crypto/elliptic: use go:embed for the precomputed p256 table
go.dev/cl/339591 changed the code generation to use a constant string,
so that the ~88KiB table can be marked read-only.
The compiled code became a lot better, but unfortunately,
the generated Go source became significantly more inefficient.
The numbers below compare "gofmt -l" and "go tool compile" of said file,
where "old" is the file as of Go 1.17, and "new" as of master in 2022/01/19:
name old time/op new time/op delta
Gofmt 22.8ms ± 6% 898.5ms ± 3% +3837.32% (p=0.000 n=8+8)
GoToolCompile 26.9ms ± 2% 371.1ms ± 2% +1278.36% (p=0.000 n=7+8)
name old user-time/op new user-time/op delta
Gofmt 25.7ms ±65% 897.1ms ± 3% +3383.86% (p=0.000 n=8+8)
GoToolCompile 35.1ms ±26% 367.2ms ± 3% +945.80% (p=0.000 n=8+8)
name old sys-time/op new sys-time/op delta
Gofmt 6.42ms ±276% 7.23ms ±38% ~ (p=0.412 n=8+6)
GoToolCompile 9.20ms ±100% 13.90ms ±53% ~ (p=0.105 n=8+8)
name old peak-RSS-bytes new peak-RSS-bytes delta
Gofmt 9.11MB ± 7% 22.79MB ± 1% +150.23% (p=0.000 n=8+8)
GoToolCompile 25.1MB ± 2% 68.6MB ± 2% +173.57% (p=0.000 n=8+8)
"+" operators are binary expressions at the syntax tree level,
which are represented by packages like go/ast as roughly:
struct {
X Expr
Op Token
Y Expr
}
Since each node is a pointer, chains of "+" operators act like linked lists.
The generated code has about 14k lines, and 8 "+" operators per line,
meaning that we end up with a linked list with over 11k elements.
This explains the slow-down in gofmt; the printer must walk said list,
and it does so more than once to work out how to format it.
It seems like the compiler is similarly affected by the huge length.
To remedy the effect of the linked list, use go:embed instead.
This results in the same string variable with the binary table,
but it greatly reduces the amount of syntax and its cost above.
We still keep the generator around, but modified so it produces the
binary file to be embedded rather than a large Go file.
Finally, we update go/build/deps_test.go to allow crypto/elliptic to
depend on embed; it's a tiny package and crypto/elliptic was already
manually embedding assets via code generation.
The change to deps_test.go was briefly discussed in the issue below.