[clang] [llvm] [RISCV] Add Qualcomm uC Xqciac (Load-Store Adress calculation) extension (PR #121037)

[Sam Elliott via llvm-commits]

================
@@ -184,6 +191,37 @@ let hasSideEffects = 0, mayLoad = 0, mayStore = 0 in {
 } // hasSideEffects = 0, mayLoad = 0, mayStore = 0
 } // Predicates = [HasVendorXqcia, IsRV32], DecoderNamespace = "Xqcia"
 
+let Predicates = [HasVendorXqciac, IsRV32], DecoderNamespace = "Xqciac" in {
+let hasSideEffects = 0, mayLoad = 0, mayStore = 0 in {
+  def QC_C_MULADDI : RVInst16CL<0b001, 0b10, (outs GPRC:$rd_wb),
+                               (ins GPRC:$rd, GPRC:$rs1, uimm5:$uimm),
+                               "qc.c.muladdi", "$rd, $rs1, $uimm"> {
+    let Constraints = "$rd = $rd_wb";
+    bits<5> uimm;
+
+    let Inst{12-10} = uimm{3-1};
+    let Inst{6} = uimm{0};
+    let Inst{5} = uimm{4};
+  }
----------------
lenary wrote:

Unlike the main RISC-V ISA, the Xqci architects did not want to split instructions into (sub-)extensions by size (as you can see, they have been split by functionality). To make the ABI much easier, they agreed that all of the sub-extensions with any 16- or 48-bit instructions would require Zca, as the compiler will need to be able to insert `c.nop` when aligning etc etc.

Not splitting instructions into extensions by size does mean that existing assumptions used by `CompressPat` no longer quite hold. At the moment, it is enough to do `.option norvc` (or an equivalent with `.option arch`) for the `C` extension to prevent compression, but you cannot disable `qc.c.muladdi` without disabling `qc.muladdi`, so we cannot just use architecture features to gate compression patterns (There are even more issues around our 48-bit instructions, which I don't want to write out here). As I said, I have been thinking about an approach for this, but I haven't had time to work through the patch, test it, and present it for review. This is why none of our `Xqci*` patches so far have had `CompressPat`s, and we do not intend to add them for the moment. Most of these issues are reasonably orthogonal to the issue that @svs-quic mentioned in the actual tablegen code for `CompressPat`s.

https://github.com/llvm/llvm-project/pull/121037