[PATCH] D19564: Update Debug Intrinsics in RewriteUsesOfClonedInstructions in LoopRotation

Wed Apr 27 16:53:56 PDT 2016

> On Apr 27, 2016, at 4:48 PM, David Blaikie <dblaikie at gmail.com> wrote:
> 
> 
> 
> On Wed, Apr 27, 2016 at 4:44 PM, Adrian Prantl <aprantl at apple.com <mailto:aprantl at apple.com>> wrote:
> 
>> On Apr 27, 2016, at 3:22 PM, David Blaikie via llvm-commits <llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>> wrote:
>> 
>> 
>> 
>> On Wed, Apr 27, 2016 at 12:46 PM, Thomas Jablin <tjablin at gmail.com <mailto:tjablin at gmail.com>> wrote:
>> The short answer is, "Yes." 
>> 
>> If a DbgInfoIntrinsic extended the live range of Value before LoopRotation, then LoopRotation may create a situation where the definition of the Value no dominates the DbgInfoIntrinsic. If so, the proposed patch will insert a PHINode.
>> 
>> If DbgInfoIntrinsics can extend the live range of a Value, then register allocation and other aspects of optimization and code generation will inevitably change. The alternative is pruning DbgInfoIntrinsics if they use Values that no longer dominate them and when they would extend the live-range of a Value.
>> 
>> I believe this (alternative) ^ would be necessary. I don't know if the right thing to do is prune the intrinsic, or just replace its value with undef, etc. I forget/don't know how we usually do it.
> 
> Let me know if I misunderstand something here, but AFAIK the register allocator goes out of its way to treat debug intrinsics as transparent.
> 
> You might have more context here than I do - it just sounded like the change meant a use of a value, and a phi, that would otherwise not exist & that seemed like it could affect the result. I could be wrong here, though. Just wanted to highlight it & was making my best-guess at what would be behavior-preserving or not.

No worries :-)
If created with DIBuilder, the arguments to debug intrinsics are created with MetadataAsValue::get(VMContext, ValueAsMetadata::get(V)) which ensures that they are never a use. They still often end up inadvertently messing with code generation just because they take up a place in the the instruction stream and then obstruct peephole optimizations, but that’s a separate story.

> 
> Totally up for other people to chime in here, I don't feel sufficiently informed to do final review on this change anyway. Was just trying to cover some parts that might've needed addressing regardless.
> 
> - Dave
>  
> 
> -- adrian
> 
>>  
>> 
>> If the DbgInfoIntrinsic is inside the live-range of a Value and does not extend it, no additional PHINodes will be need to be inserted.
>> 
>> On Wed, Apr 27, 2016 at 1:27 PM, David Blaikie <dblaikie at gmail.com <mailto:dblaikie at gmail.com>> wrote:
>> Does this cause changes in non-debug IR in the presence of debug info intrinsics? (you mention that "If there were uses of values produced by these instructions that were outside the loop, we have to insert PHI nodes to merge the two values." - that sounds like it could be a problem, as enabling debug info is not intended to change optimization/code generation)
>> 
>> On Tue, Apr 26, 2016 at 3:34 PM, Thomas Jablin via llvm-commits <llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>> wrote:
>> tjablin created this revision.
>> tjablin added reviewers: kbarton, hfinkel, cycheng.
>> tjablin added a subscriber: llvm-commits.
>> Herald added a subscriber: mzolotukhin.
>> 
>> Loop rotation clones instruction from the old header into the preheader. If there were uses of values produced by these instructions that were outside the loop, we have to insert PHI nodes to merge the two values. If the values are used by DbgIntrinsics they will be used as a MetadataAsValue of a ValueAsMetadata of the original values, and iterating all of the uses of the original value will not update the DbgIntrinsics. The new code checks if the values are used by DbgIntrinsics and if so, updates them using essentially the same logic as the original code.
>> 
>> The attached testcase demonstrates the issue. Without the fix, the DbgIntrinic outside the loop uses values computed inside the loop, even though these values do not dominate the DbgIntrinsic.
>> 
>> http://reviews.llvm.org/D19564 <http://reviews.llvm.org/D19564>
>> 
>> Files:
>>   lib/Transforms/Scalar/LoopRotation.cpp
>>   test/Transforms/LoopRotate/dbgvalue.ll
>> 
>> Index: test/Transforms/LoopRotate/dbgvalue.ll
>> ===================================================================
>> --- test/Transforms/LoopRotate/dbgvalue.ll
>> +++ test/Transforms/LoopRotate/dbgvalue.ll
>> @@ -7,6 +7,7 @@
>>  ; CHECK-LABEL: define i32 @tak(
>>  ; CHECK: entry
>>  ; CHECK-NEXT: call void @llvm.dbg.value(metadata i32 %x
>> +; CHECK: tail call void @llvm.dbg.value(metadata i32 %call
>> 
>>  entry:
>>    br label %tailrecurse
>> @@ -37,6 +38,44 @@
>>    ret i32 %z.tr <http://z.tr/>, !dbg !17
>>  }
>> 
>> +define i32 @tak2(i32 %x, i32 %y, i32 %z) nounwind ssp !dbg !0 {
>> +; CHECK-LABEL: define i32 @tak2(
>> +; CHECK: entry
>> +; CHECK: tail call void @llvm.dbg.value(metadata i32 %x.tr <http://x.tr/>
>> +; CHECK: tail call void @llvm.dbg.value(metadata i32 %x.tr <http://x.tr/>
>> +
>> +entry:
>> +  br label %tailrecurse
>> +
>> +tailrecurse:                                      ; preds = %if.then, %entry
>> +  %x.tr <http://x.tr/> = phi i32 [ %x, %entry ], [ %call, %if.then ]
>> +  %y.tr <http://y.tr/> = phi i32 [ %y, %entry ], [ %call9, %if.then ]
>> +  %z.tr <http://z.tr/> = phi i32 [ %z, %entry ], [ %call14, %if.then ]
>> +  %cmp = icmp slt i32 %y.tr <http://y.tr/>, %x.tr <http://x.tr/>, !dbg !12
>> +  br i1 %cmp, label %if.then, label %if.end, !dbg !12
>> +
>> +if.then:                                          ; preds = %tailrecurse
>> +  tail call void @llvm.dbg.value(metadata i32 %x.tr <http://x.tr/>, i64 0, metadata !6, metadata !DIExpression()), !dbg !7
>> +  tail call void @llvm.dbg.value(metadata i32 %y.tr <http://y.tr/>, i64 0, metadata !8, metadata !DIExpression()), !dbg !9
>> +  tail call void @llvm.dbg.value(metadata i32 %z.tr <http://z.tr/>, i64 0, metadata !10, metadata !DIExpression()), !dbg !11
>> +  %sub = sub nsw i32 %x.tr <http://x.tr/>, 1, !dbg !14
>> +  %call = tail call i32 @tak(i32 %sub, i32 %y.tr <http://y.tr/>, i32 %z.tr <http://z.tr/>), !dbg !14
>> +  %sub6 = sub nsw i32 %y.tr <http://y.tr/>, 1, !dbg !14
>> +  %call9 = tail call i32 @tak(i32 %sub6, i32 %z.tr <http://z.tr/>, i32 %x.tr <http://x.tr/>), !dbg !14
>> +  %sub11 = sub nsw i32 %z.tr <http://z.tr/>, 1, !dbg !14
>> +  %call14 = tail call i32 @tak(i32 %sub11, i32 %x.tr <http://x.tr/>, i32 %y.tr <http://y.tr/>), !dbg !14
>> +  br label %tailrecurse
>> +
>> +if.end:                                           ; preds = %tailrecurse
>> +  tail call void @llvm.dbg.value(metadata i32 %x.tr <http://x.tr/>, i64 0, metadata !6, metadata !DIExpression()), !dbg !7
>> +  tail call void @llvm.dbg.value(metadata i32 %y.tr <http://y.tr/>, i64 0, metadata !8, metadata !DIExpression()), !dbg !9
>> +  tail call void @llvm.dbg.value(metadata i32 %z.tr <http://z.tr/>, i64 0, metadata !10, metadata !DIExpression()), !dbg !11
>> +  br label %return, !dbg !16
>> +
>> +return:                                           ; preds = %if.end
>> +  ret i32 %z.tr <http://z.tr/>, !dbg !17
>> +}
>> +
>>  @channelColumns = external global i64
>>  @horzPlane = external global i8*, align 8
>> 
>> Index: lib/Transforms/Scalar/LoopRotation.cpp
>> ===================================================================
>> --- lib/Transforms/Scalar/LoopRotation.cpp
>> +++ lib/Transforms/Scalar/LoopRotation.cpp
>> @@ -108,6 +108,33 @@
>> 
>>        // Anything else can be handled by SSAUpdater.
>>        SSA.RewriteUse(U);
>> +
>> +      // Replace MetadataAsValue(ValueAsMetadata(OrigHeaderVal)) uses in debug
>> +      // intrinsics
>> +      LLVMContext &C = OrigHeader->getContext();
>> +      if (ValueAsMetadata *VAM = ValueAsMetadata::getIfExists(OrigHeaderVal)) {
>> +        if (MetadataAsValue *MAV = MetadataAsValue::getIfExists(C, VAM)) {
>> +          for (auto UI = MAV->use_begin(), E = MAV->use_end(); UI != E; ) {
>> +            // Grab the use before incrementing the iterator.
>> +            Use &U = *UI++;
>> +            DbgInfoIntrinsic *UserInst = cast<DbgInfoIntrinsic>(U.getUser());
>> +            BasicBlock *UserBB = UserInst->getParent();
>> +
>> +            // The original users in the OrigHeader are already using the
>> +            // original definitions.
>> +            if (UserBB == OrigHeader)
>> +              continue;
>> +
>> +            // Users in the OrigPreHeader need to use the value to which the
>> +            // original definitions are mapped and anything else can be handled
>> +            // by the SSAUpdater
>> +            Value *NewVal = UserBB == OrigPreheader ?
>> +              OrigPreHeaderVal :
>> +              SSA.GetValueInMiddleOfBlock(UserBB);
>> +            U = MetadataAsValue::get(C, ValueAsMetadata::get(NewVal));
>> +          }
>> +        }
>> +      }
>>      }
>>    }
>>  }
>> 
>> 
>> 
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits <http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits>
>> 
>> 
>> 
>> 
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits <http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits>
> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160427/5ac2a664/attachment.html>