From 5a8351d7ea38de832162e726de3b214f22dd3d01 Mon Sep 17 00:00:00 2001
From: Evghenii <egaburov@dds.nl>
Date: Fri, 24 Jan 2014 09:22:55 +0100
Subject: [PATCH] added varying new/delete

---
 builtins/util-nvptx.m4 | 54 +++++++++++++++++++++++++++++++++++++++---
 expr.cpp               |  2 +-
 tests/soa-22.ispc      |  4 +++-
 3 files changed, 55 insertions(+), 5 deletions(-)

diff --git a/builtins/util-nvptx.m4 b/builtins/util-nvptx.m4
index 62b2391c..804c07e3 100644
--- a/builtins/util-nvptx.m4
+++ b/builtins/util-nvptx.m4
@@ -2204,10 +2204,58 @@ if.end:                                           ; preds = %if.then, %entry
   ret void
 }
 
-declare <WIDTH x i64> @__new_varying32_64rt(<WIDTH x i32> %size, <WIDTH x MASK> %mask);
-declare <WIDTH x i64> @__new_varying64_64rt(<WIDTH x i64> %size, <WIDTH x MASK> %mask);
-declare void @__delete_varying_64rt(<WIDTH x i64> %ptr, <WIDTH x MASK> %mask);
+define <1 x i64> @__new_varying32_64rt(<1 x i32> %sizev, <1 x i1> %maskv)
+{
+entry:
+  %size32 = extractelement <1 x i32> %sizev, i32 0
+  %mask   = extractelement <1 x  i1> %maskv, i32 0
+  %size64 = zext i32 %size32 to i64
+  br i1 %mask, label %alloc, label %skip
 
+alloc:
+  %ptr   = tail call noalias i8* @malloc(i64 %size64) 
+  %addr1 = ptrtoint i8* %ptr to i64
+  br label %skip
+
+skip:
+  %addr64 = phi i64 [ %addr1, %alloc], [ 0, %entry ]
+  %addr   = insertelement <1 x i64> undef, i64 %addr64, i32 0
+  ret <1 x i64> %addr
+}
+
+define <1 x i64> @__new_varying64_64rt(<1 x i64> %sizev, <1 x i1> %maskv)
+{
+entry:
+  %size64 = extractelement <1 x i64> %sizev, i32 0
+  %mask   = extractelement <1 x  i1> %maskv, i32 0
+  br i1 %mask, label %alloc, label %skip
+
+alloc:
+  %ptr   = tail call noalias i8* @malloc(i64 %size64) 
+  %addr1 = ptrtoint i8* %ptr to i64
+  br label %skip
+
+skip:
+  %addr64 = phi i64 [ %addr1, %alloc], [ 0, %entry ]
+  %addr   = insertelement <1 x i64> undef, i64 %addr64, i32 0
+  ret <1 x i64> %addr
+}
+
+define void @__delete_varying_64rt(<1 x i64> %ptrv, <1 x i1> %maskv)
+{
+entry:
+  %addr64 = extractelement <1 x i64> %ptrv,  i32 0
+  %mask   = extractelement <1 x  i1> %maskv, i32 0
+  br i1 %mask, label %free, label %skip
+
+free:
+  %ptr = inttoptr i64 %addr64 to i8*
+  tail call void @free(i8* %ptr) 
+  br label %skip
+
+skip:
+  ret void
+}
 ', `
 errprint(`RUNTIME should be defined to either 32 or 64
 ')
diff --git a/expr.cpp b/expr.cpp
index e7e99b83..bfa14d22 100644
--- a/expr.cpp
+++ b/expr.cpp
@@ -8654,7 +8654,7 @@ NewExpr::TypeCheck() {
     }
     if (g->target->getISA() == Target::NVPTX && allocType->IsVaryingType())
     {
-      Error(pos, "\"sizeof\" with varying data types is not yet supported with \"nvptx\" target.");
+      Error(pos, "\"new\" with varying data types is not yet supported with \"nvptx\" target.");
       return NULL;
     }
     if (CastType<UndefinedStructType>(allocType) != NULL) {
diff --git a/tests/soa-22.ispc b/tests/soa-22.ispc
index 60448694..4d8fd299 100644
--- a/tests/soa-22.ispc
+++ b/tests/soa-22.ispc
@@ -25,8 +25,10 @@ export void f_fu(uniform float RET[], uniform float aFOO[], uniform float b) {
             }
         }
     }
-    
+   
+#ifndef __NVPTX__ 
     assert(programIndex < 80);
+#endif
     RET[programIndex] = pts[programIndex].pts[programIndex % 3][programIndex % 4].z;
 }