
unsigned int d(struct x * arg)
{ return arg->d; }
unsigned int e(struct x * arg)
{ return arg->e; }
compiled with -Os -S
results in
.file "davidm.c"
.section ".text"
.align 4
.global a
.type a, #function
.proc 016
a:
ld [%o0], %o0
jmp %o7+8
srl %o0, 28, %o0
.size a, .-a
.align 4
.global b
.type b, #function
.proc 016
b:
ld [%o0], %o0
jmp %o7+8
ext %o0,2,4,%o0
.size b, .-b
.align 4
.global c
.type c, #function
.proc 016
c:
ld [%o0], %o0
jmp %o7+8
ext %o0,3,6,%o0
.size c, .-c
.align 4
.global d
.type d, #function
.proc 016
d:
ld [%o0], %o0
jmp %o7+8
ext %o0,3,9,%o0
.size d, .-d
.align 4
.global e
.type e, #function
.proc 016
e:
ld [%o0], %g1
sethi %hi(-1048576), %o0
jmp %o7+8
andn %g1, %o0, %o0
.size e, .-e
.ident "GCC: (GNU) 4.6.1"
Just looking at functions a() and b(), a() is using the bitfields in big endian
order as it should for SPARC, but b() seems to be using the bitfields in little
endian order.