diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
index 2b457fe519d96..c96625092a76c 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -793,6 +793,15 @@ void AMDGPUTargetMachine::registerPassBuilderCallbacks(
 
         PM.addPass(createCGSCCToFunctionPassAdaptor(std::move(FPM)));
       });
+
+  PB.registerFullLinkTimeOptimizationLastEPCallback(
+      [this](ModulePassManager &PM, OptimizationLevel Level) {
+        // We want to support the -lto-partitions=N option as "best effort".
+        // For that, we need to lower LDS earlier in the pipeline before the
+        // module is partitioned for codegen.
+        if (EnableLowerModuleLDS)
+          PM.addPass(AMDGPULowerModuleLDSPass(*this));
+      });
 }
 
 int64_t AMDGPUTargetMachine::getNullPointerValue(unsigned AddrSpace) {
diff --git a/llvm/test/CodeGen/AMDGPU/lto-lower-module-lds.ll b/llvm/test/CodeGen/AMDGPU/lto-lower-module-lds.ll
new file mode 100644
index 0000000000000..b813b8047bf24
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/lto-lower-module-lds.ll
@@ -0,0 +1,47 @@
+
+; Default O0
+; RUN: opt -mtriple=amdgcn-- -mcpu=gfx1030 %s -o %t.bc
+; RUN: llvm-lto2 run -O0 -cg-opt-level 0 %t.bc -o %t.s -r %t.bc,test,px -debug-pass-manager -debug-pass=Structure 2>&1 | FileCheck %s
+
+; Unified O0
+; RUN: opt -unified-lto -thinlto-split-lto-unit -thinlto-bc -mtriple=amdgcn-- -mcpu=gfx1030 %s -o %t.bc
+; RUN: llvm-lto2 run -unified-lto=full -O0 -cg-opt-level 0 %t.bc -o %t.s -r %t.bc,test,px -debug-pass-manager -debug-pass=Structure 2>&1 | FileCheck %s
+
+; Default O1
+; RUN: opt -mtriple=amdgcn-- -mcpu=gfx1030 %s -o %t.bc
+; RUN: llvm-lto2 run -O1 -cg-opt-level 1 %t.bc -o %t.s -r %t.bc,test,px -debug-pass-manager -debug-pass=Structure 2>&1 | FileCheck %s
+
+; Unified O1
+; RUN: opt -unified-lto -thinlto-split-lto-unit -thinlto-bc -mtriple=amdgcn-- -mcpu=gfx1030 %s -o %t.bc
+; RUN: llvm-lto2 run -unified-lto=full -O1 -cg-opt-level 1 %t.bc -o %t.s -r %t.bc,test,px -debug-pass-manager -debug-pass=Structure 2>&1 | FileCheck %s
+
+; Default O2
+; RUN: opt -mtriple=amdgcn-- -mcpu=gfx1030 %s -o %t.bc
+; RUN: llvm-lto2 run -O2 -cg-opt-level 2 %t.bc -o %t.s -r %t.bc,test,px -debug-pass-manager -debug-pass=Structure 2>&1 | FileCheck %s
+
+; Unified O2
+; RUN: opt -unified-lto -thinlto-split-lto-unit -thinlto-bc -mtriple=amdgcn-- -mcpu=gfx1030 %s -o %t.bc
+; RUN: llvm-lto2 run -unified-lto=full -O2 -cg-opt-level 2 %t.bc -o %t.s -r %t.bc,test,px -debug-pass-manager -debug-pass=Structure 2>&1 | FileCheck %s
+
+; Default O3
+; RUN: opt -mtriple=amdgcn-- -mcpu=gfx1030 %s -o %t.bc
+; RUN: llvm-lto2 run -O3 -cg-opt-level 3 %t.bc -o %t.s -r %t.bc,test,px -debug-pass-manager -debug-pass=Structure 2>&1 | FileCheck %s
+
+; Unified O3
+; RUN: opt -unified-lto -thinlto-split-lto-unit -thinlto-bc -mtriple=amdgcn-- -mcpu=gfx1030 %s -o %t.bc
+; RUN: llvm-lto2 run -unified-lto=full -O3 -cg-opt-level 3 %t.bc -o %t.s -r %t.bc,test,px -debug-pass-manager -debug-pass=Structure 2>&1 | FileCheck %s
+
+; First print will be from the New PM during the full LTO pipeline.
+; Second print will be from the legacy PM during the CG pipeline.
+
+; CHECK: Running pass: AMDGPULowerModuleLDSPass on [module]
+; CHECK: ModulePass Manager
+; CHECK:   Lower uses of LDS variables from non-kernel functions
+
+@lds = internal unnamed_addr addrspace(3) global i32 undef, align 4
+
+define amdgpu_kernel void @test() {
+entry:
+  store i32 1, ptr addrspace(3) @lds
+  ret void
+}