[clang] [clang] Better bitfield access units (PR #65742)

[John McCall via cfe-commits]

================
@@ -442,79 +455,235 @@ CGRecordLowering::accumulateBitFields(RecordDecl::field_iterator Field,
     return;
   }
 
-  // Check if OffsetInRecord (the size in bits of the current run) is better
-  // as a single field run. When OffsetInRecord has legal integer width, and
-  // its bitfield offset is naturally aligned, it is better to make the
-  // bitfield a separate storage component so as it can be accessed directly
-  // with lower cost.
-  auto IsBetterAsSingleFieldRun = [&](uint64_t OffsetInRecord,
-                                      uint64_t StartBitOffset) {
-    if (!Types.getCodeGenOpts().FineGrainedBitfieldAccesses)
-      return false;
-    if (OffsetInRecord < 8 || !llvm::isPowerOf2_64(OffsetInRecord) ||
-        !DataLayout.fitsInLegalInteger(OffsetInRecord))
-      return false;
-    // Make sure StartBitOffset is naturally aligned if it is treated as an
-    // IType integer.
-    if (StartBitOffset %
-            Context.toBits(getAlignment(getIntNType(OffsetInRecord))) !=
-        0)
-      return false;
-    return true;
+  // The SysV ABI can overlap bitfield storage units with both other bitfield
+  // storage units /and/ other non-bitfield data members. Such overlap, in the
+  // absence of packing, is always complete -- one storage unit is entirely
+  // within another. However, llvm cannot represent that -- it's structures are
+  // entirely flat. We place bitfields in 'access units', which are similar to
+  // the SysV storage units, but a clang-specific concept.
+
+  // It can be advantageous to concatenate two adjacent access units, if the
+  // concenation can be read or written in a single instruction.
+
+  // We do two passes.
+
+  // a) allocate bitfields into the smallest access units they can
+  // fit. This results in a set of integral-typed access units.
+
+  // b) concatentate mergeable access units. This applies the
+  // above-mentioned optimization, and in general, requires lookahead
+  // to know the next access unit -- not merely the next bitfield.
----------------
rjmccall wrote:

As far as I can tell, this doesn't really require doing two passes; it's still an online algorithm, just over first-phase access units instead of individual bit-fields.  You could take your existing code to build up those first-phase units and then just call into your merge-or-emit logic whenever you finish one instead of adding it to a vector.

https://github.com/llvm/llvm-project/pull/65742