JIT: Optimize counted loops by reversing them #100915
Description
The JIT should be able to optimize counted loops like
private static void Foo()
{
for (int i = 0; i < 100; i++)
{
Bar();
}
}to the equivalent downwards counting loop, which is more efficient on both x64 and arm64.
Current codegen:
x64:
G_M35517_IG02: ;; offset=0x0005
xor ebx, ebx
;; size=2 bbWeight=1 PerfScore 0.25
G_M35517_IG03: ;; offset=0x0007
call [RangeCheck_Overflow:Bar()]
inc ebx
cmp ebx, 100
jl SHORT G_M35517_IG03
;; size=13 bbWeight=4 PerfScore 18.00arm64:
G_M35517_IG02: ;; offset=0x000C
mov w19, wzr
;; size=4 bbWeight=1 PerfScore 0.50
G_M35517_IG03: ;; offset=0x0010
movz x0, #0x6820 // code for Program:Bar()
movk x0, #0x6A29 LSL #16
movk x0, #0x7FF9 LSL #32
ldr x0, [x0]
blr x0
add w19, w19, #1
cmp w19, #100
blt G_M35517_IG03
;; size=32 bbWeight=4 PerfScore 30.00Expected codegen:
x64:
G_M35517_IG02: ;; offset=0x0005
mov ebx, 100
;; size=5 bbWeight=1 PerfScore 0.25
G_M35517_IG03: ;; offset=0x000A
call [RangeCheck_Overflow:Bar()]
dec ebx
jne SHORT G_M35517_IG03
;; size=10 bbWeight=4 PerfScore 17.00arm64:
G_M35517_IG02: ;; offset=0x000C
mov w19, #100
;; size=4 bbWeight=1 PerfScore 0.50
G_M35517_IG03: ;; offset=0x0010
movz x0, #0x6820 // code for Program:Bar()
movk x0, #0x6A2B LSL #16
movk x0, #0x7FF9 LSL #32
ldr x0, [x0]
blr x0
sub w19, w19, #1
cbnz w19, G_M35517_IG03
;; size=28 bbWeight=4 PerfScore 28.00