[MemProf] Don't mutate the function type when calling clone

[teresajohnson]

In rare cases the declaration of a function may not match its callsite
after function importing, when the declaration was imported from a
module where the function had void return type (presumably due to
incomplete types). Instead of using setCalledFunction() to change a call
to call its clone, which updates the call's function type as well, just
call setCalledOperand directly so the only thing changed is the function
target.

Note this can't happen for the other places where we call
setCalledFunction: FullLTO requires the cloned callee to be defined in
the same FullLTO merged module; ThinLTO memprof ICP calls an ICP
facility to first perform the promotion and that will be blocked if the
function type doesn't match the callsite (the new test explicitly tests
this latter case).

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-lto

Author: Teresa Johnson (teresajohnson)

Changes

In rare cases the declaration of a function may not match its callsite
after function importing, when the declaration was imported from a
module where the function had void return type (presumably due to
incomplete types). Instead of using setCalledFunction() to change a call
to call its clone, which updates the call's function type as well, just
call setCalledOperand directly so the only thing changed is the function
target.

Note this can't happen for the other places where we call
setCalledFunction: FullLTO requires the cloned callee to be defined in
the same FullLTO merged module; ThinLTO memprof ICP calls an ICP
facility to first perform the promotion and that will be blocked if the
function type doesn't match the callsite (the new test explicitly tests
this latter case).

---

Full diff: https://github.com/llvm/llvm-project/pull/147829.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/IPO/MemProfContextDisambiguation.cpp (+7-1)
• (added) llvm/test/ThinLTO/X86/memprof_callee_type_mismatch.ll (+63)

diff --git a/llvm/lib/Transforms/IPO/MemProfContextDisambiguation.cpp b/llvm/lib/Transforms/IPO/MemProfContextDisambiguation.cpp
index c0f84456d2b27..24dd8bcf618c8 100644
--- a/llvm/lib/Transforms/IPO/MemProfContextDisambiguation.cpp
+++ b/llvm/lib/Transforms/IPO/MemProfContextDisambiguation.cpp
@@ -5173,7 +5173,13 @@ bool MemProfContextDisambiguation::applyImport(Module &M) {
           CBClone = CB;
         else
           CBClone = cast<CallBase>((*VMaps[J - 1])[CB]);
-        CBClone->setCalledFunction(NewF);
+        // Set the called operand directly instead of calling setCalledFunction,
+        // as the latter mutates the function type on the call. In rare cases
+        // we may have a slightly different type on a callee function
+        // declaration due to it being imported from a different module with
+        // incomplete types. We really just want to change the name of the
+        // function to the clone, and not make any type changes.
+        CBClone->setCalledOperand(NewF.getCallee());
         ORE.emit(OptimizationRemark(DEBUG_TYPE, "MemprofCall", CBClone)
                  << ore::NV("Call", CBClone) << " in clone "
                  << ore::NV("Caller", CBClone->getFunction())
diff --git a/llvm/test/ThinLTO/X86/memprof_callee_type_mismatch.ll b/llvm/test/ThinLTO/X86/memprof_callee_type_mismatch.ll
new file mode 100644
index 0000000000000..60804ab52ba65
--- /dev/null
+++ b/llvm/test/ThinLTO/X86/memprof_callee_type_mismatch.ll
@@ -0,0 +1,63 @@
+;; Test to ensure the call updated to call a clone does not mutate the callee
+;; function type. In rare cases we may end up with a callee declaration that
+;; does not match the call type, because it was imported from a different
+;; module with an incomplete return type (in which case clang gives it a void
+;; return type).
+
+; RUN: rm -rf %t && split-file %s %t && cd %t
+; RUN: llvm-as src.ll -o src.o
+; RUN: llvm-as src.o.thinlto.ll -o src.o.thinlto.bc
+
+; RUN: opt -passes=memprof-context-disambiguation src.o -S -memprof-import-summary=src.o.thinlto.bc | FileCheck %s
+
+;--- src.ll
+; ModuleID = 'src.o'
+source_filename = "src.c"
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define i32 @main(ptr %b) {
+entry:
+  ;; This call is not changed as the summary specifies clone 0.
+  ; CHECK: call ptr @_Z3foov()
+  %call = call ptr @_Z3foov(), !callsite !5
+  ;; After changing this call to call a clone, the function type should still
+  ;; be ptr, despite the void on the callee declaration.
+  ; CHECK: call ptr @_Z3foov.memprof.1()
+  %call1 = call ptr @_Z3foov(), !callsite !6
+  %0 = load ptr, ptr %b, align 8
+  ;; Although the summary indicates this should call clone 1, and the VP
+  ;; metadata indicates the callee is _Z3foov, it is not updated because
+  ;; the ICP facility requires the function types to match.
+  ; CHECK: call ptr %0()
+  %call2 = call ptr %0(), !prof !7, !callsite !8
+  ret i32 0
+}
+
+;; Both the original callee function declaration and its clone have void return
+;; type.
+; CHECK: declare void @_Z3foov()
+; CHECK: declare void @_Z3foov.memprof.1()
+declare void @_Z3foov()
+
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!2, !3, !4}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus_14, file: !1, producer: "clang version 21.0.0git ([email protected]:llvm/llvm-project.git e391301e0e4d9183fe06e69602e87b0bc889aeda)", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug, splitDebugInlining: false, nameTableKind: None)
+!1 = !DIFile(filename: "src.cc", directory: "", checksumkind: CSK_MD5, checksum: "8636c46e81402013b9d54e8307d2f149")
+!2 = !{i32 7, !"Dwarf Version", i32 5}
+!3 = !{i32 2, !"Debug Info Version", i32 3}
+!4 = !{i32 1, !"EnableSplitLTOUnit", i32 0}
+!5 = !{i64 8632435727821051414}
+!6 = !{i64 -3421689549917153178}
+!7 = !{!"VP", i32 0, i64 4, i64 9191153033785521275, i64 4}
+!8 = !{i64 1234}
+
+;--- src.o.thinlto.ll
+; ModuleID = 'src.o.thinlto.bc'
+source_filename = "src.o.thinlto.bc"
+
+^0 = module: (path: "src.o", hash: (2823430083, 3994560862, 899296057, 1055405378, 2961356784))
+^1 = gv: (guid: 15822663052811949562, summaries: (function: (module: ^0, flags: (linkage: external, visibility: default, notEligibleToImport: 0, live: 1, dsoLocal: 1, canAutoHide: 0, importType: definition), insts: 3, funcFlags: (readNone: 0, readOnly: 0, noRecurse: 0, returnDoesNotAlias: 0, noInline: 1, alwaysInline: 0, noUnwind: 0, mayThrow: 0, hasUnknownCall: 0, mustBeUnreachable: 0), callsites: ((callee: null, clones: (0), stackIds: (8632435727821051414)), (callee: null, clones: (1), stackIds: (15025054523792398438)), (callee: null, clones: (1), stackIds: (1234))))))
+^2 = flags: 353
+^3 = blockcount: 0

[snehasish]

Is it possible to assert that the type of the clone is the same as the original function here?

[teresajohnson]

I started to add such an assert, however, it seems unnecessary to me as we create NewF from the function type of CalledFunction just a few lines above here. However, looking at this made me concerned about the possibility of a different issue. Where we get CalledFunction we try various ways to find the function from the CallBase if getCalledFunction returns null. One of those things is to invoke stripPointerCasts. Now if we had a pointer cast here it would not be valid (pre-opaque pointers) to simply replace the called operand. But in theory, with opaque pointers, we should not actually find any pointer casts to strip. For testing, I added some checks in that code, as well as in the module summary analysis which does the same thing for all ThinLTO bitcode generation and kicks in much more broadly, and never found (with the regression tests or with a large app) any where it found anything to actually strip. But out of an abundance of caution I have added some handling above here to make sure the called function is the called operand (or its aliasee), and skip rewriting the call otherwise (also incrementing a new statistic in case we want to watch for cases this actually doing anything). In theory we could remove the stripPointerCasts call, but that needs to stay in sync with module summary analysis, so that's a change that could be considered later.

[snehasish]

lgtm

[snehasish]

I wonder if this could be more specific to the reason it was skipped? On the other hand, is this case the only place where we may skip calls during cloning?

[teresajohnson]

I started out making it longer and more specific, but figured we'd need to look up the reason in the code if we ever encounter this in the wild anyway (it's a developer-focused stat). Any suggestions for something more specific but not too terse? The other place we might skip is if it required ICP and we couldn't legally perform that. How about "Number of calls skipped during cloning due to unexpected operand" ?

[snehasish]

Maybe the ICP case is good to cover with this statistic too? Then we can keep the name and description as is.

Not a strong opinion, feel free to just update the description as you suggested as an alternative.

[teresajohnson]

I think I will update the description for this one as I suggested, since I'd like to know specifically if/when we hit this case. I'll add a separate stat for the other case(s) when I get a chance.