[llvm] [AMDGPU] Account for existing SDWA selections (PR #123221)

[Frederik Harwath via llvm-commits]

================
@@ -88,6 +88,42 @@ class SIPeepholeSDWALegacy : public MachineFunctionPass {
 
 using namespace AMDGPU::SDWA;
 
+/// Check that the SDWA selections \p Sel and \p OperandSel
+/// are suitable for being combined by combineSdwaSel.
+bool compatibleSelections(SdwaSel Sel, SdwaSel OperandSel) {
+  return Sel == SdwaSel::DWORD || OperandSel == Sel ||
+         (Sel != SdwaSel::WORD_1 && Sel != SdwaSel::BYTE_2 &&
+          Sel != SdwaSel::BYTE_3 &&
+          (OperandSel == SdwaSel::WORD_0 || OperandSel == SdwaSel::WORD_1));
+}
+
+/// Combine an SDWA instruction's existing SDWA selection \p Sel with
+/// the SDWA selection \p OpSel of its operand which must be
+/// compatible.
+/// For example, if we have Sel = BYTE_0 Sel and OperandSel = WORD_1:
+///     BYTE_0 Sel (WORD_1 Sel (%X)) -> BYTE_2 Sel (%X)
+SdwaSel combineSdwaSel(SdwaSel Sel, SdwaSel OperandSel) {
----------------
frederik-h wrote:

I have now changed it to `static` and moved it out of the anonymous namespace as per the LLVM coding standards.

https://github.com/llvm/llvm-project/pull/123221