cmd/compile: optimize float32/float64 conversions used with `math` package

[dominikh]

Most functions in the math package operate on float64s. For users who work with float32s this leads to a lot of code like float32(math.Floor(float64(v)) where var v float32 (where Floor could be any one of the many math functions). On amd64 with GOAMD64=v2, that example compiles to

CVTSS2SD  X0, X1
ROUNDSD   $1, X1, X1
CVTSD2SS  X1, X1

which is the literal interpretation of the code. I suggest that this sequence gets optimized to a use of ROUNDSS instead. The example uses ROUNDSD, but of course I'd like this to work for all relevant instructions.

[mrkfrmn]

@golang/compiler

[adonovan]

Just curious: what are you using float32 for?

[dominikh]

@adonovan 2D vector graphics renderer. Among other things, colors are represented as [4]float32.