[clang-repl] Duplicate definition error for symbols while run in C mode

[anutosh491]

Something like this fails

Example 1

anutosh491@Anutoshs-MacBook-Air bin % ./clang-repl --Xcc=-x --Xcc=c --Xcc=-std=c23 
clang-repl> #include <stdio.h> 
clang-repl> char *p; 
clang-repl> printf("ok\n"); 
error: Duplicate definition of symbol '_p'

OR

Example 2

anutosh491@Anutoshs-MacBook-Air bin % ./clang-repl --Xcc=-x --Xcc=c --Xcc=-std=c23 
clang-repl> int x; 
clang-repl> int y; 
error: Duplicate definition of symbol '_x'

[anutosh491]

Observation : Building in debug mode and adding some logs like (dumping TranslationUnitDecl for incoming input) I see this for example 2

./clang-repl --Xcc=-x --Xcc=c --Xcc=-std=c23 -debug-only=clang-repl, orc
clang-repl> int x; 

=== Dumping current TranslationUnitDecl ===
TranslationUnitDecl 0x13c91e350 prev 0x13c8fdd60 <<invalid sloc>> <invalid sloc>
`-VarDecl 0x13d019e00 <input_line_1:1:1, col:5> col:5 x 'int'
=== End TU dump ===
execute-ptu 1: [TU=0x13c91e350, M=0x13d809930 (incr_module_1)]
; ModuleID = 'incr_module_1'
source_filename = "incr_module_1"
target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-n32:64-S128-Fn32"
target triple = "arm64-apple-macosx15.0.0"

@x = global i32 0, align 4

!llvm.module.flags = !{!0, !1, !2, !3}
!llvm.ident = !{!4}

!0 = !{i32 1, !"wchar_size", i32 4}
!1 = !{i32 8, !"PIC Level", i32 2}
!2 = !{i32 7, !"uwtable", i32 1}
!3 = !{i32 7, !"frame-pointer", i32 1}
!4 = !{!"clang version 22.0.0git (git@github.com:anutosh491/llvm-project.git ef46f8a7d73c1657b2448fc2f3f41ff6eecc4c0f)"}
Defining MU incr_module_1 for main (tracker: 0x13c61bfb0)
  { ("_x", [Data]) }
LLJIT running initializers for JITDylib "main"
GenericLLVMIRPlatformSupport getting initializers to run
Issuing init-symbol lookup:
JITDylib init order is [ "<Process Symbols>" "<Platform>" "main" ]
Looking up init functions:
Issuing init-symbol lookup:
GenericLLVMIRPlatformSupport running initializers
clang-repl> int y;

=== Dumping current TranslationUnitDecl ===
TranslationUnitDecl 0x13d019e70 prev 0x13c91e350 <<invalid sloc>> <invalid sloc>
`-VarDecl 0x13d019ef0 <input_line_2:1:1, col:5> col:5 y 'int'
=== End TU dump ===
execute-ptu 2: [TU=0x13d019e70, M=0x13c6233c0 (incr_module_2)]
; ModuleID = 'incr_module_2'
source_filename = "incr_module_2"
target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-n32:64-S128-Fn32"
target triple = "arm64-apple-macosx15.0.0"

@x = global i32 0, align 4
@y = global i32 0, align 4

!llvm.module.flags = !{!0, !1, !2, !3}
!llvm.ident = !{!4}

!0 = !{i32 1, !"wchar_size", i32 4}
!1 = !{i32 8, !"PIC Level", i32 2}
!2 = !{i32 7, !"uwtable", i32 1}
!3 = !{i32 7, !"frame-pointer", i32 1}
!4 = !{!"clang version 22.0.0git (git@github.com:anutosh491/llvm-project.git ef46f8a7d73c1657b2448fc2f3f41ff6eecc4c0f)"}
Defining MU incr_module_2 for main (tracker: 0x12b3042b0)
  { ("_y", [Data]), ("_x", [Data]) }
  Error: Duplicate symbols { _x }
error: In incr_module_2, duplicate definition of symbol '_x'
clang-repl> 

As can be the seen the error is reported while defining the MaterializationUnit for incr_module_2, but the Module generated for input 2 itself is corrupt (as has references to both x & y) . But the TranslationUnitDecl for each input is appropriate (TUDECL for input 1 mentions x, for input 2 mentions y)

[anutosh491]

Why this happens : C vs C++ semantics that weren't considered

• C: A file-scope int x; without initializer is a tentative definition. C requires it to become a zero-initialized definition by end of TU if no real definition appears.

• C++: There is no tentative definition rule. int x; is a declaration until defined; no end-of-TU synthesis.

So basically ParseTopLevelDecl hits tok::annot_repl_input_end and calls Sema's ActOnEndOfTranslationUnit that calls Consumer's CompleteTentativeDefinition which follows (BackendConsumer → CodeGenerator → EmitTentativeDefinition)

So adding more logs (dumping ADECl on which ParseTopLevelDecl is being called and dumping the source of the tentative definitions) I see the following (skipping most logs from above)

clang-repl> int x;

=== Parsed top-level DeclGroup ===
Var x
[ActOnEndOfTranslationUnit] Completing tentative definition for x from input_line_1:1:5
[EmitTentativeDefinition] x  from input_line_1:1:5

=== Dumping current TranslationUnitDecl ===
TranslationUnitDecl 0x12c90db50 prev 0x12c8e1d60 <<invalid sloc>> <invalid sloc>
`-VarDecl 0x12e809400 <input_line_1:1:1, col:5> col:5 x 'int'
=== End TU dump ===
execute-ptu 1: [TU=0x12c90db50, M=0x12bf44860 (incr_module_1)]
; ModuleID = 'incr_module_1'

@x = global i32 0, align 4

!llvm.module.flags = !{!0, !1, !2, !3}
!llvm.ident = !{!4}

clang-repl> int y;

=== Parsed top-level DeclGroup ===
Var y
[ActOnEndOfTranslationUnit] Completing tentative definition for y from input_line_2:1:5
[EmitTentativeDefinition] x  from input_line_1:1:5
[EmitTentativeDefinition] y  from input_line_2:1:5

=== Dumping current TranslationUnitDecl ===
TranslationUnitDecl 0x12e809470 prev 0x12c90db50 <<invalid sloc>> <invalid sloc>
`-VarDecl 0x12e8094f0 <input_line_2:1:1, col:5> col:5 y 'int'
=== End TU dump ===
execute-ptu 2: [TU=0x12e809470, M=0x12d10e670 (incr_module_2)]
; ModuleID = 'incr_module_2'

@x = global i32 0, align 4
@y = global i32 0, align 4

!llvm.module.flags = !{!0, !1, !2, !3}
!llvm.ident = !{!4}

error: In incr_module_2, duplicate definition of symbol '_x'

As can be seen input 2 also holds the Tentative Definition from previous inputs

[anutosh491]

Solution Proposed

Only complete tentative definitions belonging to the current PTU’s TUDecl in incremental mode. Prior PTUs’ tentatives must be ignored.

@@ void Sema::ActOnEndOfTranslationUnit() {
     if (!VD || VD->isInvalidDecl() || !Seen.insert(VD).second)
       continue;

+    // In incremental mode (REPL), do not complete tentative definitions
+    // that belong to prior PTUs. Each PTU has its own TranslationUnitDecl.
+    if (PP.isIncrementalProcessingEnabled() &&
+        VD->getTranslationUnitDecl() != Context.getTranslationUnitDecl())
+      continue;

This way, we preserves C semantics( tentative definitions from the current PTU still get completed at end-of-PTU) and prevent cross ptu re-emission.

C++ mode is unaffected because TentativeDefinitions is empty there.

This should now give us

clang-repl> int x;

=== Parsed top-level DeclGroup ===
Var x
[ActOnEndOfTranslationUnit] Completing tentative definition for x from input_line_1:1:5
[EmitTentativeDefinition] x  from input_line_1:1:5

=== Dumping current TranslationUnitDecl ===
TranslationUnitDecl 0x12d12af50 prev 0x12d0fa760 <<invalid sloc>> <invalid sloc>
`-VarDecl 0x12f008200 <input_line_1:1:1, col:5> col:5 x 'int'
=== End TU dump ===
clang-repl> int y;

=== Parsed top-level DeclGroup ===
Var y
[ActOnEndOfTranslationUnit] Completing tentative definition for y from input_line_2:1:5
[EmitTentativeDefinition] y  from input_line_2:1:5

=== Dumping current TranslationUnitDecl ===
TranslationUnitDecl 0x12f008270 prev 0x12d12af50 <<invalid sloc>> <invalid sloc>
`-VarDecl 0x12f0082f0 <input_line_2:1:1, col:5> col:5 y 'int'
=== End TU dump ===
clang-repl> char *p;

=== Parsed top-level DeclGroup ===
Var p
[ActOnEndOfTranslationUnit] Completing tentative definition for p from input_line_3:1:7
[EmitTentativeDefinition] p  from input_line_3:1:7

=== Dumping current TranslationUnitDecl ===
TranslationUnitDecl 0x12f008360 prev 0x12f008270 <<invalid sloc>> <invalid sloc>
`-VarDecl 0x12f0083e0 <input_line_3:1:1, col:7> col:7 p 'char *'
=== End TU dump ===

[anutosh491]

Link to PR : #164597