[llvm] [AMDGPU] Guard more intrinsics with target features (PR #203956)

[Shilei Tian via llvm-commits]

================
@@ -475,6 +476,7 @@ def int_amdgcn_fmul_legacy : ClangBuiltin<"__builtin_amdgcn_fmul_legacy">,
 // intended for use on subtargets that have the v_fma_legacy_f32 and/or
 // v_fmac_legacy_f32 instructions. (Note that v_fma_legacy_f16 is unrelated and
 // has a completely different kind of legacy behaviour.)
+let TargetFeatures = "gfx10-3-insts" in
----------------
shiltian wrote:

I've added a new target feature for legacy FMA, but hasCompressedExport is trickier. It effectively means "as long as the target doesn't have gfx11-insts, it's supported." If we're going to represent that as a target feature, we'd need to add it to every target, including future ones. That comes with a fairly high maintenance cost.

https://github.com/llvm/llvm-project/pull/203956