[PATCH] D108407: [CodeGen][WIP] Avoid generating Record layouts for pointee types

[Raphael Isemann via Phabricator via llvm-commits]

teemperor created this revision.
teemperor added reviewers: dblaikie, rjmccall, rsmith, v.g.vassilev.
teemperor added a project: clang.
Herald added a subscriber: dexonsmith.
teemperor requested review of this revision.
Herald added subscribers: llvm-commits, cfe-commits.
Herald added a project: LLVM.

This is a WIP patch that tries to avoid creating a RecordLayout in Clang and instead just emit an opaque structure type
as if we only had a forward declarations. The main motivation for this patch is actually just supporting a use case in LLDB
where laying out types can be very expensive as it usually triggers parsing of debug information.

The changes in this patch can be summarized as:

- `CodeGenTypes::ConvertRecordDeclType` (and related funcs) have a new parameter that tells us if we need the definition. It's currently only set to false for Clang pointer types.
- There are a few new places where I added (temporary) calls to `ConvertTypeForMem()` on some pointee types. The reason is that the code after is usually creating GEP instructions where we need a non-opaque source type. We can't do this automatically from the GEP factory methods as they would need to know the clang::Type to automatically do this (and they only have the llvm::Type that can't be mapped back to a clang::Type from what I can see, but that might be incorrect).
- A few test that needed to be adjusted as they relied on e.g. `Foo *x` to be enough to force `Foo` to be laid out/emitted.

There are still about a dozen more tests failing but from what I can see they all just need to be adjusted to force specific types to be emitted. I'll fix those up once there is consensus that this patch is going in the right direction.

Some benchmarks: I did a stage2 build of LLVM+Clang with my patch and those are the stats:

  current ToT Clang:
  2232421 - total amount of struct types created
    94911 - of which are opaque struct types
  
  with this patch:
  1715074 - total amount of struct types created (-23%)
   173127 - of which are opaque struct types (+82%)

I built a part of Clang (the last 300 source files in the compile_commands.json) and the average time on my 64 core machine changes like this (as per hyperfine):

  Benchmark #1: parallel --progress -j63 -a ToT-clang
    Time (mean ± σ):     27.703 s ±  0.168 s    [User: 1434.619 s, System: 66.687 s]
    Range (min … max):   27.459 s … 27.891 s    10 runs
   
  Benchmark #2: parallel --progress -j63 -a with-patch
    Time (mean ± σ):     27.439 s ±  0.111 s    [User: 1427.739 s, System: 66.220 s]
    Range (min … max):   27.300 s … 27.625 s    10 runs
   
  Summary
    'parallel --progress -j63 -a with-patch' ran
      1.01 ± 0.01 times faster than 'parallel --progress -j63 -a ToT-clang'


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D108407

Files:
  clang/include/clang/AST/Type.h
  clang/lib/CodeGen/CGExpr.cpp
  clang/lib/CodeGen/CGExprCXX.cpp
  clang/lib/CodeGen/CGExprScalar.cpp
  clang/lib/CodeGen/CGOpenMPRuntime.cpp
  clang/lib/CodeGen/CodeGenTypes.cpp
  clang/lib/CodeGen/CodeGenTypes.h
  clang/test/CodeGen/c11atomics.c
  clang/test/CodeGenCXX/class-layout.cpp
  clang/test/CodeGenCXX/pr18962.cpp
  clang/test/CodeGenCXX/warn-padded-packed.cpp
  llvm/include/llvm/IR/Instructions.h

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D108407.367590.patch
Type: text/x-patch
Size: 12605 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210819/5340324a/attachment.bin>