------- Comment #2 from vda dot linux at googlemail dot com 2007-07-25 15:05
-------
Created an attachment (id=13973)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=13973&action=view)
Fix: adjust div cost for -Os on i386
Patch was tested with 4.2.1, I guess it will apply to other versions of gcc as
it is quite trivial.
Test program with lots or randomly-generated constant divisions and %'s was
compiled by patched gcc with different costs of division:
text data bss dec hex filename
257731 0 0 257731 3eec3 421.org/t-Os.o
256788 0 0 256788 3eb14 421.7/t-Os.o
256788 0 0 256788 3eb14 421.8/t-Os.o
257377 0 0 257377 3ed61 421.9/t-Os.o
257825 0 0 257825 3ef21 421.10/t-Os.o
Seems like (at least on 4.2.1) cost of 8 is giving smallest code.
Among 15000 divisions in test program, only signed divisions by power-of-two
grew in size (but they become MUCH faster, as they don't even use multiply now,
let alone div):
@@ -1703 +1703 @@
-0000000f T id_x_16
+00000012 T id_x_16
@@ -1836 +1836 @@
-0000000f T id_x_2
+0000000e T id_x_2
@@ -2030 +2030 @@
-0000000f T id_x_32
+00000012 T id_x_32
id_x_16 was:
movl 4(%esp), %eax
movl $16, %edx
movl %edx, %ecx
cltd
idivl %ecx
ret
Now it is:
movl 4(%esp), %edx
movl %edx, %eax
sarl $31, %eax
andl $15, %eax
addl %edx, %eax
sarl $4, %eax
ret
and also unsigned_x / 28, unsigned_x / 13952, unsigned_x / 56 grew by 1 byte.
The rest either were not changed (and still use div insn) or shrank (typically
by 1 byte, record holders are "unsigned_x / 641" and "unsigned_x / 6700417" -
shrank by 4 bytes).
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30354
