[clang] [CIR] Add inline function attributes (PR #162866)

[Andy Kaylor via cfe-commits]

================
@@ -1885,6 +1885,87 @@ void CIRGenModule::setFunctionAttributes(GlobalDecl globalDecl,
   }
 }
 
+void CIRGenModule::setCIRFunctionAttributesForDefinition(
+    const clang::FunctionDecl *decl, cir::FuncOp f) {
+  assert(!cir::MissingFeatures::opFuncUnwindTablesAttr());
+  assert(!cir::MissingFeatures::stackProtector());
+
+  auto existingInlineKind = f.getInlineKind();
+  bool isNoInline =
+      existingInlineKind && *existingInlineKind == cir::InlineKind::NoInline;
+  bool isAlwaysInline = existingInlineKind &&
+                        *existingInlineKind == cir::InlineKind::AlwaysInline;
+
+  if (!decl) {
+    assert(!cir::MissingFeatures::hlsl());
+
+    if (!isAlwaysInline &&
+        codeGenOpts.getInlining() == CodeGenOptions::OnlyAlwaysInlining) {
+      // If we don't have a declaration to control inlining, the function isn't
+      // explicitly marked as alwaysinline for semantic reasons, and inlining is
+      // disabled, mark the function as noinline.
+      f.setInlineKindAttr(
+          cir::InlineAttr::get(&getMLIRContext(), cir::InlineKind::NoInline));
+    }
+
+    return;
+  }
+
+  assert(!cir::MissingFeatures::opFuncArmStreamingAttr());
+  assert(!cir::MissingFeatures::opFuncArmNewAttr());
+  assert(!cir::MissingFeatures::opFuncOptNoneAttr());
+  assert(!cir::MissingFeatures::opFuncMinSizeAttr());
+  assert(!cir::MissingFeatures::opFuncNakedAttr());
+  assert(!cir::MissingFeatures::opFuncNoDuplicateAttr());
+  assert(!cir::MissingFeatures::hlsl());
+
+  // Handle inline attributes
+  if (decl->hasAttr<NoInlineAttr>() && !isAlwaysInline) {
+    // Add noinline if the function isn't always_inline.
+    f.setInlineKindAttr(
+        cir::InlineAttr::get(&getMLIRContext(), cir::InlineKind::NoInline));
+  } else if (decl->hasAttr<AlwaysInlineAttr>() && !isNoInline) {
+    // (noinline wins over always_inline, and we can't specify both in IR)
----------------
andykaylor wrote:

This comment doesn't seem quite right. If the declaration has `NoInlineAttr` and the function previously had the `cir::InlineKind::AlwaysInline` attribute, the case above does not set `noinline`. So it seems that the existing inline kind attribute wins (unless it's 'InlineHint').

https://github.com/llvm/llvm-project/pull/162866