This adds a penalty of 4 to the post-reload branch costs.
Purpose is reduce the number of out-of-line blocks like in
unsigned long variant5 (unsigned in)
{
unsigned long out = 0;
if (in & (1 << 0)) out |= 0xful << (4*0);
if (in & (1 << 1)) out |= 0xful << (4*1);
if (in & (1 << 2)) out |= 0xful << (4*2);
if (in & (1 << 3)) out |= 0xful << (4*3);
return out;
}
without the patch, code is
variant5:
mov r18,r24 ; 67 movqi_insn/1 [length = 1]
sbrs r24,0 ; 10 *sbrx_branchhi [length = 2]
rjmp .L6
ldi r22,lo8(15) ; 5 *movsi/5 [length = 4]
ldi r23,0
ldi r24,0
ldi r25,0
.L2:
<some_code_and_epilogue>
.L6:
ldi r22,0 ; 4 *movsi/2 [length = 3]
ldi r23,0
movw r24,r22
rjmp .L2 ; 74 jump [length = 1]
and with the patch it reads:
variant5:
mov r18,r24 ; 67 movqi_insn/1 [length = 1]
ldi r22,lo8(15) ; 5 *movsi/5 [length = 4]
ldi r23,0
ldi r24,0
ldi r25,0
sbrc r18,0 ; 10 *sbrx_branchhi [length = 2]
rjmp .L2
ldi r22,0 ; 4 *movsi/2 [length = 3]
ldi r23,0
movw r24,r22
.L2:
<some_code_and_epilogue>
Using fall-through safes insn 74.
Main blocker for not increasing default branch costs in general
is do_store_flag which is a heap of assertions not using
rtx_costs and which gives the best results with the old
default of 0, which is not changed.
Tested without regressions.
Ok for trunk?
Johann
* config/avr/avr.h (BRANCH_COST) [reload_completed]: Increase by 4.
Index: config/avr/avr.h
===================================================================
--- config/avr/avr.h (revision 244001)
+++ config/avr/avr.h (working copy)
@@ -360,7 +360,12 @@ typedef struct avr_args
} \
} while (0)
-#define BRANCH_COST(speed_p, predictable_p) avr_branch_cost
+/* We increase branch costs after reload in order to keep basic-block
+ reordering from introducing out-of-line jumps and to prefer fall-through
+ edges instead. The default branch costs are 0, mainly because otherwise
+ do_store_flag might come up with bloated code. */
+#define BRANCH_COST(speed_p, predictable_p) \
+ (avr_branch_cost + (reload_completed ? 4 : 0))
#define SLOW_BYTE_ACCESS 0
