[llvm] [AMDGPU] Create an AMDGPUIfConverter pass (PR #106415)

[Juan Manuel Martinez Caamaño via llvm-commits]

================
@@ -199,11 +199,6 @@ static VGPRRegisterRegAlloc fastRegAllocVGPR(
   "fast", "fast register allocator", createFastVGPRRegisterAllocator);
 } // anonymous namespace
 
-static cl::opt<bool>
-EnableEarlyIfConversion("amdgpu-early-ifcvt", cl::Hidden,
-                        cl::desc("Run early if-conversion"),
-                        cl::init(false));
----------------
jmmartinez wrote:

I have one main concern about doing this at IR level: How to handle instructions accessing memory.
I guess it could be done through masked load/store operations. But what if we have some other built-in?

I'm also not sure about how to implement the cost model in the LLVM-IR, but I guess it could be possible.

https://github.com/llvm/llvm-project/pull/106415