[llvm] [SPIR-V] Align memory scopes with SPIR-V specification and LLVM expectations and change default mem scope value (PR #108528)

[Vyacheslav Levytskyy via llvm-commits]

https://github.com/VyacheslavLevytskyy created https://github.com/llvm/llvm-project/pull/108528

This PR is an attempt to align memory scopes with SPIR-V specification and LLVM requirements/expectations with respect to usage of "singlethread" / "" memory scopes. It also changes default mem scope value to the most conservative/safe choice that is the System.

This PR is to partially implement ideas and code review suggestions from https://github.com/llvm/llvm-project/pull/106429 and if this PR is to be merged, the merge must be postponed until after Clang-related changes from https://github.com/llvm/llvm-project/pull/106429 are merged to keep OpenCL default mem scope intact (see clang/lib/CodeGen/CGAtomic.cpp changes https://github.com/llvm/llvm-project/pull/106429 and https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#atomic-functions for definition of default mem scope in the OpenCL spec).

See https://github.com/llvm/llvm-project/pull/106429 for the detailed discussion.

>From 744d1c90710ef510d88d97918c716a71890025fc Mon Sep 17 00:00:00 2001
From: "Levytskyy, Vyacheslav" <vyacheslav.levytskyy at intel.com>
Date: Fri, 13 Sep 2024 03:32:13 -0700
Subject: [PATCH] another approach to name mem scope/use defaults

---
 .../Target/SPIRV/SPIRVInstructionSelector.cpp | 44 +++++++++----------
 1 file changed, 21 insertions(+), 23 deletions(-)

diff --git a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
index b526c9f29f1e6a..599954c9c5c0a8 100644
--- a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
@@ -36,19 +36,22 @@
 namespace {
 
 struct SyncScopeIDs {
-  llvm::SyncScope::ID Work_ItemSSID;
+  // Init SubGroupSSID with the reserved (invalid for SubGroupSSID) value to be
+  // able to reason about valid/non-initialized states of the structure.
+  llvm::SyncScope::ID SubGroupSSID = SyncScope::SingleThread;
   llvm::SyncScope::ID WorkGroupSSID;
   llvm::SyncScope::ID DeviceSSID;
-  llvm::SyncScope::ID AllSVMDevicesSSID;
-  llvm::SyncScope::ID SubGroupSSID;
 
   SyncScopeIDs() {}
   SyncScopeIDs(llvm::LLVMContext &Context) {
-    Work_ItemSSID = Context.getOrInsertSyncScopeID("work_item");
+    // Named by
+    // https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_scope_id
+    // We don't need aliases for Invocation and CrossDevice, as we already have
+    // them covered by "singlethread" and "" strings respectively (see
+    // implementation of LLVMContext::LLVMContext()).
+    SubGroupSSID = Context.getOrInsertSyncScopeID("subgroup");
     WorkGroupSSID = Context.getOrInsertSyncScopeID("workgroup");
     DeviceSSID = Context.getOrInsertSyncScopeID("device");
-    AllSVMDevicesSSID = Context.getOrInsertSyncScopeID("all_svm_devices");
-    SubGroupSSID = Context.getOrInsertSyncScopeID("sub_group");
   }
 };
 
@@ -305,7 +308,9 @@ void SPIRVInstructionSelector::setupMF(MachineFunction &MF, GISelKnownBits *KB,
                                        CodeGenCoverage *CoverageInfo,
                                        ProfileSummaryInfo *PSI,
                                        BlockFrequencyInfo *BFI) {
-  SSIDs = SyncScopeIDs(MF.getFunction().getContext());
+  // Init SPIR-V specific SyncScopeID's if needed.
+  if (SSIDs.SubGroupSSID == SyncScope::SingleThread)
+    SSIDs = SyncScopeIDs(MF.getFunction().getContext());
   MRI = &MF.getRegInfo();
   GR.setCurrentFunc(MF);
   InstructionSelector::setupMF(MF, KB, CoverageInfo, PSI, BFI);
@@ -847,25 +852,18 @@ bool SPIRVInstructionSelector::selectBitcast(Register ResVReg,
 
 static SPIRV::Scope::Scope getScope(SyncScope::ID Ord,
                                     const SyncScopeIDs &SSIDs) {
-  if (Ord == SyncScope::SingleThread || Ord == SSIDs.Work_ItemSSID)
+  if (Ord == SyncScope::SingleThread)
     return SPIRV::Scope::Invocation;
-  else if (Ord == SyncScope::System || Ord == SSIDs.DeviceSSID)
-    return SPIRV::Scope::Device;
-  else if (Ord == SSIDs.WorkGroupSSID)
-    return SPIRV::Scope::Workgroup;
-  else if (Ord == SSIDs.AllSVMDevicesSSID)
-    return SPIRV::Scope::CrossDevice;
-  else if (Ord == SSIDs.SubGroupSSID)
+  if (Ord == SSIDs.SubGroupSSID)
     return SPIRV::Scope::Subgroup;
-  else
-    // OpenCL approach is: "The functions that do not have memory_scope argument
-    // have the same semantics as the corresponding functions with the
-    // memory_scope argument set to memory_scope_device." See ref.: //
-    // https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#atomic-functions
-    // In our case if the scope is unknown, assuming that SPIR-V code is to be
-    // consumed in an OpenCL environment, we use the same approach and set the
-    // scope to memory_scope_device.
+  if (Ord == SSIDs.WorkGroupSSID)
+    return SPIRV::Scope::Workgroup;
+  if (Ord == SSIDs.DeviceSSID)
     return SPIRV::Scope::Device;
+  if (Ord == SyncScope::System)
+    return SPIRV::Scope::CrossDevice;
+  // Use the most conservative/safe choice as the default.
+  return SPIRV::Scope::CrossDevice;
 }
 
 static void addMemoryOperands(MachineMemOperand *MemOp,