[WPD][ThinLTO]Add cutoff option for WPD

[mingmingl-llvm]

This option applies for import WPD (i.e., when DevirtModule pass de-virtualizes according to an imported summary, in ThinLTO backend pipeline). It's meant for debugging (e.g., bisection).

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Mingming Liu (minglotus-6)

Changes

This option applies for import WPD (i.e., when DevirtModule pass de-virtualizes according to an imported summary, in ThinLTO backend pipeline).

---

Full diff: https://github.com/llvm/llvm-project/pull/113383.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/IPO/WholeProgramDevirt.cpp (+12)
• (modified) llvm/test/Transforms/WholeProgramDevirt/import.ll (+6)

diff --git a/llvm/lib/Transforms/IPO/WholeProgramDevirt.cpp b/llvm/lib/Transforms/IPO/WholeProgramDevirt.cpp
index 36a1841b363463..f193bb11a06319 100644
--- a/llvm/lib/Transforms/IPO/WholeProgramDevirt.cpp
+++ b/llvm/lib/Transforms/IPO/WholeProgramDevirt.cpp
@@ -168,6 +168,14 @@ static cl::list<std::string>
                       cl::desc("Prevent function(s) from being devirtualized"),
                       cl::Hidden, cl::CommaSeparated);
 
+/// If this value is other than 0, the devirt module pass will stop
+/// transformation once the total number of devirtualizations reach the cutoff
+/// value.
+static cl::opt<unsigned> WholeProgramDevirtCutoff(
+    "wholeprogramdevirt-cutoff",
+    cl::desc("Max number of devirtualization for devirt module pass"),
+    cl::init(0));
+
 /// Mechanism to add runtime checking of devirtualization decisions, optionally
 /// trapping or falling back to indirect call on any that are not correct.
 /// Trapping mode is useful for debugging undefined behavior leading to failures
@@ -1169,6 +1177,10 @@ void DevirtModule::applySingleImplDevirt(VTableSlotInfo &SlotInfo,
       if (!OptimizedCalls.insert(&VCallSite.CB).second)
         continue;
 
+      if (WholeProgramDevirtCutoff.getNumOccurrences() > 0 &&
+          NumSingleImpl >= WholeProgramDevirtCutoff)
+        return;
+
       if (RemarksEnabled)
         VCallSite.emitRemark("single-impl",
                              TheFn->stripPointerCasts()->getName(), OREGetter);
diff --git a/llvm/test/Transforms/WholeProgramDevirt/import.ll b/llvm/test/Transforms/WholeProgramDevirt/import.ll
index 4e4df5f35a777d..de25bc10a7c123 100644
--- a/llvm/test/Transforms/WholeProgramDevirt/import.ll
+++ b/llvm/test/Transforms/WholeProgramDevirt/import.ll
@@ -8,6 +8,12 @@
 ; RUN: opt -S -passes=wholeprogramdevirt -wholeprogramdevirt-summary-action=import -wholeprogramdevirt-read-summary=%S/Inputs/import-vcp-branch-funnel.yaml < %s | FileCheck --check-prefixes=CHECK,VCP,VCP-X86,VCP64,BRANCH-FUNNEL %s
 ; RUN: opt -S -passes=wholeprogramdevirt -wholeprogramdevirt-summary-action=import -wholeprogramdevirt-read-summary=%S/Inputs/import-branch-funnel.yaml < %s | FileCheck --check-prefixes=CHECK,BRANCH-FUNNEL,BRANCH-FUNNEL-NOVCP %s
 
+; Cutoff value is not explicitly set. Expect 3 remark messages.
+; RUN: opt -S -passes=wholeprogramdevirt -wholeprogramdevirt-summary-action=import -pass-remarks=wholeprogramdevirt -wholeprogramdevirt-read-summary=%S/Inputs/import-single-impl.yaml < %s  2>&1 | grep "single-impl" | count 3
+; Cutoff value is set to 1. Expect one remark messages.
+; RUN: opt -S -passes=wholeprogramdevirt -wholeprogramdevirt-summary-action=import -pass-remarks=wholeprogramdevirt -wholeprogramdevirt-cutoff=1  -wholeprogramdevirt-read-summary=%S/Inputs/import-single-impl.yaml < %s  2>&1 | grep "single-impl" | count 1
+; Cutoff value is explicitly set to zero. Expect no remark message.
+; RUN: opt -S -passes=wholeprogramdevirt -wholeprogramdevirt-summary-action=import -pass-remarks=wholeprogramdevirt -wholeprogramdevirt-cutoff=0  -wholeprogramdevirt-read-summary=%S/Inputs/import-single-impl.yaml < %s 2>&1  | FileCheck -implicit-check-not="remark" %s
 target datalayout = "e-p:64:64"
 target triple = "x86_64-unknown-linux-gnu"
 

[teresajohnson]

s/devirtualization/devirtualizations/

[mingmingl-llvm]

done.

[teresajohnson]

NumSingleImpl is a STATISTIC - I'm not sure if it is safe to use this here as I believe the compiler can be built with statistics disabled, in which case this variable would not be correctly tracking the number.

[mingmingl-llvm]

Thanks for the catch!

I added static unsigned NumDevirtCalls to track the number of de-virtualized calls.

[mingmingl-llvm]

I tested the updated PR (using static unsigned NumDevirtCalls) using clang and verified the new option works

• -wholeprogramdevirt-cutoff=0 gives no devirtualization
• -wholeprogramdevirt-cutoff=1 allows at most one in one clang invocation (and catches the potentially interesting call-site)

[teresajohnson]

I believe the way the code is written, value 0 is not special and will cause no devirts to happen - assuming the option was specified (i.e. getNumOccurrences() > 0). And in fact we probably want a way to say "do 0 devirtualizations".

[mingmingl-llvm]

I updated the comment, and yes it makes sense to do 0 devirtualization when this option is explicitly set to zero.