[llvm-branch-commits] MC: Refactor FT_Align fragments when linker relaxation is enabled (PR #149465)

[Fangrui Song via llvm-branch-commits]

================
@@ -230,22 +230,24 @@ uint64_t MCAssembler::computeFragmentSize(const MCFragment &F) const {
   case MCFragment::FT_Align: {
     unsigned Offset = F.Offset + F.getFixedSize();
     unsigned Size = offsetToAlignment(Offset, F.getAlignment());
-
-    // Insert extra Nops for code alignment if the target define
-    // shouldInsertExtraNopBytesForCodeAlign target hook.
-    if (F.getParent()->useCodeAlign() && F.hasAlignEmitNops() &&
-        getBackend().shouldInsertExtraNopBytesForCodeAlign(F, Size))
-      return F.getFixedSize() + Size;
-
-    // If we are padding with nops, force the padding to be larger than the
-    // minimum nop size.
-    if (Size > 0 && F.hasAlignEmitNops()) {
-      while (Size % getBackend().getMinimumNopSize())
-        Size += F.getAlignment().value();
+    auto &Frag = const_cast<MCFragment &>(F);
+    // In the nops mode, RISC-V style linker relaxation might adjust the size
+    // and add a fixup, even if `Size` is originally 0.
+    bool AlignFixup = false;
+    if (F.hasAlignEmitNops()) {
+      AlignFixup = getBackend().relaxAlign(Frag, Size);
----------------
MaskRay wrote:

Thanks for the feedback! There is currently an oddity in the layout algorithm

* https://github.com/llvm/llvm-project/blob/a9147e64aa751caaa106953fded2d0f7223bb167/llvm/lib/MC/MCAssembler.cpp#L676 This `layoutSection` assigns offsets to fragments, including FT_Align (using `computeFragmentSize`)
* `relaxOnce` checks FT_Relaxable (which should probably be renamed to FT_Insn) but not FT_Align

Future `computeFragmentSize` calls should not modify the member variable. I should perhaps move this FT_Align handling to `layoutSection` so that `computeFragmentSize` can simply return `getSize()`.

https://github.com/llvm/llvm-project/pull/149465