[Sprint 3] Cost increase after applying load vectorize

[rorhcdream]

Problem

matmul4 cost is increased about 15% (and 10% for matmul1) after load vectorize is applied. The main reason is that variables are assigned to the stack due to a lack of registers. Temperature and load/store stack costs are increased a lot.

Because vectorize pass sinks all loads use, they take registers for longer period. InlinerPass should be removed, and an outliner pass is needed to increase performance.

input8 costs

• sprint 2

  
  Returned: 0
  Cost: 69154896.7062
  Max heap usage (bytes): 240000

temp: 31535562.6993 addsub: 19174294.3978 muldiv: 5442088.0000 load_stack: 5343756.0000 load_heap: 3040000.0000 store_stack: 1812512.0000 logical: 1337705.6000 store_heap: 1003200.0000 comp: 168833.0000 brcond_true: 84336.0000 bruncond: 67656.0000 brcond_false: 63399.0000 read: 40002.0000 write: 30000.0000 vstore_heap: 11520.0000 malloc: 24.0000 call_arg: 4.0000 ret: 2.0000 call: 2.0000 vtemp: 0.0000 vstore_stack: 0.0000 vload_stack: 0.0000 vload_heap: 0.0000 ternary: 0.0000 switch: 0.0000 free: 0.0000 cool: 0.0000

* after load vectorize

Returned: 0 Cost: 79827059.6099 Max heap usage (bytes): 240000

temp: 40608343.5996 addsub: 19120534.3978 load_stack: 6563756.0000 muldiv: 5428890.0000 load_heap: 3001600.0000 store_stack: 2315012.0000 logical: 1314185.6000 store_heap: 1003200.0000 comp: 168833.0000 brcond_true: 84336.0000 bruncond: 67656.0000 brcond_false: 63399.0000 read: 40002.0000 write: 30000.0000 vstore_heap: 11520.0000 vload_heap: 5760.0000 malloc: 24.0000 call_arg: 4.0000 ret: 2.0000 call: 2.0000 vtemp: 0.0000 vstore_stack: 0.0000 vload_stack: 0.0000 ternary: 0.0000 switch: 0.0000 free: 0.0000 cool: 0.0000

### matmul4.opt.ll codes
* sprint 2
```c++
...
define dso_local void @matmul(i32 %dim, i64* %c, i64* %a, i64* %b) #0 {
entry:
  %color35 = alloca i64, align 8
  %color32 = alloca i64, align 8
  %color31 = alloca i64, align 8
  %color30 = alloca i64, align 8
  %color28 = alloca i64, align 8
  %color27 = alloca i64, align 8
  %color24 = alloca i64, align 8
...

• after load vectorize

  ...
  define dso_local void @matmul(i32 %dim, i64* %c, i64* %a, i64* %b) #0 {
  entry:
  %color45 = alloca i64, align 8
  %color44 = alloca i64, align 8
  %color41 = alloca i64, align 8
  %color39 = alloca i64, align 8
  %color37 = alloca i64, align 8
  %color35 = alloca i64, align 8
  %color34 = alloca i64, align 8
  %color29 = alloca i64, align 8
  %color23 = alloca i64, align 8
  %color20 = alloca i64, align 8
  %color19 = alloca i64, align 8
  %color15 = alloca i64, align 8
  %color14 = alloca i64, align 8
  %color13 = alloca i64, align 8
  %color11 = alloca i64, align 8
  ...