cmd/compile: large, simple array 75x slower to compile with -race

[josharian]

Compile this file with and without -race:

https://gist.githubusercontent.com/josharian/7e19da2d4c68d004ffebf510aeb8ef69/raw/5c475f906ce7a515d45d8bb5ae5c701469682b8a/x.go

On my laptop, it compiles without -race in 1.5s. With -race, it takes over 100s.

This is extracted and simplified from real code that was causing repeated CI build timeouts, meaning it was taking upwards of 10 minutes to build on a slow CI machine.

Reproduced with both 1.19.4 and 1.20rc2.

I haven't investigated at all.

[cuonglm]

It seems to me that -race forces new(T) to be copied, causing quadratic behavior of cmd/compile/internal/ssa.Func.computeZeroMap. Disable writebarrier brings the speed back:

time go tool compile -race -wb=false p.go
=====
JOB go tool compile -race -wb=false p.go
128%    cpu
1.80s real
2.18s user
0.12s sys

Func.computeZeroMap was introduced in CL 156363 cc @randall77

[gopherbot]

Change https://go.dev/cl/461142 mentions this issue: cmd/compile: lazy compute zero region map

[gopherbot]

Change https://go.dev/cl/461142 mentions this issue: cmd/compile: computing zero region map lazily

[cuonglm]

Ops @gopherbot, why did you post two comments 😃

[randall77]

Looks like the quadratic behavior is in IsNewObject.

   143.75s    144.54s (flat, cum) 96.47% of Total
         .          .    589:func IsNewObject(v *Value) (mem *Value, ok bool) {
      20ms       20ms    590:	f := v.Block.Func
      10ms       10ms    591:	c := f.Config
         .          .    592:	if f.ABIDefault == f.ABI1 && len(c.intParamRegs) >= 1 {
         .          .    593:		if v.Op != OpSelectN || v.AuxInt != 0 {
         .          .    594:			return nil, false
         .          .    595:		}
         .          .    596:		// Find the memory
     4.47s      4.49s    597:		for _, w := range v.Block.Values {
   139.25s    140.01s    598:			if w.Op == OpSelectN && w.AuxInt == 1 && w.Args[0] == v.Args[0] {
         .          .    599:				mem = w
         .          .    600:				break
         .          .    601:			}
         .          .    602:		}

That's called from computeZeroMap.
This was introduced in https://go-review.googlesource.com/c/go/+/308709 @cherrymui

Probably there is a better way to get from SelectN(0) to SelectN(1).

[gopherbot]

Change https://go.dev/cl/461080 mentions this issue: cmd/compile: prevent IsNewObject from taking quadratic time