GLSL FS backend: Extremely poor code generation for loops that reference constant arrays!

[chrisforbes]

Things like this generate a horrendous mess currently. The nonconstant index is lowered to a massive conditional tree for all possible indexes.

If the loop were just unrolled instead, or even better, induction variables were allowed as array indexes, things would be quite a lot better.

const float xs[16] = { ... }

float z = 0.0f;
for(int i=0; i<16; i++)
    z += xs[i];

[chrisforbes]

Here's a real example, from KSP. It's a 60-tap poisson disc filter, butchered horribly.

http://sprunge.us/cSKL

[chrisforbes]

In D3D you'd write that as a cbuffer -- which behaves more or less like our UBO. (In fact, the NV paper on this shadow technique provides a sample HLSL shader using a cbuffer for exactly this purpose.

Suggests that it's both acceptable (from a perf perspective) and possible to shunt the block of constants as a whole, into a block of uniforms.

[chrisforbes]

Looks like the lowering pass can be tweaked slightly to avoid lowering references into constant arrays. Unsure what else would have to happen -- probably shipping the array out to pull constants [as we do with uniforms] needs to be handled nicely in FS too.