[llvm] [ElimAvailExtern] Add an option to allow to convert global variables in a specified address space to local (PR #144287)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: Shilei Tian (shiltian)

<details>
<summary>Changes</summary>

Currently, the `EliminateAvailableExternallyPass` only converts certain
available externally functions to local. For global variables, it only drops
their initializers.

This PR adds an option to allow the pass to convert global variables in a
specified address space to local. The motivation for this change is to correctly
support lowering of LDS variables (`__shared__` variables, in more generic
terminology) when ThinLTO is enabled for AMDGPU.

A `__shared__` variable is lowered to a hidden global variable in a particular
address space by the frontend, which is roughly same as a `static` local
variable. To properly lower it in the backend, the compiler needs to check all
its uses. Enabling ThinLTO currently breaks this when a function containing a
`__shared__` variable is imported from another module. Even though the global
variable is imported along with its associated function, and the function is
privatized by the `EliminateAvailableExternallyPass`, the global variable itself
is not.

It's safe to privatize such global variables, because they're _local_ to their
associated functions. If the function itself is privatized, its associated
global variables should also be privatized accordingly.

---
Full diff: https://github.com/llvm/llvm-project/pull/144287.diff


2 Files Affected:

- (modified) llvm/lib/Transforms/IPO/ElimAvailExtern.cpp (+30-1) 
- (added) llvm/test/Transforms/EliminateAvailableExternally/convert-global-variables-to-local.ll (+21) 


``````````diff
diff --git a/llvm/lib/Transforms/IPO/ElimAvailExtern.cpp b/llvm/lib/Transforms/IPO/ElimAvailExtern.cpp
index 718452fc02764..a015ba8ccfd4a 100644
--- a/llvm/lib/Transforms/IPO/ElimAvailExtern.cpp
+++ b/llvm/lib/Transforms/IPO/ElimAvailExtern.cpp
@@ -35,8 +35,14 @@ static cl::opt<bool> ConvertToLocal(
     cl::desc("Convert available_externally into locals, renaming them "
              "to avoid link-time clashes."));
 
+static cl::opt<unsigned> ConvertGlobalVariableInAddrSpace(
+    "avail-extern-gv-in-addrspace-to-local", cl::Hidden,
+    cl::desc(
+        "Convert available_externally global variables into locals if they are "
+        "in specificed addrspace, renaming them to avoid link-time clashes."));
+
 STATISTIC(NumRemovals, "Number of functions removed");
-STATISTIC(NumConversions, "Number of functions converted");
+STATISTIC(NumConversions, "Number of functions and globalbs converted");
 STATISTIC(NumVariables, "Number of global variables removed");
 
 void deleteFunction(Function &F) {
@@ -88,9 +94,32 @@ static void convertToLocalCopy(Module &M, Function &F) {
   ++NumConversions;
 }
 
+static void convertToLocalCopy(Module &M, GlobalValue &GV) {
+  assert(GV.hasAvailableExternallyLinkage());
+  std::string OrigName = GV.getName().str();
+  std::string NewName = OrigName + ".__uniq" + getUniqueModuleId(&M);
+  GV.setName(NewName);
+  GV.setLinkage(GlobalValue::InternalLinkage);
+  ++NumConversions;
+}
+
 static bool eliminateAvailableExternally(Module &M, bool Convert) {
   bool Changed = false;
 
+  // Convert global variables in specified address space before changing it to
+  // external linkage below.
+  if (ConvertGlobalVariableInAddrSpace.getNumOccurrences()) {
+    for (GlobalVariable &GV : M.globals()) {
+      if (!GV.hasAvailableExternallyLinkage() || GV.use_empty())
+        continue;
+
+      if (GV.getAddressSpace() == ConvertGlobalVariableInAddrSpace)
+        convertToLocalCopy(M, GV);
+
+      Changed = true;
+    }
+  }
+
   // Drop initializers of available externally global variables.
   for (GlobalVariable &GV : M.globals()) {
     if (!GV.hasAvailableExternallyLinkage())
diff --git a/llvm/test/Transforms/EliminateAvailableExternally/convert-global-variables-to-local.ll b/llvm/test/Transforms/EliminateAvailableExternally/convert-global-variables-to-local.ll
new file mode 100644
index 0000000000000..6995b97e79887
--- /dev/null
+++ b/llvm/test/Transforms/EliminateAvailableExternally/convert-global-variables-to-local.ll
@@ -0,0 +1,21 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals all --version 5
+; RUN: opt -S -passes=elim-avail-extern -avail-extern-gv-in-addrspace-to-local=3 %s -o - | FileCheck %s
+
+ at shared = internal addrspace(3) global i32 undef, align 4
+ at shared.imported = available_externally hidden unnamed_addr addrspace(3) global i32 undef, align 4
+
+;.
+; CHECK: @shared = internal addrspace(3) global i32 undef, align 4
+; CHECK: @shared.imported.__uniq.[[UUID:.*]] = internal unnamed_addr addrspace(3) global i32 undef, align 4
+;.
+define void @foo(i32 %v) {
+; CHECK-LABEL: define void @foo(
+; CHECK-SAME: i32 [[V:%.*]]) {
+; CHECK-NEXT:    store i32 [[V]], ptr addrspace(3) @shared, align 4
+; CHECK-NEXT:    store i32 [[V]], ptr addrspace(3) @shared.imported.__uniq.[[UUID]], align 4
+; CHECK-NEXT:    ret void
+;
+  store i32 %v, ptr addrspace(3) @shared, align 4
+  store i32 %v, ptr addrspace(3) @shared.imported, align 4
+  ret void
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/144287