Le 14/08/2019 à 04:08, Paul Mackerras a écrit :
On Tue, Aug 13, 2019 at 09:59:35AM +0000, Christophe Leroy wrote:
[snip]
+.macro __LOAD_REG_IMMEDIATE r, x
+ .if \x & ~0xffffffff != 0
+ __LOAD_REG_IMMEDIATE_32 \r, (\x) >> 32
+ rldicr \r, \r, 32, 31
+ .if (\x) & 0xffff0000 != 0
+ oris \r, \r, (\x)@__AS_ATHIGH
+ .endif
+ .if (\x) & 0xffff != 0
+ oris \r, \r, (\x)@l
+ .endif
+ .else
+ __LOAD_REG_IMMEDIATE_32 \r, \x
+ .endif
+.endm
Doesn't this force all negative constants, even small ones, to use
the long sequence? For example, __LOAD_REG_IMMEDIATE r3, -1 will
generate (as far as I can see):
li r3, -1
rldicr r3, r3, 32, 31
oris r3, r3, 0xffff
ori r3, r3, 0xffff
which seems suboptimal.
Ah yes, thanks. And it is also buggy when \x is over 0x80000000 because
lis is a signed ops
I'll send v2
Christophe
