uint256: minor improvement for Mul

[AaronChen0]

Minor tweek of the code. Not sure where the boost comes from. But it does improve around 10% in my two x86-64 machines.

godbolt links:
old: https://go.godbolt.org/z/z8vqhG9r7
new: https://go.godbolt.org/z/M17Ycfzxx

Test

go test ./...

ok  	github.com/holiman/uint256	1.380s

Benchmark

goos: linux
goarch: amd64
pkg: github.com/holiman/uint256
cpu: AMD Ryzen 7 7735H with Radeon Graphics         
                      │     old     │                 new                 │
                      │   sec/op    │   sec/op     vs base                │
Mul/single/uint256-16   4.239n ± 2%   3.725n ± 1%  -12.14% (p=0.000 n=10)
Mul/single/big-16       37.05n ± 1%   36.96n ± 2%        ~ (p=0.403 n=10)
geomean                 12.53n        11.73n        -6.38%

goos: linux
goarch: amd64
pkg: github.com/holiman/uint256
cpu: Intel Core Processor (Broadwell, no TSX, IBRS)
                   │     old      │                new                 │
                   │    sec/op    │   sec/op     vs base               │
Mul/single/uint256   10.125n ± 4%   9.346n ± 2%  -7.69% (p=0.000 n=10)
Mul/single/big        79.24n ± 3%   79.11n ± 5%       ~ (p=0.853 n=10)
geomean               28.32n        27.19n       -4.00%

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 100.00%. Comparing base (5ecf78c) to head (3cbed10).

Additional details and impacted files

@@            Coverage Diff            @@
##            master      #167   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files            5         5           
  Lines         1628      1626    -2     
=========================================
- Hits          1628      1626    -2     

[holiman]

Interesting!

name                  old time/op  new time/op  delta
Mul/single/uint256-8  15.1ns ± 0%  10.6ns ±35%  -29.68%  (p=0.000 n=8+10)

@chfast lgty?

[AaronChen0]

New: https://go.godbolt.org/z/5E8TahK9j
Old: https://go.godbolt.org/z/nY31bsxdb

With Go 1.22.1, the new code has 7 less MOVQ instructions. Other instructions are the same amount.
Possibly, this is the reason.

[chfast]

LGTM as this only moves the values to different variables and changes the order of operations.

[chfast]

New: https://go.godbolt.org/z/5E8TahK9j Old: https://go.godbolt.org/z/nY31bsxdb

With Go 1.22.1, the new code has 7 less MOVQ instructions. Other instructions are the same amount. Possibly, this is the reason.

godbolt has also the "diff view": https://go.godbolt.org/z/ddcK6TzKK

It looks the new code spills less values from registers to stack. Good change.