[llvm] AMDGPU: Fix packed 16-bit inline constants (PR #76522)

[Nicolai Hähnle via llvm-commits]

================
@@ -234,51 +232,143 @@ bool SIFoldOperands::tryFoldImmWithOpSel(FoldCandidate &Fold) const {
   MachineOperand &Old = MI->getOperand(Fold.UseOpNo);
   unsigned Opcode = MI->getOpcode();
   int OpNo = MI->getOperandNo(&Old);
+  uint8_t OpType = TII->get(Opcode).operands()[OpNo].OperandType;
+
+  // If the literal can be inlined as-is, apply it and short-circuit the
+  // tests below. The main motivation for this is to avoid unintuitive
+  // uses of opsel.
+  if (AMDGPU::isInlinableLiteralV216(Fold.ImmToFold, OpType)) {
+    Old.ChangeToImmediate(Fold.ImmToFold);
+    return true;
+  }
 
-  // Set op_sel/op_sel_hi on this operand or bail out if op_sel is
-  // already set.
+  // Refer to op_sel/op_sel_hi and check if we can change the immediate and
+  // op_sel in a way that allows an inline constant.
   int ModIdx = -1;
-  if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0))
+  unsigned SrcIdx = ~0;
+  if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0)) {
     ModIdx = AMDGPU::OpName::src0_modifiers;
-  else if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1))
+    SrcIdx = 0;
+  } else if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1)) {
     ModIdx = AMDGPU::OpName::src1_modifiers;
-  else if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2))
+    SrcIdx = 1;
+  } else if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2)) {
     ModIdx = AMDGPU::OpName::src2_modifiers;
+    SrcIdx = 2;
+  }
   assert(ModIdx != -1);
   ModIdx = AMDGPU::getNamedOperandIdx(Opcode, ModIdx);
   MachineOperand &Mod = MI->getOperand(ModIdx);
-  unsigned Val = Mod.getImm();
-  if ((Val & SISrcMods::OP_SEL_0) || !(Val & SISrcMods::OP_SEL_1))
+  unsigned ModVal = Mod.getImm();
+
+  uint16_t ImmLo = static_cast<uint16_t>(
----------------
nhaehnle wrote:

I actually started out with having some of these changes separate, but there was a rather silly amount of test churn, and there were temporary code quality regressions that made this annoying. If you feel strongly about it I can try to separate the changes out again, but I couldn't figure out how to do it without temporary code quality regressions.

https://github.com/llvm/llvm-project/pull/76522