AMDGPU: support folding 64-bit immediates

[jayfoad]

SIFoldOperands should support folding 64-bit immediates into literal operands.

Before GFX10 this was rarely a problem because most 64-bit instructions are VOP3 only and VOP3 did not allow a literal operand. One exception is V_FMAC_F64 on gfx90a+. Here's an example where we generate v_fmac_f64_e32 v[0:1], s[4:5], v[2:3] which could be folded to v_fmac_f64_e32 v[0:1], 0x41200000, v[2:3]: https://godbolt.org/z/EMTjfj6jb

On GFX10+ VOP3 can take a literal operand, so there are lots of 64-bit instructions that could be folded.

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

SIFoldOperands should support folding 64-bit immediates into literal operands.

Before GFX10 this was rarely a problem because most 64-bit instructions are VOP3 only and VOP3 did not allow a literal operand. One exception is V_FMAC_F64 on gfx90a+. Here's an example where we generate v_fmac_f64_e32 v[0:1], s[4:5], v[2:3] which could be folded to v_fmac_f64_e32 v[0:1], 0x41200000, v[2:3]: https://godbolt.org/z/EMTjfj6jb

On GFX10+ VOP3 can take a literal operand, so there are lots of 64-bit instructions that could be folded.

[rampitec]

I can see 3 parts of the problem:

1. We are only storing high 32 bits of an fp64 literal in the operand. If we continue that way it creates 3 sorts of problems: 1) every time we adding such immediate to an instruction we need to modify its value 2) in case if we ever unfold it from an instruction we need to convert it back, or its correct value will be lost 3) in case if we ever do any constant folding we need to do the same conversion. An obvious fix for that is to keep full fp64 value in an operand and instead adjust code emitter to handle it.
2. Folding itself, this shall be fairly trivial if we solve the next one.
3. We need something to fold, and this may be problematic. It should start with a move immediate selection, but we do not have 64-bit immediate moves with non-zero high 32 bits. We could use S_MOV_B64_IMM_PSEUDO/V_MOV_B64_PSEUDO instead, although it may pessimize resulting code if a literal is not folded. One solution could be to only select it this way if a value is used in an fp64 context later which may be tricky.

[rampitec]

1. is done: https://reviews.llvm.org/rGab6c3d50345f