ARM: 8745/1: get rid of __memzero()

The __memzero assembly code is almost identical to memset's except for
two orr instructions. The runtime performance of __memset(p, n) and
memset(p, 0, n) is accordingly almost identical.

However, the memset() macro used to guard against a zero length and to
call __memzero at compile time when the fill value is a constant zero
interferes with compiler optimizations.

Arnd found tha the test against a zero length brings up some new
warnings with gcc v8:

  https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82103

And successively rremoving the test against a zero length and the call
to __memzero optimization produces the following kernel sizes for
defconfig with gcc 6:

    text     data     bss       dec       hex  filename
12248142  6278960  413588  18940690   1210312  vmlinux.orig
12244474  6278960  413588  18937022   120f4be  vmlinux.no_zero_test
12239160  6278960  413588  18931708   120dffc  vmlinux.no_memzero

So it is probably not worth keeping __memzero around given that the
compiler can do a better job at inlining trivial memset(p,0,n) on its
own. And the memset code already handles a zero length just fine.

Suggested-by: Arnd Bergmann <[email protected]>
Signed-off-by: Nicolas Pitre <[email protected]>
Acked-by: Ard Biesheuvel <[email protected]>
Acked-by: Arnd Bergmann <[email protected]>
Signed-off-by: Russell King <[email protected]>
This commit is contained in:
Nicolas Pitre 2018-01-19 18:17:46 +01:00 • committed by Russell King
commit ff5fdafc9e
7 changed files with 5 additions and 161 deletions

View file

@ -765,7 +765,7 @@ static int imxdma_alloc_chan_resources(struct dma_chan *chan)
desc = kzalloc(sizeof(*desc), GFP_KERNEL);
if (!desc)
break;
__memzero(&desc->desc, sizeof(struct dma_async_tx_descriptor));
memset(&desc->desc, 0, sizeof(struct dma_async_tx_descriptor));
dma_async_tx_descriptor_init(&desc->desc, chan);
desc->desc.tx_submit = imxdma_tx_submit;
/* txd.flags will be overwritten in prep funcs */