redundant conversion emitted?

[GoogleCodeExporter]

while debugging issue 32 (loop unrolling in combination with heavy-duty
optimization), I came upon this code (produced with -O3 + -funroll-loops,
though it's probably the same with -O0 as well; comments by me):

    ld1b    $s32,0(,$s39) # $s32 <- src[0]
    sll $s32,$s32,56  # $s32 <- $s32 << 56
    srax    $s32,$s32,56  # $s32 <- $s32 >> 56
    st1b    $s32,0(,$s38) # dst[0] <- $s32

it seems that gcc uses sll+srax combination to make sure that top 56 bits
of the register to/from which chars are loaded/stored are 0. this is imho
redundant as ld1b "loads 1 byte, zero extended". since ld1b is not sign
extended, sll+srax after it is redundant and should not be emitted by the
compiler.

saving 2 cycles here and 2 cycles there makes for a whole lot of cycles ... ;)

Original issue reported on code.google.com by jmoc...@gmail.com on 26 Nov 2008 at 3:14

[GoogleCodeExporter]

hrumph ... sll + sra are exactly a sign extension. and this is actually desired 
for
the compiled code.

Original comment by jmoc...@gmail.com on 6 Mar 2009 at 12:24

• Changed state: Invalid