cmd/compile: unnecessary shift-check

[JAicewizard]

What version of Go are you using (go version)?

$ go version
1.17.0

Does this issue reproduce with the latest release?

yes

What did you do?

See https://go.godbolt.org/z/o5PqscasE for code.
On line 11/12 panics when endreg, is bigger than 29.
On line 15 it is checked that h is negative.
On line 16 there is a branch that panics on negative -h.

This might make sense if h could be -128, but it cant because of the check on line 11.

What did you expect to see?

I expected the go compiler to combine these checks and to avoid the extra branch+2 instructions.

What did you see instead?

Out of the assembly I gather that the go compiler fails to link line 15 with 16.
The gc also check the same thing twice, it also fails to see that h < 0 is the same as -h >= 0.

https://go.godbolt.org/z/hsMnMa1W1 does what I expect, and the gc manages to proof that -h is less than 0x1f, but more importantly also that -h is non-negative.

[JAicewizard]

another example:
https://godbolt.org/z/Ge9887jEr

[JAicewizard]

Ive been investigating this second example.

Posit32es2.Int64 method(Posit32es2) func() int64
  b1:
    v1 = InitMem <mem>
    v6 = Arg <Posit32es2> {a} (a[Posit32es2])
    v8 = Const64 <int64> [0]
    v9 = StructSelect <uint32> [0] v6
    v10 = Const64 <uint> [2]
    v11 = Lsh32x64 <uint32> [false] v9 v10
    v12 = Const32 <uint32> [-2147483648]
    v13 = Or32 <uint32> v11 v12 (afracP1[uint32])
    v16 = Const32 <int32> [0]
    v17 = Less32 <bool> v9 v16
    v26 = Const32 <int32> [-1]
    If v17 -> b3 b2
  b2: <- b1
    v31 = MakeResult <int64,mem> v8 v1
    Ret v31
  b3: <- b1
    v19 = Neg32 <int32> v9
    v21 = Add32 <int32> v26 v19 (sft[int32])
    v23 = Leq32 <bool> v16 v21
    If v23 -> b4 b5 (likely)
  b4: <- b3
    v27 = Rsh32Ux32 <uint32> [false] v13 v21 (ret[uint32])
    v28 = ZeroExt32to64 <int64> v27
    v29 = MakeResult <int64,mem> v28 v1
    Ret v29
  b5: <- b3
    v25 = StaticLECall <mem> {AuxCall{runtime.panicshift}} v1
    Exit v25

at b5 it has a proof tree as follows:

           extra
          /    \
        32      9
       /  \\  //  \
 {empty}   8/16   {empty}

It fails to utilize the intermediary values.

I added have a small patch that enables the compiler to add more data to the prove DAG, its very specific ATM but could be more general:
https://github.com/JAicewizard/go/tree/add_backwards_checks_ssa.

This adds checks in the SSA pass where if the first argument to a conditional is a 0 constant, and the second happens to be inverted from another variable that happened to already be compared to 0, it adds the knowlege that the inverted value it is </> than 0 (and only for 32 bits, and when its the 3rd block :P). This could be added for more constants, and in more advanced transformations.

I dont know how to actually run benchmarks in the gc, I get a bunch of "cannot access internal package" errors.

[gopherbot]

Change https://go.dev/cl/599096 mentions this issue: cmd/compile: rewrite the constant parts of the prove pass