I noticed that the fast/linear DAG scheduler doesn't lower DBG_VALUEs except for function entry nodes. This copies what appears to be the relevant code, and it does seem to work for most of the simple testing I've done.
Thanks for working on this.
|791 ↗||(On Diff #133242)|
Nit: could you used a range-for loop here and run clang-format over the diff?
|29 ↗||(On Diff #133242)|
I'm not able to follow what's changed based on these check lines. Could you attach llc's pre-patch/post-patch output, so we can see how the output changes?
|797 ↗||(On Diff #133242)|
Don't put braces around a single-line body.
if (auto *DbgMI = Emitter.EmitDbgValue(DVs[i], VRBaseMap)) BB->insert(InsertPos, DbgMI);
Effectively, before this patch, the only dbg.value entries that get added are those that have to do with function arguments. The ones in the middle of the function are completely ignored.
This is in line with how we handle debug values in ScheduleDAGSDNodes::EmitSchedule (see ProcessSDDbgValues).
I don't see any issues with this patch but am not familiar enough with ScheduleDAGFast to give an explicit lgtm. I'll CC some folks who might be able to.
|29 ↗||(On Diff #134158)|
It would be better to write this as
CHECK: DBG_VALUE ![[X:[0-9]+]], ![[EXPR:[0-9]+]] CHECK: [[EXPR]] = !DIExpression() ...
etc. Hard-coding the metadata numbering is just asking for trouble :-)
There's something weird with the test output:
bb.3.for.body: successors: %bb.2(0x04000000), %bb.3(0x7c000000) %2:gr64 = PHI %8, %bb.1, %7, %bb.3 %3:gr64 = PHI %0, %bb.1, %6, %bb.3 %4:gr32 = PHI %11, %bb.1, %5, %bb.3 %5:gr32 = ADD32rm %4, %2, 1, $noreg, 0, $noreg, implicit-def dead $eflags, debug-location !33 :: (load 4 from %ir.lsr.iv1, !tbaa !29) DBG_VALUE debug-use %5, debug-use $noreg, !14, !DIExpression(), debug-location !19 DBG_VALUE debug-use %5, debug-use $noreg, !14, !DIExpression(), debug-location !19 %6:gr64 = DEC64r %3, implicit-def $eflags, debug-location !21 %7:gr64 = ADD64ri8 %2, 4, implicit-def dead $eflags, debug-location !21 JE_1 %bb.2, implicit $eflags, debug-location !24 JMP_1 %bb.3, debug-location !24
Why is this DBG_VALUE inserted twice?