given something like:
long long foo(long long X, unsigned int Y) { return X << Y; }
or
long long foo(long long X, unsigned int Y) { return X >> Y; }
We currently generate a libcall to do the shift in 32 bit ppc.
On the contrary, XLC will *always* expand these shifts to an instruction
sequence (LLVM now does the same as of a few days ago), rather than a library
call.
The first will generate with xlc -O3 -qarch=ppc970
.foo:
rlwinm r0,r5,25,0,6
subfic r6,r5,32
slw r3,r3,r5
addi r7,r5,-32
srw r6,r4,r6
xoris r0,r0,0x8000
slw r7,r4,r7
slw r4,r4,r5
or r5,r6,r7
srawi r0,r0,31
or r3,r3,r5
and r4,r0,r4
and r3,r0,r3
and the second:
.foo:
subfic r0,r5,32
addi r8,r5,-32
rlwinm r6,r5,26,0,0
rlwinm r7,r5,25,0,6
srw r8,r3,r8
slw r9,r3,r0
xoris r7,r7,0x8000
srw r0,r4,r5
and r4,r3,r6
or r9,r9,r8
or r0,r0,r9
srawi r6,r7,31
srawi r8,r3,31
rlwinm r7,r5,0,27,31
sraw r4,r4,r7
sraw r3,r3,r5
or r0,r0,r4
andc r5,r8,r6
and r3,r6,r3
and r0,r6,r0
or r4,r5,r0
or r3,r5,r3
(O5 will give slightly better code for both, but let's start somewhere :P)
--
Summary: 64 bit shift by non-constant implemented as libcall on
PPC32
Product: gcc
Version: 4.1.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P2
Component: target
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: dberlin at gcc dot gnu dot org
CC: gcc-bugs at gcc dot gnu dot org
GCC build triplet: powerpc-linux
GCC host triplet: powerpc-linux
GCC target triplet: powerpc-linux
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=20781
