Vector of bool trips LLVM assertion "Invalid vector size"

[gwenzek]

Zig Version

0.9.0+my_fork https://github.com/gwenzek/zig/tree/llvm14
(To work on #11274 )

Steps to Reproduce

Simplified version of test/behavior/vector.zig

const std = @import("std");
const Vector = std.meta.Vector;

export fn vector_bitcast() u8 {
    const a: Vector(4, bool) = [_]bool{ true, false, true, false };
    const result_array: [4]bool = a;
    return if (result_array[0]) 0 else 1;
}

$ /home/guw/github/zig-bootstrap/out/build-zig-host/zig build-obj -I test --zig-lib-dir lib -Denable-llvm -Dconfig_h=/home/guw/github/zig-bootstrap/out/build-zig-host/config.h  test/behavior/vector.zig
LLVM Emit Object... zig: /home/guw/github/zig-bootstrap/llvm/lib/CodeGen/AsmPrinter/DwarfUnit.cpp:1475: bool hasVectorBeenPadded(const llvm::DICompositeType*): Assertion `ActualSize >= (NumVecElements * ElementSize) && "Invalid vector size"' failed.

Note: I'm using LLVM14 built with assertions on, but I don't think that LLVM14 is the main issue since this assertion is 4 years old
https://github.com/llvm/llvm-project/blame/75f9e83ace52773af65dcebca543005ec8a2705d/llvm/lib/CodeGen/AsmPrinter/DwarfUnit.cpp#L1473

A reduced version of the LLVM IR is:

define i8 @vector_bitcast() !dbg !5 {
Entry:
  store i8 0, i8* null, align 1, !dbg !19
  ret i8 0
}

!llvm.module.flags = !{!0, !1}
!llvm.dbg.cu = !{!2}

!0 = !{i32 2, !"Debug Info Version", i32 3}
!1 = !{i32 2, !"Dwarf Version", i32 4}
!2 = distinct !DICompileUnit(language: DW_LANG_C99, file: !3, producer: "zig 0.9.0", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, enums: !4, globals: !4)
!3 = !DIFile(filename: "vector", directory: "test/behavior")
!4 = !{}
!5 = distinct !DISubprogram(name: "vector_bitcast", scope: !6, file: !6, line: 22, type: !7, scopeLine: 22, flags: DIFlagStaticMember, spFlags: DISPFlagDefinition, unit: !2, retainedNodes: !10)
!6 = !DIFile(filename: "vector.zig", directory: "/home/guw/github/zig-bootstrap/zig/test/behavior")
!7 = !DISubroutineType(types: !8)
!8 = !{!9}
!9 = !DIBasicType(name: "u8", size: 8, encoding: DW_ATE_unsigned_char)
!10 = !{!11, !17}
!11 = !DILocalVariable(name: "a", scope: !12, file: !6, line: 23, type: !13)
!12 = distinct !DILexicalBlock(scope: !5, file: !6, line: 22, column: 31)
!13 = !DICompositeType(tag: DW_TAG_array_type, baseType: !14, size: 8, align: 1, flags: DIFlagVector, elements: !15)
!14 = !DIBasicType(name: "bool", size: 8, encoding: DW_ATE_boolean)
!15 = !{!16}
!16 = !DISubrange(count: 4, lowerBound: 0)
!17 = !DILocalVariable(name: "result_array", scope: !12, file: !6, line: 24, type: !18)
!18 = !DICompositeType(tag: DW_TAG_array_type, baseType: !14, size: 32, align: 32, elements: !15)
!19 = !DILocation(line: 25, column: 33, scope: !12)

Calling llc on it will also fail with the same error (unless built without assertions):

[guw@jinx-fedora zig]$ /home/guw/github/zig-bootstrap/out/host/bin/llc vector.ll 
llc: /home/guw/github/zig-bootstrap/llvm/lib/CodeGen/AsmPrinter/DwarfUnit.cpp:1475: bool hasVectorBeenPadded(const llvm::DICompositeType*): Assertion `ActualSize >= (NumVecElements * ElementSize) && "Invalid vector size"' failed.

As you can see from the IR, the bug is only due to the debug info and not to the actual code.
The issue is that:

• a is a vector of 4 booleans
• a has a size of 8 bits
• booleans have a size of 8 bits too
• DwarfUnit.cpp doesn't understand how we can fit 4 8-bits booleans inside an 8 bits vector.

Expected Behavior

I'm not sure what's the correct way of fixing this, but we shouldn't have an assertion error while writing the Dwarf Debug Info

Either:

• Zig should say that the vector is made of 1-bit booleans
• DwarfUnit.cpp shouldn't try to assert the vector size (I guess the reader of debug info should make this kind of checks if it's relevant to them)

Actual Behavior

/home/guw/github/zig-bootstrap/llvm/lib/CodeGen/AsmPrinter/DwarfUnit.cpp:1475: bool hasVectorBeenPadded(const llvm::DICompositeType*): Assertion `ActualSize >= (NumVecElements * ElementSize) && "Invalid vector size"' failed.

[andrewrk]

I confirmed this affects stage1 but not stage2.

[andrewrk]

Given that this works fine if you change it to a vector of u8:

export fn vector_bitcast() u8 {
    const a: @Vector(4, u8) = [_]u8{ 1, 2, 3, 4 };
    const result_array: [4]u8 = a;
    return result_array[0];
}

I think this is actually llvm/llvm-project#55522.

I also tried this patch which had no effect:

--- a/src/codegen/llvm.zig
+++ b/src/codegen/llvm.zig
@@ -1476,10 +1476,17 @@ pub const Object = struct {
                 return array_di_ty;
             },
             .Vector => {
+                // Lower vector of bool with the element type being a u1 instead so that
+                // LLVM debug info does not get confused and think that the element type has
+                // 8 bits instead of 1.
+                const elem_di_ty = if (ty.childType().zigTypeTag() == .Bool)
+                    try o.lowerDebugType(Type.u1, .full)
+                else
+                    try o.lowerDebugType(ty.childType(), .full);
                 const vector_di_ty = dib.createVectorType(
                     ty.abiSize(target) * 8,
                     ty.abiAlignment(target) * 8,
-                    try o.lowerDebugType(ty.childType(), .full),
+                    elem_di_ty,
                     ty.vectorLen(),
                 );
                 // The recursive call to `lowerDebugType` means we can't use `gop` anymore.
--- a/src/stage1/analyze.cpp
+++ b/src/stage1/analyze.cpp
@@ -9933,8 +9933,14 @@ static void resolve_llvm_types(CodeGen *g, ZigType *type, ResolveStatus wanted_r
             if (type->llvm_di_type != nullptr) return;
 
             type->llvm_type = LLVMVectorType(get_llvm_type(g, type->data.vector.elem_type), type->data.vector.len);
-            type->llvm_di_type = ZigLLVMDIBuilderCreateVectorType(g->dbuilder, 8 * type->abi_size,
-                    type->abi_align, get_llvm_di_type(g, type->data.vector.elem_type), type->data.vector.len);
+            ZigLLVMDIType *elem_di_ty = type->data.vector.elem_type->id == ZigTypeIdBool ?
+                get_llvm_di_type(g, get_int_type(g, false, 1)) :
+                get_llvm_di_type(g, type->data.vector.elem_type) ;
+            type->llvm_di_type = ZigLLVMDIBuilderCreateVectorType(g->dbuilder,
+                    8 * type->abi_size,
+                    8 * type->abi_align,
+                    elem_di_ty,
+                    type->data.vector.len);
             return;
         }
         case ZigTypeIdFnFrame:

We don't actually have a zig issue specifically open to track that upstream issue, however, so this one can be it.

[andrewrk]

Hmm looks like actually this is a separate issue from that LLVM issue because this diff did solve the issue:

--- a/src/stage1/analyze.cpp
+++ b/src/stage1/analyze.cpp
@@ -9933,8 +9933,15 @@ static void resolve_llvm_types(CodeGen *g, ZigType *type, ResolveStatus wanted_r
             if (type->llvm_di_type != nullptr) return;
 
             type->llvm_type = LLVMVectorType(get_llvm_type(g, type->data.vector.elem_type), type->data.vector.len);
-            type->llvm_di_type = ZigLLVMDIBuilderCreateVectorType(g->dbuilder, 8 * type->abi_size,
-                    type->abi_align, get_llvm_di_type(g, type->data.vector.elem_type), type->data.vector.len);
+            ZigLLVMDIType *elem_di_ty = type->data.vector.elem_type->id == ZigTypeIdBool ?
+                ZigLLVMCreateDebugBasicType(g->dbuilder, "bool", 1,
+                        ZigLLVMEncoding_DW_ATE_unsigned()) :
+                get_llvm_di_type(g, type->data.vector.elem_type) ;
+            type->llvm_di_type = ZigLLVMDIBuilderCreateVectorType(g->dbuilder,
+                    8 * type->abi_size,
+                    8 * type->abi_align,
+                    elem_di_ty,
+                    type->data.vector.len);
             return;
         }
         case ZigTypeIdFnFrame:

stage2 does not have this issue because integer and boolean types use the proper bit size rather than abi_size * 8.

[gwenzek]

@andrewrk here is how I was planning to solve this issue: in codegen.cpp@add_fp_entry use LLVMSizeOfTypeInBits instead of 8 * LLVMStoreSizeOfType

https://github.com/gwenzek/zig/pull/4/files#r912686791