From 9704b35663db346c081ac02032cb4ab669a2a0f1 Mon Sep 17 00:00:00 2001
From: Alexey Sachkov <alexey.sachkov@intel.com>
Date: Tue, 12 May 2020 17:12:24 +0300
Subject: [PATCH] Improve handling of OpMemoryBarrier

Added support for non-constant 'Semantics' operand
---
 llvm-spirv/lib/SPIRV/SPIRVToOCL20.cpp         | 33 +++++++++++--------
 .../test/mem_fence_explicit_arguments.spt     | 32 ++++++++++++++++--
 2 files changed, 48 insertions(+), 17 deletions(-)

diff --git a/llvm-spirv/lib/SPIRV/SPIRVToOCL20.cpp b/llvm-spirv/lib/SPIRV/SPIRVToOCL20.cpp
index 6686d886e50f4..86eae4bcbd6b6 100644
--- a/llvm-spirv/lib/SPIRV/SPIRVToOCL20.cpp
+++ b/llvm-spirv/lib/SPIRV/SPIRVToOCL20.cpp
@@ -107,20 +107,25 @@ bool SPIRVToOCL20::runOnModule(Module &Module) {
 
 void SPIRVToOCL20::visitCallSPIRVMemoryBarrier(CallInst *CI) {
   AttributeList Attrs = CI->getCalledFunction()->getAttributes();
-  mutateCallInstOCL(M, CI,
-                    [=](CallInst *, std::vector<Value *> &Args) {
-                      auto GetArg = [=](unsigned I) {
-                        return cast<ConstantInt>(Args[I])->getZExtValue();
-                      };
-                      auto MScope = static_cast<Scope>(GetArg(0));
-                      auto Sema = mapSPIRVMemSemanticToOCL(GetArg(1));
-                      Args.resize(3);
-                      Args[0] = getInt32(M, Sema.first);
-                      Args[1] = getInt32(M, Sema.second);
-                      Args[2] = getInt32(M, rmap<OCLScopeKind>(MScope));
-                      return kOCLBuiltinName::AtomicWorkItemFence;
-                    },
-                    &Attrs);
+  mutateCallInstOCL(
+      M, CI,
+      [=](CallInst *, std::vector<Value *> &Args) {
+        Value *MemScope =
+            getInt32(M, rmap<OCLScopeKind>(static_cast<Scope>(
+                            cast<ConstantInt>(Args[0])->getZExtValue())));
+        Value *MemFenceFlags =
+            SPIRV::transSPIRVMemorySemanticsIntoOCLMemFenceFlags(Args[1], CI);
+        Value *MemOrder =
+            SPIRV::transSPIRVMemorySemanticsIntoOCLMemoryOrder(Args[1], CI);
+
+        Args.resize(3);
+        Args[0] = MemFenceFlags;
+        Args[1] = MemOrder;
+        Args[2] = MemScope;
+
+        return kOCLBuiltinName::AtomicWorkItemFence;
+      },
+      &Attrs);
 }
 
 void SPIRVToOCL20::visitCallSPIRVControlBarrier(CallInst *CI) {
diff --git a/llvm-spirv/test/mem_fence_explicit_arguments.spt b/llvm-spirv/test/mem_fence_explicit_arguments.spt
index 1518f7bf4b6ce..46838bcaec9fc 100644
--- a/llvm-spirv/test/mem_fence_explicit_arguments.spt
+++ b/llvm-spirv/test/mem_fence_explicit_arguments.spt
@@ -26,11 +26,15 @@
 ; RUN: llvm-spirv %s -to-binary -o %t.spv
 ; RUN: spirv-val %t.spv
 ; RUN: llvm-spirv -r %t.spv -o %t.bc
-; RUN: llvm-dis < %t.bc | FileCheck %s
+; RUN: llvm-dis < %t.bc | FileCheck %s --check-prefixes CHECK,CHECK-12
+; RUN: llvm-spirv -r %t.spv --spirv-target-env=CL2.0 -o %t.bc
+; RUN: llvm-dis < %t.bc | FileCheck %s --check-prefixes CHECK,CHECK-20
 
 ; CHECK: define spir_func void @test(i32 %val)
 ; CHECK: %call = call spir_func i32 @__translate_spirv_memory_fence(i32 %val)
-; CHECK: call spir_func void @_Z9mem_fencej(i32 %call)
+; CHECK-12: call spir_func void @_Z9mem_fencej(i32 %call)
+; CHECK-20: %call1 = call spir_func i32 @__translate_spirv_memory_order(i32 %val)
+; CHECK-20: call spir_func void @_Z22atomic_work_item_fencej12memory_order12memory_scope(i32 %call, i32 %call1
 ; CHECK: define private spir_func i32 @__translate_spirv_memory_fence(i32 %key)
 ; CHECK: entry:
 ; CHECK:   %key.masked = and i32 2816, %key
@@ -60,4 +64,26 @@
 ; CHECK: case.2816:
 ; CHECK:   ret i32 7
 ; CHECK: }
-; CHECK: declare spir_func void @_Z9mem_fencej(i32)
+; CHECK-20: define private spir_func i32 @__translate_spirv_memory_order(i32 %key) {
+; CHECK-20: entry:
+; CHECK-20:   %key.masked = and i32 30, %key
+; CHECK-20:   switch i32 %key.masked, label %default [
+; CHECK-20:     i32 0, label %case.0
+; CHECK-20:     i32 2, label %case.2
+; CHECK-20:     i32 4, label %case.4
+; CHECK-20:     i32 8, label %case.8
+; CHECK-20:     i32 16, label %case.16
+; CHECK-20:   ]
+; CHECK-20: default:
+; CHECK-20:   unreachable
+; CHECK-20: case.0:
+; CHECK-20:   ret i32 0
+; CHECK-20: case.2:
+; CHECK-20:   ret i32 2
+; CHECK-20: case.4:
+; CHECK-20:   ret i32 3
+; CHECK-20: case.8:
+; CHECK-20:   ret i32 4
+; CHECK-20: case.16:
+; CHECK-20:   ret i32 5
+; CHECK-20: }