Hi,
On 20/5/2021 下午 9:08, Segher Boessenkool wrote:
Hi!
On Thu, May 20, 2021 at 04:29:07PM +0800, HAO CHEN GUI wrote:
On 19/5/2021 下午 9:20, Segher Boessenkool wrote:
On Wed, May 19, 2021 at 04:36:00PM +0800, HAO CHEN GUI wrote:
-/* Define this macro if it is advisable to hold scalars in registers
- in a wider mode than that declared by the program. In such cases,
- the value is constrained to be within the bounds of the declared
- type, but kept valid in the wider mode. The signedness of the
- extension may differ from that of the type. */
-
-#define PROMOTE_MODE(MODE,UNSIGNEDP,TYPE) \
- if (GET_MODE_CLASS (MODE) == MODE_INT \
- && GET_MODE_SIZE (MODE) < (TARGET_32BIT ? 4 : 8)) \
- (MODE) = TARGET_32BIT ? SImode : DImode;
-
And this part needs some more words in the commit message :-)
[ ... ]
Also, how about something like
#define PROMOTE_MODE(MODE,UNSIGNEDP,TYPE) \
if (GET_MODE_CLASS (MODE) == MODE_INT \
&& GET_MODE_SIZE (MODE) < 4) \
(MODE) = SImode;
(that is, promoting modes smaller than SImode to SImode). How does that
compare?
It hits an ICE when assigning a function return value to a variable. The
mode of variable is promoted to SImode, but the mode of return value is
promoted to DImode. Current GCC logical is either the mode of set_dest
is unchanged or the mode of set_dest matches the mode of function return
value.
So that sounds like a bug in the generic code?
By the way, I tested the patch on ppc64 with m32 option. The SPECint
shows a little improvement(0.4%). So it's better not do any mode promotions.
Interesting! Do you have an example that shows improved code for that,
just like that "SI->DI promotion gets a superfluous extension" testcase
you sent?
perlbench and omnetpp_r got 3-4% performance improvement with the patch
on ppc64 m32. I tried to find an example in these two benchmarks, but I
failed. The improvement of perlbench comes from the different sequence
of basic block. For omnetpp, one of improvement comes from condition
register not spilling. But it is not directly related to the mode
promotion.
By examming the ISA, I found there is no instructions for QI and HI
mode(except load and store). So QI and HI are extended to SI mode
automatically when they're used in an expression with the patch. So with
the patch, gcc does implicit mode promotion for QI and HI. With original
gcc, it does explicit mode promotion. They should have almost the same
behavior.
Segher
