Diego Novillo
2013-Nov-26 12:47 UTC
[LLVMdev] Finding the first line number of a function (-g vs -gline-tables-only)
For the sample profiler, I need to determine the location of the first user instruction in the function. This can be the LOC where the function header is or the LOC of the actual first instruction, it doesn't really matter. But it needs to be consistent wrt -g and -gline-tables-only. Currently, I'm scanning for the first non-zero line number from the top of the function. This skips over the frame setup instructions emitted by the compiler, that have no LOC, but it gives me different results depending on whether -g or -gline-tables-only is used, and whether the function has arguments: - If -g is used, and the function has arguments, the first non-zero loc instruction is the debug declaration call for the first argument: call void @llvm.dbg.declare(...) - If -g is used, and the function has no arguments, the first non-zero loc instruction is the first user instruction. - Similarly, if -gline-tables-only is used, we do not emit debug declarations for the arguments, so the first non-zero loc is the first user instruction. I want to modify the scanning I do in my pass to skip over all the debug setup instruction at the start of the function and latch on the first user instruction. Is there a straightforward way to do that? Right now I'm tempted to recognize calls to llvm.dbg.declare(), but I don't know if that leaves other cases unhandled. Thanks. Diego.
Diego Novillo
2013-Nov-26 13:05 UTC
[LLVMdev] Finding the first line number of a function (-g vs -gline-tables-only)
On Tue, Nov 26, 2013 at 7:47 AM, Diego Novillo <dnovillo at google.com> wrote:> I want to modify the scanning I do in my pass to skip over all the > debug setup instruction at the start of the function and latch on the > first user instruction. Is there a straightforward way to do that? > Right now I'm tempted to recognize calls to llvm.dbg.declare(), but I > don't know if that leaves other cases unhandled.I just found something that might be what I was looking for: if (Loc.getLine() != 0 && !isa<DbgInfoIntrinsic>(Inst)) Will this do? Thanks. Diego.
Diego Novillo
2013-Nov-26 14:01 UTC
[LLVMdev] Finding the first line number of a function (-g vs -gline-tables-only)
On Tue, Nov 26, 2013 at 7:47 AM, Diego Novillo <dnovillo at google.com> wrote:> For the sample profiler, I need to determine the location of the first > user instruction in the function. This can be the LOC where the > function header is or the LOC of the actual first instruction, it > doesn't really matter.Actually, I lied. It turns out that it DOES matter. I need the source location of the function header itself. So: 1 void 2 3 foo( 4 int x 5 6 ) 7 8 { 9 ... I need '3' to be my base for all the relative offsets I get from the profiler. So, how do I find out the line number of the symbol 'foo' itself? Thanks. Diego.
Renato Golin
2013-Nov-26 14:42 UTC
[LLVMdev] Finding the first line number of a function (-g vs -gline-tables-only)
On 26 November 2013 14:01, Diego Novillo <dnovillo at google.com> wrote:> I need '3' to be my base for all the relative offsets I get from the > profiler. So, how do I find out the line number of the symbol 'foo' > itself? >Hi Diego, In the metadata. Look for DW_TAG_subprogram: define i32 @main() #0 { ... } ... !4 = metadata !{i32 786478, metadata !1, metadata !5, metadata !"main", metadata !"main", metadata !"", i32 6, metadata !6, i1 false, i1 true, i32 0, i32 0, null, i32 256, i1 false, i32 ()* @main, null, null, metadata !2, i32 6} ; [ DW_TAG_subprogram ] [line 6] [def] [main] cheers, --renato -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20131126/53f97636/attachment.html>
Apparently Analagous Threads
- [LLVMdev] Finding the first line number of a function (-g vs -gline-tables-only)
- [LLVMdev] Finding the first line number of a function (-g vs -gline-tables-only)
- [LLVMdev] Recent changes in -gmlt break sample profiling
- [LLVMdev] Recent changes in -gmlt break sample profiling
- [LLVMdev] Recent changes in -gmlt break sample profiling