[SYCL][CUDA] Add IPSCCP pass to O0 by default

llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp

@@ -33,6 +33,7 @@
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Target/TargetMachine.h"
 #include "llvm/Target/TargetOptions.h"
+#include "llvm/Transforms/IPO.h"
 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
 #include "llvm/Transforms/Scalar.h"
 #include "llvm/Transforms/Scalar/GVN.h"
@@ -63,6 +64,12 @@ static cl::opt<bool> UseShortPointersOpt(
         "Use 32-bit pointers for accessing const/local/shared address spaces."),
     cl::init(false), cl::Hidden);
 
+static cl::opt<bool>
+    UseIPSCCPO0("use-ipsccp-nvptx-O0",
+                cl::desc("Use IPSCCP pass at O0 as a temp solution for "
+                         "nvvm-reflect dead-code errors."),
+                cl::init(true), cl::Hidden);
+
 namespace llvm {
 
 void initializeLocalAccessorToSharedMemoryPass(PassRegistry &);
@@ -327,6 +334,10 @@ void NVPTXPassConfig::addIRPasses() {
   const NVPTXSubtarget &ST = *getTM<NVPTXTargetMachine>().getSubtargetImpl();
   addPass(createNVVMReflectPass(ST.getSmVersion()));
 
+  if (getOptLevel() == CodeGenOpt::None && UseIPSCCPO0) {
+    addPass(createIPSCCPPass());
+  }
+
   // FIXME: should the target triple check be done by the pass itself?
   // See createNVPTXLowerArgsPass as an example
   if (getTM<NVPTXTargetMachine>().getTargetTriple().getOS() == Triple::CUDA) {

llvm/test/CodeGen/NVPTX/param-load-store.ll

@@ -1,5 +1,5 @@
 ; Verifies correctness of load/store of parameters and return values.
-; RUN: llc < %s -march=nvptx64 -mcpu=sm_35 -O0 -verify-machineinstrs | FileCheck -allow-deprecated-dag-overlap %s
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_35 -O0 -verify-machineinstrs -use-ipsccp-nvptx-O0=false | FileCheck -allow-deprecated-dag-overlap %s
 
 %s_i1 = type { i1 }
 %s_i8 = type { i8 }

sycl/doc/GetStartedGuide.md

@@ -825,6 +825,18 @@ which contains all the symbols required.
   significantly slower but matches the default precision used by `nvcc`, and
   this `clang++` flag is equivalent to the `nvcc` `-prec-sqrt` flag, except that
   it defaults to `false`.
+* No Opt (O0) uses the IPSCCP compiler pass by default, although the IPSCCP pass
+  can be switched off at O0 using the `-mllvm -use-ipsccp-nvptx-O0=false` flag at
+  the user's discretion.
+  The reason that the IPSCCP pass is used by default even at O0 is that there is
+  currently an unresolved issue with the nvvm-reflect compiler pass: This pass is
+  used to pick the correct branches depending on the SM version which can be
+  optionally specified by the `--cuda-gpu-arch` flag.
+  If the arch flag is not specified by the user, the default value, SM 50, is used.
+  Without the execution of the IPSCCP pass at -O0 when using a low SM version,
+  dead instructions which require a higher SM version can remain. Since
+  corresponding issues occur in other backends future work will aim for a
+  universal solution to these issues.
 
 ### HIP back-end limitations