MTLVertexCache constantly creates/delete MTLBuffer objects to pass small amount of vertices. It noticeable degrades performance of AA rendering and font rendering where we operate with lot's of tiles. This logic can be replaced with just passing raw bytes to Vertex buffer.

Here is some performance measurements with improved RenderPerfTest

$java -Dsun.java2d.metal=true -jar dist/RenderPerfTest.jar testFlatBubblesAA testFlatQuadAA testTextBubblesGray testTextBubblesNoAA testTextBubblesLCD testWhiteTextBubblesGray testWhiteTextBubblesNoAA testWhiteTextBubblesLCD

Without fix:

FlatOvalAA : 2.13 FPS
FlatQuadAA : 0.45 FPS
TextGray : 3.44 FPS
TextNoAA : 4.15 FPS
TextLCD : 4.23 FPS
WhiteTextGray : 4.19 FPS
WhiteTextNoAA : 4.22 FPS
WhiteTextLCD : 4.21 FPS

With the fix applied:

FlatOvalAA : 3.37 FPS
FlatQuadAA : 0.85 FPS
TextGray : 5.93 FPS
TextNoAA : 5.83 FPS
TextLCD : 5.94 FPS
WhiteTextGray : 5.48 FPS
WhiteTextNoAA : 5.80 FPS
WhiteTextLCD : 6.07 FPS