[llvm] [GlobalISel][AArch64][AMDGPU] Expand FPOWI into series of multiplication (PR #95217)

[Matt Arsenault via llvm-commits]

================
@@ -7356,6 +7356,38 @@ void CombinerHelper::applyBuildFnMO(const MachineOperand &MO,
   Root->eraseFromParent();
 }
 
+void CombinerHelper::applyExpandFPowI(MachineInstr &MI) {
+  auto [Dst, Base, Exp] = MI.getFirst3Regs();
+  LLT Ty = MRI.getType(Dst);
+
+  std::optional<int64_t> ConstantExpValue = getIConstantVRegSExtVal(Exp, MRI);
+  assert(ConstantExpValue && "matched non-const FPOWI?");
+
+  int64_t OriginalExprVal = *ConstantExpValue;
+  int64_t ExpVal = OriginalExprVal;
+
+  if (ExpVal == 0) {
+    Builder.buildFConstant(Dst, 1.0);
+    MI.removeFromParent();
+    return;
+  }
+
+  if (ExpVal < 0)
+    ExpVal = -ExpVal;
+
----------------
arsenm wrote:

Can you preserve the comment from the DAG version about "We use the simple binary decomposition method..."

https://github.com/llvm/llvm-project/pull/95217