Mergefunc removes functions that are on llvm.used / llvm.compiler.used list

[tmiasko]

For example, running:

; ModuleID = 'a.ll'
source_filename = "a.38475855-cgu.0"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

@llvm.compiler.used = appending global [2 x i8*] [
  i8* bitcast (void ()* @f to i8*), 
  i8* bitcast (void ()* @g to i8*)
], section "llvm.metadata"

define void @f() unnamed_addr #0 {
start:
  ret void
}

define internal void @g() unnamed_addr #0 {
start:
  ret void
}

attributes #0 = { "target-cpu"="x86-64" }

through opt -S -mergefunc, results in g being removed, even though it was present on llvm.compiler.used.list:

; ModuleID = 'a.ll'
source_filename = "a.38475855-cgu.0"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

@llvm.compiler.used = appending global [2 x i8*] [i8* bitcast (void ()* @f to i8*), i8* bitcast (void ()* @f to i8*)], section "llvm.metadata"

define void @f() unnamed_addr #0 {
start:
  ret void
}

attributes #0 = { "target-cpu"="x86-64" }

[nikic]

I'm not sure this is a bug, as long as it doesn't happen for cases where @g is observably different, e.g. by having a specific section assigned.

[tmiasko]

Consider the example extended with module asm "call g" (or any other reference to the symbol that compiler cannot see).

[tmiasko]

The issue seems to be limited to non-interposable functions with insignificant address, for which mergefunc directly replaces all uses of one function with another:

llvm-project/llvm/lib/Transforms/IPO/MergeFunctions.cpp

Lines 827 to 835 in 6c2a01c

	if (!G->isInterposable() && !MergeFunctionsPDI) {
	if (G->hasGlobalUnnamedAddr()) {
	// G might have been a key in our GlobalNumberState, and it's illegal
	// to replace a key in ValueMap<GlobalValue *> with a non-global.
	GlobalNumbers.erase(G);
	// If G's address is not significant, replace it entirely.
	Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
	removeUsers(G);
	G->replaceAllUsesWith(BitcastF);

This is incorrect if function was on any of llvm used lists. The occurrence on the llvm used list will be replaced, and as a result the original symbol is no longer being retained (in fact mergefunc itself removes the replaced function a few lines later). At the same time references to the replaced symbol from within module asm or inline asm are left unmodified and undefined.

[tmiasko]

Fixed in https://reviews.llvm.org/D127751.