On Tue, Nov 9, 2010 at 9:04 AM, Devang Patel <dpatel at apple.com> wrote:> > > On Nov 8, 2010, at 10:52 PM, Talin <viridia at gmail.com> wrote: > > On Mon, Nov 8, 2010 at 9:56 AM, Devang Patel < <dpatel at apple.com> > dpatel at apple.com> wrote: > >> >> On Nov 6, 2010, at 7:35 PM, Talin wrote: >> >> After to speaking to Devang and a number of other people at the >> developer's conference, I was able to make some forward progress on getting >> debugging to work. I'm now able to actually single-step through my program >> and set breakpoints, and examine function parameters. >> >> However, I'm also seeing a lot of new problems which weren't exposed >> before. After spending the better part of two days working on them, I'd like >> to describe them here and get some advice on what I might be doing wrong. >> >> #1) Class sizes coming out as zero. >> >> In my frontend, I call DebugFactory::CreateCompositeTypeEx as follows: >> >> DICompositeType di = dbgFactory_.CreateCompositeTypeEx( >> type->typeClass() == Type::Class ? dwarf::DW_TAG_class_type : >> dwarf::DW_TAG_structure_type, >> dbgCompileUnit_, >> type->typeDefn()->linkageName().c_str(), >> genDIFile(type->typeDefn()), >> getSourceLineNumber(type->typeDefn()->location()), >> getSizeOfInBits(type->irType()), >> getAlignOfInBits(type->irType()), >> getInt64Val(0), 0, >> DIType(), >> dbgFactory_.GetOrCreateArray(members.data(), members.size())); >> >> The 'getSizeOfInBits()' function and the others like it basically call >> llvm::ConstantExpr::getSizeOf() and then multiple the result by 8. Looking >> at the generated IR, I can see that the expressions look correct to me (I've >> added some line breaks for clarity): >> >> !15 = metadata !{i32 589826, metadata !2, metadata !"tart.core.Object", >> metadata !12, i32 9, >> i64 mul (i64 ptrtoint (%tart.core.Object* getelementptr >> (%tart.core.Object* null, i32 1) to i64), i64 8), >> >> >> >> This does not work. The DebugInfo accessors uses getSizeInBits() to get an >> integer for the size. You will have to modify it to get the value. >> > > I'm not sure I understand what you are saying - isn't > llvm::ConstantExpr::getSizeOf() and llvm::ConstantExpr::getAlignOf() the > proper way to compute the size and alignment of a field? If those functions > can't be used for debug info, then that's really inconvenient - I don't want > to have to write my own calculations for it, since (a) I'm likely to get it > wrong in some subtle way, and (b) I'd be duplicating code, since the same > calculations need to be done in a lot of other places. > > > > > llvm::Constant* refers to llvm struct and fields. They do not necessarily > map directly to source lang structs and fields. Debugging information cares > about size, offset and alignment as per lang. That said, if anyone prepares > appropriate DebugInfo and Dwarfwriter patches then I won't oppose. > > First, apologies for responding to this after such a long interval.So I think I understand what you are saying here, although the way that I would say it is "Dwarfwriter doesn't understand constant expressions, only simple constants". At least, when I changed my code to use simple constants instead of ConstantExpr::getSizeOf(), it worked. Also, in my original message in this thread, I listed 3 issues. The one we've been discussing is issue #1 (class sizes coming out as zero) which I will consider solved. However, I never did figure out the answers to the other two issues I mentioned: 2) llvm.dbg.declare() not working for local variables, and 3) dwarfdump printing lots of errors when trying to dump a module produced by my frontend - even though when I check the debug info manually it all appears to be correct. Devang> > >> i6 4 mul (i64 ptrtoint (%tart.core.Object* getelementptr ({ i1, >> %tart.core.Object }* null, i64 0, i32 1) to i64), i64 8), >> i64 0, i32 0, null, metadata !16, i32 0, null} ; [ DW_TAG_class_type >> ] >> >> However, when I print out the debugging information via dwarfdump, you can >> see that the size of the object is zero: >> >> <1>< 2233> DW_TAG_class_type >> DW_AT_sibling <2295> >> DW_AT_name tart.core.Object >> * DW_AT_byte_size 0* >> DW_AT_decl_file 65 >> DW_AT_decl_line 9 >> <2>< 2258> DW_TAG_member >> DW_AT_name __tib >> DW_AT_type <2228> >> DW_AT_decl_file 65 >> DW_AT_decl_line 10 >> DW_AT_data_member_location DW_OP_plus_uconst 0 >> <2>< 2274> DW_TAG_member >> DW_AT_name __gcstate >> DW_AT_type <156> >> DW_AT_decl_file 65 >> DW_AT_decl_line 11 >> DW_AT_data_member_location DW_OP_plus_uconst 0 >> >> In addition, it appears that all of the field offsets are zero as well - >> see the DIEs immediately following the class definition. >> >> #2) I can't seem to get llvm.dbg.declare() to work for local variables, >> although it appears to work fine for function parameters. For example, look >> at the following snippet of IR: >> >> define {} @ReflectionTest.testModuleReflection(%ReflectionTest* %self) gc >> "tart-gc" { >> prologue: >> %m = alloca %tart.reflect.Module* >> call void @llvm.dbg.value(metadata !{%ReflectionTest* %self}, i64 0, >> metadata !1176) >> call void @llvm.dbg.declare(metadata !{%tart.reflect.Module** %m}, >> metadata !1177) >> %0 = bitcast %tart.reflect.Module** %m to i8**, !dbg !1178 >> call void @llvm.gcroot(i8** %0, i8* null), !dbg !1178 >> >> >> In the debugger, I can examine the value of 'self', however when I attempt >> to print the value of 'm', gdb reports 'No symbol "m" in current context.' >> >> >> Are using -O0? You do not have line number info with dbg.declare >> intrinsic. FWIW, following works ... >> >> define void @foo() nounwind ssp { >> entry: >> %i = alloca i32, align 4 >> call void @llvm.dbg.declare(metadata !{i32* %i}, metadata !5), !dbg !8 >> store i32 2, i32* %i, align 4, !dbg !9 >> ret void, !dbg !10 >> } >> >> - >> Devang >> >> >> > > > -- > -- Talin > >-- -- Talin -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20101125/eb3e174d/attachment.html>
On Thu, Nov 25, 2010 at 5:55 PM, Talin <viridia at gmail.com> wrote:> On Tue, Nov 9, 2010 at 9:04 AM, Devang Patel <dpatel at apple.com> wrote: > >> >> >> On Nov 8, 2010, at 10:52 PM, Talin <viridia at gmail.com> wrote: >> >> On Mon, Nov 8, 2010 at 9:56 AM, Devang Patel < <dpatel at apple.com> >> dpatel at apple.com> wrote: >> >>> >>> On Nov 6, 2010, at 7:35 PM, Talin wrote: >>> >>> After to speaking to Devang and a number of other people at the >>> developer's conference, I was able to make some forward progress on getting >>> debugging to work. I'm now able to actually single-step through my program >>> and set breakpoints, and examine function parameters. >>> >>> However, I'm also seeing a lot of new problems which weren't exposed >>> before. After spending the better part of two days working on them, I'd like >>> to describe them here and get some advice on what I might be doing wrong. >>> >>> #1) Class sizes coming out as zero. >>> >>> In my frontend, I call DebugFactory::CreateCompositeTypeEx as follows: >>> >>> DICompositeType di = dbgFactory_.CreateCompositeTypeEx( >>> type->typeClass() == Type::Class ? dwarf::DW_TAG_class_type : >>> dwarf::DW_TAG_structure_type, >>> dbgCompileUnit_, >>> type->typeDefn()->linkageName().c_str(), >>> genDIFile(type->typeDefn()), >>> getSourceLineNumber(type->typeDefn()->location()), >>> getSizeOfInBits(type->irType()), >>> getAlignOfInBits(type->irType()), >>> getInt64Val(0), 0, >>> DIType(), >>> dbgFactory_.GetOrCreateArray(members.data(), members.size())); >>> >>> The 'getSizeOfInBits()' function and the others like it basically call >>> llvm::ConstantExpr::getSizeOf() and then multiple the result by 8. Looking >>> at the generated IR, I can see that the expressions look correct to me (I've >>> added some line breaks for clarity): >>> >>> !15 = metadata !{i32 589826, metadata !2, metadata >>> !"tart.core.Object", metadata !12, i32 9, >>> i64 mul (i64 ptrtoint (%tart.core.Object* getelementptr >>> (%tart.core.Object* null, i32 1) to i64), i64 8), >>> >>> >>> >>> This does not work. The DebugInfo accessors uses getSizeInBits() to get >>> an integer for the size. You will have to modify it to get the value. >>> >> >> I'm not sure I understand what you are saying - isn't >> llvm::ConstantExpr::getSizeOf() and llvm::ConstantExpr::getAlignOf() the >> proper way to compute the size and alignment of a field? If those functions >> can't be used for debug info, then that's really inconvenient - I don't want >> to have to write my own calculations for it, since (a) I'm likely to get it >> wrong in some subtle way, and (b) I'd be duplicating code, since the same >> calculations need to be done in a lot of other places. >> >> >> >> >> llvm::Constant* refers to llvm struct and fields. They do not necessarily >> map directly to source lang structs and fields. Debugging information cares >> about size, offset and alignment as per lang. That said, if anyone prepares >> appropriate DebugInfo and Dwarfwriter patches then I won't oppose. >> >> First, apologies for responding to this after such a long interval. > > So I think I understand what you are saying here, although the way that I > would say it is "Dwarfwriter doesn't understand constant expressions, only > simple constants". At least, when I changed my code to use simple constants > instead of ConstantExpr::getSizeOf(), it worked. > > Also, in my original message in this thread, I listed 3 issues. The one > we've been discussing is issue #1 (class sizes coming out as zero) which I > will consider solved. However, I never did figure out the answers to the > other two issues I mentioned: 2) llvm.dbg.declare() not working for local > variables, and 3) dwarfdump printing lots of errors when trying to dump a > module produced by my frontend - even though when I check the debug info > manually it all appears to be correct. > > Also, there's another issue I wanted to mention: I have noticed thatdebugging and inlining don't play very well together. Even if other optimizations are disabled, if I turn on inlining, I get null pointer errors in this piece of code in DwarfDebug::constructInlinedScopeDIE(): *if* (!Scope->getScopeNode()) *return* NULL; DIScope DS(Scope->getScopeNode()); DIE *ScopeDIE = *new* DIE(dwarf::DW_TAG_inlined_subroutine); DISubprogram InlinedSP = getDISubprogram(DS); CompileUnit *TheCU = getCompileUnit(InlinedSP); DIE *OriginDIE = TheCU->getDIE(InlinedSP); assert(OriginDIE && *"Unable to find Origin DIE!"*); addDIEEntry(ScopeDIE, dwarf::DW_AT_abstract_origin, *dwarf*::DW_FORM_ref4, OriginDIE); What's happening is that InlinedSP.DbgNode is NULL, which causes getCompileUnit() to assert. Note that this occurs in both llc and my linker (tartln) because they both use the same set of passes to generate machine code. So basically whenever I want to debug, I have to disable inlining. Devang>> >> >>> i6 4 mul (i64 ptrtoint (%tart.core.Object* getelementptr ({ i1, >>> %tart.core.Object }* null, i64 0, i32 1) to i64), i64 8), >>> i64 0, i32 0, null, metadata !16, i32 0, null} ; [ >>> DW_TAG_class_type ] >>> >>> However, when I print out the debugging information via dwarfdump, you >>> can see that the size of the object is zero: >>> >>> <1>< 2233> DW_TAG_class_type >>> DW_AT_sibling <2295> >>> DW_AT_name tart.core.Object >>> * DW_AT_byte_size 0* >>> DW_AT_decl_file 65 >>> DW_AT_decl_line 9 >>> <2>< 2258> DW_TAG_member >>> DW_AT_name __tib >>> DW_AT_type <2228> >>> DW_AT_decl_file 65 >>> DW_AT_decl_line 10 >>> DW_AT_data_member_location DW_OP_plus_uconst 0 >>> <2>< 2274> DW_TAG_member >>> DW_AT_name __gcstate >>> DW_AT_type <156> >>> DW_AT_decl_file 65 >>> DW_AT_decl_line 11 >>> DW_AT_data_member_location DW_OP_plus_uconst 0 >>> >>> In addition, it appears that all of the field offsets are zero as well - >>> see the DIEs immediately following the class definition. >>> >>> #2) I can't seem to get llvm.dbg.declare() to work for local variables, >>> although it appears to work fine for function parameters. For example, look >>> at the following snippet of IR: >>> >>> define {} @ReflectionTest.testModuleReflection(%ReflectionTest* %self) gc >>> "tart-gc" { >>> prologue: >>> %m = alloca %tart.reflect.Module* >>> call void @llvm.dbg.value(metadata !{%ReflectionTest* %self}, i64 0, >>> metadata !1176) >>> call void @llvm.dbg.declare(metadata !{%tart.reflect.Module** %m}, >>> metadata !1177) >>> %0 = bitcast %tart.reflect.Module** %m to i8**, !dbg !1178 >>> call void @llvm.gcroot(i8** %0, i8* null), !dbg !1178 >>> >>> >>> In the debugger, I can examine the value of 'self', however when I >>> attempt to print the value of 'm', gdb reports 'No symbol "m" in current >>> context.' >>> >>> >>> Are using -O0? You do not have line number info with dbg.declare >>> intrinsic. FWIW, following works ... >>> >>> define void @foo() nounwind ssp { >>> entry: >>> %i = alloca i32, align 4 >>> call void @llvm.dbg.declare(metadata !{i32* %i}, metadata !5), !dbg !8 >>> store i32 2, i32* %i, align 4, !dbg !9 >>> ret void, !dbg !10 >>> } >>> >>> - >>> Devang >>> >>> >>> >> >> >> -- >> -- Talin >> >> > > > -- > -- Talin >-- -- Talin -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20101125/4410f255/attachment.html>
On Nov 25, 2010, at 5:55 PM, Talin wrote:> Also, in my original message in this thread, I listed 3 issues. The one we've been discussing is issue #1 (class sizes coming out as zero) which I will consider solved.ok> However, I never did figure out the answers to the other two issues I mentioned: 2) llvm.dbg.declare() not working for local variables,I suggest, you try looking at LLVM IR generated for small c test cases by Clang to understand what is missing link in your FE generated LLVM IR.> and 3) dwarfdump printing lots of errors when trying to dump a module produced by my frontend - even though when I check the debug info manually it all appears to be correct.I am assuming same dwarfdump error is repeated multiple times. Only thing I can suggest is get as small as possible LLVM IR where dwarfdump produces one or two errors and attach it to bugzilla PR. - Devang -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20101129/7e114abf/attachment.html>
On Nov 25, 2010, at 6:08 PM, Talin wrote:> Also, there's another issue I wanted to mention: I have noticed that debugging and inlining don't play very well together. Even if other optimizations are disabled, if I turn on inlining, I get null pointer errors in this piece of code in DwarfDebug::constructInlinedScopeDIE(): > if (!Scope->getScopeNode()) > return NULL; > DIScope DS(Scope->getScopeNode()); > DIE *ScopeDIE = new DIE(dwarf::DW_TAG_inlined_subroutine); > > DISubprogram InlinedSP = getDISubprogram(DS); > CompileUnit *TheCU = getCompileUnit(InlinedSP); > DIE *OriginDIE = TheCU->getDIE(InlinedSP); > assert(OriginDIE && "Unable to find Origin DIE!"); > addDIEEntry(ScopeDIE, dwarf::DW_AT_abstract_origin, > dwarf::DW_FORM_ref4, OriginDIE); > What's happening is that InlinedSP.DbgNode is NULL, which causes getCompileUnit() to assert.Please file a bugzilla PR with a reproduce testcase. A small LLVM IR which I can feed to llc to reproduce this crash should be good enough. Thanks, - Devang -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20101129/18e9dae6/attachment.html>