[LLVMdev] Coverage mapping issue: Malformed profile data

[Dmitry Volyntsev]

Here is probably a related bug: http://llvm.org/bugs/show_bug.cgi?id=22575

On Tue, Feb 10, 2015 at 7:47 PM, Justin Bogner <mail at justinbogner.com> wrote:
> Dmitry Volyntsev <xeioexception at gmail.com> writes:
>> It seems I came across on issue with coverage mapping
>> (http://www.llvm.org/docs/CoverageMappingFormat.html)
>>
>> check on:
>> llvm revision: r228136
>> clang  Last Changed Rev: 228121
>> build: Debug+Asserts
>> OS: ubuntu 14.04
>>
>>
>> Here is simple snippets
>>
>> test1.c: NOT OK
>> ==================
>> #include <stdio.h>
>>
>> static int foo() { return 42; }
>> int main() {
>>  return 0;
>> }
>> ==================
>> cp src/test1.c src/test.c; rm -fr default.profraw test.profdata ;
>> clang -c -o obj/test.o src/test.c -I -g -O0 -fprofile-instr-generate
>> -fcoverage-mapping; clang -o bin/test obj/test.o -I -g -O0
>> -fprofile-instr-generate -fcoverage-mapping; ./bin/test ;
>> llvm-profdata merge -o test.profdata default.profraw; llvm-cov show
>> ./bin/test -instr-profile=test.profdata
>>
>> error: Failed to load coverage: Malformed profile data
>>
>>
>>
>> test2.c: OK
>> ==================
>> #include <stdio.h>
>>
>> static int foo() { return 42; }
>> int main() {
>>  return foo();
>> }
>> ==================
>> cp src/test2.c src/test.c; rm -fr default.profraw test.profdata ;
>> clang -c -o obj/test.o src/test.c -I -g -O0 -fprofile-instr-generate
>> -fcoverage-mapping; clang -o bin/test obj/test.o -I -g -O0
>> -fprofile-instr-generate -fcoverage-mapping; ./bin/test ;
>> llvm-profdata merge -o test.profdata default.profraw; llvm-cov show
>> ./bin/test -instr-profile=test.profdata
>>
>> /home/xeioex/workspace/clang/experiments/llvm-cov/test/src/test.c:
>>        |    1|#include <stdio.h>
>>        |    2|
>>       1|    3|static int foo() { return 42; }
>>       1|    4|int main() {
>>       1|    5|  return foo();
>>       1|    6|}
>>
>> test3.c: OK
>> ==================
>> #include <stdio.h>
>>
>> int foo() { return 42; }
>> int main() {
>>  return 0;
>> }
>> ==================
>> cp src/test3.c src/test.c; rm -fr default.profraw test.profdata ;
>> clang -c -o obj/test.o src/test.c -I -g -O0 -fprofile-instr-generate
>> -fcoverage-mapping; clang -o bin/test obj/test.o -I -g -O0
>> -fprofile-instr-generate -fcoverage-mapping; ./bin/test ;
>> llvm-profdata merge -o test.profdata default.profraw; llvm-cov show
>> ./bin/test -instr-profile=test.profdata
>>
>> /home/xeioex/workspace/clang/experiments/llvm-cov/test/src/test.c:
>>        |    1|#include <stdio.h>
>>        |    2|
>>       0|    3|int foo() { return 42; }
>>       1|    4|int main() {
>>       1|    5|  return 0;
>>       1|    6|}
>>
>>
>> As far I can understand issue somehow related to unused static functions.
>
> It looks like this hits when a function isn't emitted by the frontend
> but we emit an "empty" coverage map for it.
>
>> Also, as I can track down the problem raised because of
>> lib/ProfileData/CoverageMappingReader.cpp (SectionData::get) expects
>> that function's names located inside __llvm_prf_names section but it
>> not true in case of test1.c, instead some function names located
>> inside .rodata section.
>
> The problem is that these don't go through the InstrProfiling machinery,
> so the names aren't lowered properly. I'd noticed a related issue not
> long ago, but fixed a bug by inspection that happened to hide it
> (r226764).
>
> Feel free to file a bug (http://llvm.org/bugs/) and I'll to get this
> fixed in the next day or two.



-- 
Be happy,
Best regards,
Dmitry Volyntsev