BugZilla: https://bugs.llvm.org/show_bug.cgi?id=48206
The LCSSA pass makes use of a function insertDebugValuesForPHIs() to propogate dbg.value() intrinsics to newly inserted PHI instructions.
Faulty behaviour occurs when the associated parent PHI of a newly inserted PHI is not the most recent assignment to a source variable as can be seen below:
; input.ll:
%S2 = ...
br label %loop.interior
loop.interior:
br i1 %S2, label %if.true, label %if.false
if.true:
%X1 = ...
@llvm.dbg.value(var = %X1)
br label %post.if
if.false:
%X2 = ...
@llvm.dbg.value(var = %X2)
br label %post.if
post.if:
%X3 = phi(X1, X2)
@llvm.dbg.value(var = %X3)
%Y1 = ...
@llvm.dbg.value(var = %Y1)
br i1 %S2, label %loop.exit, label %loop.interior
loop.exit:
... = X3 + 4
; output.ll:
%S2 = ...
br label %loop.interior
loop.interior:
br i1 %S2, label %if.true, label %if.false
if.true:
%X1 = ...
@llvm.dbg.value(var = %X1)
br label %post.if
if.false:
%X2 = ...
@llvm.dbg.value(var = %X2)
br label %post.if
post.if:
%X3 = phi(X1, X2)
@llvm.dbg.value(var = %X3)
%Y1 = ...
@llvm.dbg.value(var = %Y1)
br i1 %S2, label %loop.exit, label %loop.interior
loop.exit:
%X3.lcssa = phi(X3)
@llvm.dbg.value(var = %X3.lcssa) <---- Incorrect!
%X4 = %X3.lcssa + 3In the pseudo-IR above, insertDebugValuesForPHIs() propogates the incorrect dbg.value() intrinsic. The intrinsic that should be propogated is the one associated with %Y1 because that is the most recent assignment to var, but it is simply ignored. This results in incorrect debugging information as it is effectively re-ordering assignments.
This change removes the call to insertDebugValuesForPHIs() from LCSSA, preventing incorrect dbg.value intrinsics from being propagated. As you can see in the locstats output below (generated by building Clang 3.4 with RelWithDebInfo with and without my change), even though the number of variables in the 100% bucket has gone down, the number of variables in the 0% bucket has also gone down. Apart from the fact that coverage has increased, incorrect debug information has also been removed (as seen in the example above). This is because even though LCSSA no longer inserts debug intrinsics for new PHIs, not inserting an intrinsic at all allows the last dominating dbg.value to continue dominating later instructions which results in the correct behaviour.
Before:
=================================================
Debug Location Statistics
=================================================
cov% samples percentage(~)
-------------------------------------------------
0% 841882 25%
(0%,10%) 41547 1%
[10%,20%) 49777 1%
[20%,30%) 49478 1%
[30%,40%) 43095 1%
[40%,50%) 45196 1%
[50%,60%) 56624 1%
[60%,70%) 53013 1%
[70%,80%) 65112 1%
[80%,90%) 71479 2%
[90%,100%) 100112 2%
100% 1935854 57%
=================================================
-the number of debug variables processed: 3353169
-PC ranges covered: 56%
-------------------------------------------------
-total availability: 61%
=================================================After:
=================================================
Debug Location Statistics
=================================================
cov% samples percentage(~)
-------------------------------------------------
0% 841874 25%
(0%,10%) 41652 1%
[10%,20%) 49847 1%
[20%,30%) 49592 1%
[30%,40%) 43230 1%
[40%,50%) 45257 1%
[50%,60%) 56594 1%
[60%,70%) 52978 1%
[70%,80%) 65097 1%
[80%,90%) 71391 2%
[90%,100%) 99888 2%
100% 1935817 57%
=================================================
-the number of debug variables processed: 3353217
-PC ranges covered: 56%
-------------------------------------------------
-total availability: 61%
=================================================(I've analyzed samples of both increases and decreases and confirmed that these changes are caused by my change in LCSSA and that the changes result in the correct behaviour)