From 268be7f0b5012031e5fa0c435b3dd5bd44f3b230 Mon Sep 17 00:00:00 2001
From: Evghenii <egaburov@dds.nl>
Date: Wed, 13 Nov 2013 11:19:10 +0100
Subject: [PATCH] fixed ISPCSync functionality

---
 builtins/target-nvptx64.ll     |  3 ++-
 ctx.cpp                        | 48 ++++++++++++++++++++++------------
 examples_cuda/aobench/ao1.ispc |  1 -
 3 files changed, 33 insertions(+), 19 deletions(-)

diff --git a/builtins/target-nvptx64.ll b/builtins/target-nvptx64.ll
index 2f2f004a..2484eb1e 100644
--- a/builtins/target-nvptx64.ll
+++ b/builtins/target-nvptx64.ll
@@ -102,7 +102,8 @@ define i32 @__lanemask_lt() nounwind readnone alwaysinline
 
 define i8* @ISPCAlloc(i8**, i64, i32) nounwind alwaysinline
 {
-  ret i8* null
+  %ptr = inttoptr i64 1 to i8*
+  ret i8* %ptr
 }
 
 ;; this call allocate parameter buffer for kernel launch
diff --git a/ctx.cpp b/ctx.cpp
index f33044f8..1ce6588e 100644
--- a/ctx.cpp
+++ b/ctx.cpp
@@ -3673,29 +3673,43 @@ FunctionEmitContext::LaunchInst(llvm::Value *callee,
 
 void
 FunctionEmitContext::SyncInst() {
-    llvm::Value *launchGroupHandle = LoadInst(launchGroupHandlePtr);
-    llvm::Value *nullPtrValue =
+    if (!g->target->isPTX())
+    {
+      llvm::Value *launchGroupHandle = LoadInst(launchGroupHandlePtr);
+      llvm::Value *nullPtrValue =
         llvm::Constant::getNullValue(LLVMTypes::VoidPointerType);
-    llvm::Value *nonNull = CmpInst(llvm::Instruction::ICmp,
-                                   llvm::CmpInst::ICMP_NE,
-                                   launchGroupHandle, nullPtrValue);
-    llvm::BasicBlock *bSync = CreateBasicBlock("call_sync");
-    llvm::BasicBlock *bPostSync = CreateBasicBlock("post_sync");
-    BranchInst(bSync, bPostSync, nonNull);
+      llvm::Value *nonNull = CmpInst(llvm::Instruction::ICmp,
+          llvm::CmpInst::ICMP_NE,
+          launchGroupHandle, nullPtrValue);
+      llvm::BasicBlock *bSync = CreateBasicBlock("call_sync");
+      llvm::BasicBlock *bPostSync = CreateBasicBlock("post_sync");
+      BranchInst(bSync, bPostSync, nonNull);
 
-    SetCurrentBasicBlock(bSync);
-    llvm::Function *fsync = m->module->getFunction("ISPCSync");
-    if (fsync == NULL)
+      SetCurrentBasicBlock(bSync);
+      llvm::Function *fsync = m->module->getFunction("ISPCSync");
+      if (fsync == NULL)
         FATAL("Couldn't find ISPCSync declaration?!");
-    CallInst(fsync, NULL, launchGroupHandle, "");
+      CallInst(fsync, NULL, launchGroupHandle, "");
 
-    // zero out the handle so that if ISPCLaunch is called again in this
-    // function, it knows it's starting out from scratch
-    StoreInst(nullPtrValue, launchGroupHandlePtr);
+      // zero out the handle so that if ISPCLaunch is called again in this
+      // function, it knows it's starting out from scratch
+      StoreInst(nullPtrValue, launchGroupHandlePtr);
 
-    BranchInst(bPostSync);
+      BranchInst(bPostSync);
 
-    SetCurrentBasicBlock(bPostSync);
+      SetCurrentBasicBlock(bPostSync);
+    }
+    else
+    {
+      llvm::Value *launchGroupHandle = LoadInst(launchGroupHandlePtr);
+      llvm::Value *nullPtrValue =
+        llvm::Constant::getNullValue(LLVMTypes::VoidPointerType);
+      llvm::Function *fsync = m->module->getFunction("ISPCSync");
+      if (fsync == NULL)
+        FATAL("Couldn't find ISPCSync declaration?!");
+      CallInst(fsync, NULL, launchGroupHandle, "");
+      StoreInst(nullPtrValue, launchGroupHandlePtr);
+    }
 }
 
 
diff --git a/examples_cuda/aobench/ao1.ispc b/examples_cuda/aobench/ao1.ispc
index a6ed8c85..cccb3d7b 100644
--- a/examples_cuda/aobench/ao1.ispc
+++ b/examples_cuda/aobench/ao1.ispc
@@ -313,5 +313,4 @@ export void ao_ispc_tasks(uniform int w, uniform int h, uniform int nsubsamples,
   const uniform int ntilex = (w+TILEX-1)/TILEX;
   const uniform int ntiley = (h+TILEY-1)/TILEY;
   launch[ntilex,ntiley] ao_task(w, h, nsubsamples, image);
-  sync;
 }