[TBAA] Handle bitfields when generating !tbaa.struct metadata. #82922

fhahn · 2024-02-25T13:30:45Z

At the moment, clang generates what I believe are incorrect !tbaa.struct fields for named bitfields. At the moment, the base type size is used for named bifields (e.g. sizeof(int)) instead of the bifield width per field. This results in overalpping fields in !tbaa.struct metadata.

This causes incorrect results when extracting individual copied fields from !tbaa.struct as in added in dc85719.

This patch fixes that by skipping by combining adjacent named bitfields in fields with correct size.

If this understanding is correct, I can also extend the verifier to check that !tbaa.struct fields aren't overlapping.

Fixes #82586

At the moment, clang generates what I believe are incorrect !tbaa.struct fields for named bitfields. At the moment, the base type size is used for named bifields (e.g. sizeof(int)) instead of the bifield width per field. This results in overalpping fields in !tbaa.struct metadata. This causes incorrect results when extracting individual copied fields from !tbaa.struct as in added in dc85719. This patch fixes that by skipping all bitfields, not only unnamed ones (note that CollectFields has a TODO to support bitfields). As bitfields specify their widths in bits, while !tbaa metadata uses bytes for sizes and offsets, I don't think we would be able to generate correct metadata for them in general. If this understanding is correct, I can also extend the verifier to check that !tbaa.struct fields aren't overlapping. Fixes llvm#82586

llvmbot · 2024-02-25T13:31:17Z

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-clang-codegen

Author: Florian Hahn (fhahn)

Changes

At the moment, clang generates what I believe are incorrect !tbaa.struct fields for named bitfields. At the moment, the base type size is used for named bifields (e.g. sizeof(int)) instead of the bifield width per field. This results in overalpping fields in !tbaa.struct metadata.

This causes incorrect results when extracting individual copied fields from !tbaa.struct as in added in dc85719.

This patch fixes that by skipping all bitfields, not only unnamed ones (note that CollectFields has a TODO to support bitfields). As bitfields specify their widths in bits, while !tbaa metadata uses bytes for sizes and offsets, I don't think we would be able to generate correct metadata for them in general.

If this understanding is correct, I can also extend the verifier to check that !tbaa.struct fields aren't overlapping.

Fixes #82586

Full diff: https://github.com/llvm/llvm-project/pull/82922.diff

2 Files Affected:

(modified) clang/lib/CodeGen/CodeGenTBAA.cpp (+1-1)
(modified) clang/test/CodeGen/tbaa-struct.cpp (+1-1)

diff --git a/clang/lib/CodeGen/CodeGenTBAA.cpp b/clang/lib/CodeGen/CodeGenTBAA.cpp
index dc288bc3f6157a..43a1aee3d73823 100644
--- a/clang/lib/CodeGen/CodeGenTBAA.cpp
+++ b/clang/lib/CodeGen/CodeGenTBAA.cpp
@@ -298,7 +298,7 @@ CodeGenTBAA::CollectFields(uint64_t BaseOffset,
     unsigned idx = 0;
     for (RecordDecl::field_iterator i = RD->field_begin(),
          e = RD->field_end(); i != e; ++i, ++idx) {
-      if ((*i)->isZeroSize(Context) || (*i)->isUnnamedBitfield())
+      if ((*i)->isZeroSize(Context) || (*i)->isBitField())
         continue;
       uint64_t Offset = BaseOffset +
                         Layout.getFieldOffset(idx) / Context.getCharWidth();
diff --git a/clang/test/CodeGen/tbaa-struct.cpp b/clang/test/CodeGen/tbaa-struct.cpp
index ff5521fcf3f604..17c9d6bf6a7260 100644
--- a/clang/test/CodeGen/tbaa-struct.cpp
+++ b/clang/test/CodeGen/tbaa-struct.cpp
@@ -130,7 +130,7 @@ void copy8(NamedBitfields *a1, NamedBitfields *a2) {
 // CHECK-OLD: [[TS3]] = !{i64 0, i64 8, !{{.*}}, i64 0, i64 2, !{{.*}}, i64 4, i64 8, !{{.*}}}
 // CHECK-OLD: [[TS4]] = !{i64 0, i64 1, [[TAG_CHAR]], i64 1, i64 1, [[TAG_CHAR]], i64 2, i64 1, [[TAG_CHAR]]}
 // CHECK-OLD: [[TS5]] = !{i64 0, i64 1, [[TAG_CHAR]], i64 4, i64 1, [[TAG_CHAR]], i64 5, i64 1, [[TAG_CHAR]]}
-// CHECK-OLD: [[TS6]] = !{i64 0, i64 4, [[TAG_INT]], i64 1, i64 4, [[TAG_INT]], i64 2, i64 1, [[TAG_CHAR]], i64 8, i64 8, [[TAG_DOUBLE:!.+]]}
+// CHECK-OLD: [[TS6]] = !{i64 2, i64 1, [[TAG_CHAR]], i64 8, i64 8, [[TAG_DOUBLE:!.+]]}
 // CHECK-OLD: [[TAG_DOUBLE]] = !{[[DOUBLE:!.+]], [[DOUBLE]], i64 0}
 // CHECK-OLD  [[DOUBLE]] = !{!"double", [[CHAR]], i64 0}

efriedma-quic · 2024-02-26T04:52:57Z

This seems like it messes up the metadata in a different way: we have no representation of the bitfield at all, so it looks like it's padding.

I think we want to treat multiple adjacent bitfields as a single "field". So NamedBitfields from the testcase would have three fields in the LLVM TBAA data: one "field" containing both bitfields, followed by a field for the char, followed by a field for the double. You should be able to use CGBitFieldInfo for this, I think?

…ed-bitfields

github-actions · 2024-02-26T10:22:49Z

✅ With the latest revision this PR passed the C/C++ code formatter.

Additional test for #82922.

…ed-bitfields

fhahn · 2024-02-26T13:02:28Z

This seems like it messes up the metadata in a different way: we have no representation of the bitfield at all, so it looks like it's padding.

I think we want to treat multiple adjacent bitfields as a single "field". So NamedBitfields from the testcase would have three fields in the LLVM TBAA data: one "field" containing both bitfields, followed by a field for the char, followed by a field for the double. You should be able to use CGBitFieldInfo for this, I think?

Thanks, updated the patch as suggested.

efriedma-quic · 2024-02-26T17:43:36Z

clang/lib/CodeGen/CodeGenTBAA.cpp

+      if ((*i)->isBitField() && !(*i)->isUnnamedBitfield()) {
+        unsigned CurrentBitFieldSize = 0;
+        unsigned CurrentBitFieldOffset = CGRL.getBitFieldInfo(*i).Offset;
+        while (i != e && (*i)->isBitField() && !(*i)->isUnnamedBitfield()) {


I don't think this handles non-zero-length unnamed bitfields correctly; they count as part of a series of bitfields.

Yes, test showing that added in 54cff50, thanks!

efriedma-quic · 2024-02-26T17:44:44Z

clang/lib/CodeGen/CodeGenTBAA.cpp

+          const CGBitFieldInfo &Info = CGRL.getBitFieldInfo(*i);
+          if (CurrentBitFieldSize + CurrentBitFieldOffset != Info.Offset)
+            break;
+          CurrentBitFieldSize += Info.Size;


Can we pull the initial offset and total length of the whole run bitfields out of the CGBitFieldInfo, instead of trying to add up the sizes of the individual fields? I think it should be StorageOffset/StorageSize.

Thanks, updated the patch to add a field for the first bitfield in the run using its StorageSize, and skip bitfields where the offset in the run isn't 0 (Offset).

…ed-bitfields

Extra tests with unnamed bitfields for #82922.

…ed-bitfields

efriedma-quic

Whitespace is weird in a few places; otherwise looks fine.

…ed-bitfields

fhahn · 2024-02-27T13:15:32Z

Whitespace is weird in a few places; otherwise looks fine.

Argh yes, should be fixed in the latest version. Still need to figure out how to make sure all changes are formatted once the branch contains merges (before the format checker complains and shows the commits)

efriedma-quic

LGTM

fhahn · 2024-02-27T20:11:21Z

Thanks!

Additional test for llvm#82922. (cherry-picked from f290c00)

Extra tests with unnamed bitfields for llvm#82922. (cherry-picked from 54cff50)

…82922) At the moment, clang generates what I believe are incorrect !tbaa.struct fields for named bitfields. At the moment, the base type size is used for named bifields (e.g. sizeof(int)) instead of the bifield width per field. This results in overalpping fields in !tbaa.struct metadata. This causes incorrect results when extracting individual copied fields from !tbaa.struct as in added in dc85719. This patch fixes that by skipping by combining adjacent bitfields in fields with correct sizes. Fixes llvm#82586 (cherry-picked from d2a9df2)

fhahn requested review from rjmccall, wsmoses, efriedma-quic and dobbelaj-snps February 25, 2024 13:30

llvmbot added clang Clang issues not falling into any other category clang:codegen IR generation bugs: mangling, exceptions, etc. labels Feb 25, 2024

efriedma-quic requested a review from urnathan February 26, 2024 04:40

fhahn added 2 commits February 26, 2024 07:42

Merge remote-tracking branch 'origin/main' into clang-tbaa-struct-nam…

da9a942

…ed-bitfields

WIP use CGBitFieldInfo.

32be3b7

fhahn added 2 commits February 26, 2024 11:53

Extra test.

98b7045

fixup use bitfield info

e4e4614

fhahn added a commit that referenced this pull request Feb 26, 2024

[TBAA] Add additional bitfield tests.

f290c00

Additional test for #82922.

fhahn added 2 commits February 26, 2024 12:36

Merge remote-tracking branch 'origin/main' into clang-tbaa-struct-nam…

8267594

…ed-bitfields

!fixup

a2dde72

fhahn changed the title ~~[TBAA] Skip all bitfields when generating !tbaa.struct metadata.~~ [TBAA] Handle bitfields when generating !tbaa.struct metadata. Feb 26, 2024

efriedma-quic reviewed Feb 26, 2024

View reviewed changes

fhahn added 2 commits February 26, 2024 22:08

Merge remote-tracking branch 'origin/main' into clang-tbaa-struct-nam…

aa896c9

…ed-bitfields

Adjust

f372ca9

fhahn added a commit that referenced this pull request Feb 26, 2024

[TBAA] Add !tbaa.struct test with unnamed bitfields.

54cff50

Extra tests with unnamed bitfields for #82922.

Merge remote-tracking branch 'origin/main' into clang-tbaa-struct-nam…

ef8a494

…ed-bitfields

efriedma-quic reviewed Feb 26, 2024

View reviewed changes

fhahn added 2 commits February 27, 2024 09:00

!fixup fix formatting, update new test.

2f7f2d3

Merge remote-tracking branch 'origin/main' into clang-tbaa-struct-nam…

c4bcfeb

…ed-bitfields

efriedma-quic approved these changes Feb 27, 2024

View reviewed changes

fhahn merged commit d2a9df2 into llvm:main Feb 27, 2024

fhahn deleted the clang-tbaa-struct-named-bitfields branch February 27, 2024 20:09

fhahn added a commit to fhahn/llvm-project that referenced this pull request Mar 1, 2024

[TBAA] Add additional bitfield tests.

8b4ab71

Additional test for llvm#82922. (cherry-picked from f290c00)

fhahn added a commit to fhahn/llvm-project that referenced this pull request Mar 1, 2024

[TBAA] Add !tbaa.struct test with unnamed bitfields.

b60e76e

Extra tests with unnamed bitfields for llvm#82922. (cherry-picked from 54cff50)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[TBAA] Handle bitfields when generating !tbaa.struct metadata. #82922

[TBAA] Handle bitfields when generating !tbaa.struct metadata. #82922

fhahn commented Feb 25, 2024 •

edited

Loading

llvmbot commented Feb 25, 2024 •

edited

Loading

efriedma-quic commented Feb 26, 2024

github-actions bot commented Feb 26, 2024 •

edited

Loading

fhahn commented Feb 26, 2024

efriedma-quic Feb 26, 2024

fhahn Feb 26, 2024

efriedma-quic Feb 26, 2024

fhahn Feb 26, 2024

efriedma-quic left a comment

fhahn commented Feb 27, 2024

efriedma-quic left a comment

fhahn commented Feb 27, 2024

[TBAA] Handle bitfields when generating !tbaa.struct metadata. #82922

[TBAA] Handle bitfields when generating !tbaa.struct metadata. #82922

Conversation

fhahn commented Feb 25, 2024 • edited Loading

llvmbot commented Feb 25, 2024 • edited Loading

efriedma-quic commented Feb 26, 2024

github-actions bot commented Feb 26, 2024 • edited Loading

fhahn commented Feb 26, 2024

efriedma-quic Feb 26, 2024

Choose a reason for hiding this comment

fhahn Feb 26, 2024

Choose a reason for hiding this comment

efriedma-quic Feb 26, 2024

Choose a reason for hiding this comment

fhahn Feb 26, 2024

Choose a reason for hiding this comment

efriedma-quic left a comment

Choose a reason for hiding this comment

fhahn commented Feb 27, 2024

efriedma-quic left a comment

Choose a reason for hiding this comment

fhahn commented Feb 27, 2024

fhahn commented Feb 25, 2024 •

edited

Loading

llvmbot commented Feb 25, 2024 •

edited

Loading

github-actions bot commented Feb 26, 2024 •

edited

Loading