http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46453
Summary: MIPS backend is not using special instructions for
__builtin_bswap32
Product: gcc
Version: 4.5.1
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: target
AssignedTo: [email protected]
ReportedBy: [email protected]
Host: i686
Target: mips-elf
Build: ../configure --disable-libssp --prefix=/usr/local
--target=mips-elf
MIPS32 Relase 2 introduced a special instruction called wsbh that can be used
for 32- and 16-bit byteswaps. However GCC never does produce this instruction.
all code-snippets are produced with "-march=mips32r2 -O3" (which should enable
the wsbh, rotr and ins instructions and optimized code using them). the
assembly assumes big-endian endian.
* __builtin_bswap32:
eg. "v0 = __builtin_bswap32(a0);" should result in
wsbh v0, a0
rotr v0, v0, 16
* 16bit byteswaps:
similarly "v0 = ((a0 >> 8) | (a0 << 8));" (with a0,v0 being 16 bit uints)
should result in:
wsbh v0, a0
as it is now, the __builtin_bswap32 will always result in a function call
(already atleast 2 instructions) and the implementation which uses 9
instruction.
a 16bit bswap results in 4 instructions. So that would be nice savings
especially in the 32bit case
* Unaligned loads:
More unimportantly unaligned 16bit loads could be optimized a bit aswell if
the ins instruction is available:
--Code sample (unaligned 16bit load):
#pragma pack(push,1)
union Unaligned {
unsigned char c[2];
unsigned short u16;
};
#pragma pack(pop)
unsigned short readUnaligned16(const void *ptr) {
return ((const union Unaligned *)(ptr))->u16;
}
-- Code sample
results in this sequence:
# a0 = ptr, v0 = return value
lbu v0,0(a0)
lbu v1,1(a0)
sll v0,v0,0x8
or v0,v1,v0
better would be:
lbu v0,0(a0)
lbu v1,1(a0)
ins v1,v0,8,8
Generating this sequences for unaligend 16bit loads would be a nice start. But
there could be generic optimizations with sequences of left-shift and or being
replaced with ins instructions, aslong it can be verified that registers have
enough explicitly zeroed bits so they dont "overlap".
similarly right-shift and masking could be replaced by ext instructions. eg. v0
= ((a0 >> 8) & 0xFF) equals to ext v0,a0,8,8.