AMDGPU: Refactor lowering of s_barrier to split barriers (#154648)

Let's do the lowering of non-split into split barriers in a new IR pass,
AMDGPULowerIntrinsics. That way, there is no code duplication between
SelectionDAG and GlobalISel. This simplifies some upcoming extensions to
the code.

1 files changed, 4 insertions, 1 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
index e969f9ec8889..4a2f0a13b132 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -577,6 +577,7 @@ extern "C" LLVM_ABI LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUTarget() {
   initializeAMDGPURemoveIncompatibleFunctionsLegacyPass(*PR);
   initializeAMDGPULowerModuleLDSLegacyPass(*PR);
   initializeAMDGPULowerBufferFatPointersPass(*PR);
+  initializeAMDGPULowerIntrinsicsLegacyPass(*PR);
   initializeAMDGPUReserveWWMRegsLegacyPass(*PR);
   initializeAMDGPURewriteAGPRCopyMFMALegacyPass(*PR);
   initializeAMDGPURewriteOutArgumentsPass(*PR);
@@ -1418,6 +1419,7 @@ void AMDGPUPassConfig::addCodeGenPrepare() {
     // nodes out of the graph, which leads to function-level passes not
     // being run on them, which causes crashes in the resource usage analysis).
     addPass(createAMDGPULowerBufferFatPointersPass());
+    addPass(createAMDGPULowerIntrinsicsLegacyPass());
     // In accordance with the above FIXME, manually force all the
     // function-level passes into a CGSCCPassManager.
     addPass(new DummyCGSCCPass());
@@ -2155,9 +2157,10 @@ void AMDGPUCodeGenPassBuilder::addCodeGenPrepare(AddIRPass &addPass) const {
   // nodes out of the graph, which leads to function-level passes not
   // being run on them, which causes crashes in the resource usage analysis).
   addPass(AMDGPULowerBufferFatPointersPass(TM));
-
   addPass.requireCGSCCOrder();
 
+  addPass(AMDGPULowerIntrinsicsPass(TM));
+
   Base::addCodeGenPrepare(addPass);
 
   if (isPassEnabled(EnableLoadStoreVectorizer))