math/big: broken mips and mips64 builders

[randall77]

After https://go-review.googlesource.com/c/go/+/664938 the linux/mips and linux/mips64 builders are failing, with errors like

../../go/src/internal/runtime/math/math.go:9:20: ^uintptr(0) (constant 8589934591 of type uintptr) overflows uintptr

I suspect something is wrong with the mips & mips64 assembly. That is causing the compiler's constant computations, which use math/big, to be wrong.

Cross-compiling mips binaries from linux/amd64 works fine. I think the toolchain itself needs to be running on mips to demonstrate the problem.

@rsc

[adonovan]

../../go/src/internal/runtime/math/math.go:9:20: ^uintptr(0) (constant 8589934591 of type uintptr) overflows uintptr

This value is 1fff... instead of ffff..., so it looks like ^uintptr(0) might have been miscomputed. Assuming the compiler calls constant.UnaryOp(XOR), then the relevant expressions are:

• z.Not(big.NewInt(int64(y))), which calls Int.Not. Int.Not calls nat.add -> addVV and addVW (we can ignore the negated case); and
• z.AndNot(z, newInt().Lsh(big.NewInt(-1), prec)), which calls Int.AndNot (which is pure Go) and nat.{sub,add}; and Int.Lsh, which calls nat.lsh -> lshVU

So the assembly routines I would audit first are addVV, addVW, their sub counterparts, and lshVU. Or just write a test for ^uintptr(0) and run it on the mips builder.

[dmitshur]

CC @golang/mips.

[gopherbot]

Change https://go.dev/cl/667816 mentions this issue: math/big: fix incorrect register allocation for mipsx/mips64x