Improve compile time performance.

[csyonghe]

1. Add caching for implicit conversion overload resolution results.
2. Handle invalidation of type checking cache upon extension registration more gracefully.
3. Add a setting to simplifyIR to disable redundancyRemoval pass (still enabled by default).

[jkwak-work]

These comments are little confusing.
I assume that they are copied from another file?
If yes, you may need to modify these.

[csyonghe]

Good catch, thanks!

[jkwak-work]

Looks good

[jkwak-work]

I have some timing numbers here.

Without this change, slangc takes 47.19s ~ 48.18s for compiling 40 TSR shaders.
With this change, slangc takes slightly less time, 46.19s ~ 47.71s for the same TSR shaders.

Without this change, FXC took 90.28s ~ 91.69s for compiling 6 TSR shaders.
With this change, FXC took little longer, 94.42s ~ 95.10s for the same TSR shaders.

I noticed that with this changes, the number of HLSL lines generated from Slangc is about 3% less.
I am not sure if it is expected.

[csyonghe]

What about the perf numbers from dxc, and does shader perf change?

[jkwak-work]

I am sorry. I thought I was using FXC but it was DXC.
All of the data I shared here are based on DXC.

[jkwak-work]

Code looks good

[csyonghe]

Given your measurements, I am not sure if we should be merging these changes, since it doesn't seem to be a net win.

[csyonghe]

Update: the changes here resulted 20% speedup in frontend when processing typical shader workloads, so I am going to merge it.