From 06313e0ec396c03ba557b1006c30464260a800b9 Mon Sep 17 00:00:00 2001
From: Evghenii <egaburov@dds.nl>
Date: Wed, 22 Jan 2014 22:12:51 +0100
Subject: [PATCH] exclusive_scan_and is supported, but must be called outside
 if-statements. in pricniple other must do the same

---
 builtins/target-nvptx.ll        |  9 ++++-----
 examples_ptx/ptxcc/ptxcc        |  2 +-
 tests/exclusive-scan-and-1.ispc |  8 ++++++++
 tests/exclusive-scan-and-2.ispc | 10 ++++++++++
 4 files changed, 23 insertions(+), 6 deletions(-)

diff --git a/builtins/target-nvptx.ll b/builtins/target-nvptx.ll
index d631f35d..840134c8 100644
--- a/builtins/target-nvptx.ll
+++ b/builtins/target-nvptx.ll
@@ -1210,14 +1210,13 @@ define <1 x i32> @__exclusive_scan_and_i32(<1 x i32>, <1 x i1>) nounwind readnon
 {
   %shft = tail call <1 x i32> @__shift_i32(<1 x i32> %0, i32 -1)
   %v0   = extractelement <1 x i32> %shft, i32 0
-  %m0   = extractelement <1 x i1 > %1,    i32 0
+  %mask = extractelement <1 x i1 > %1, i32 0
+  %v1   = select i1 %mask, i32 %v0, i32 -1
 
   %tid  = tail call i32 @__tid_x()
   %lane = and i32 %tid, 31
-  %m1   = icmp eq i32 %lane, 0
-
-  %mask = and i1 %m0, %m1
-  %v    = select i1 %mask, i32 %v0, i32 -1
+  %c    = icmp eq i32 %lane, 0
+  %v    = select i1 %c, i32 -1, i32 %v1
   
   %s1 = tail call i32 @__shfl_scan_and_step_i32(i32 %v,  i32  1);
   %s2 = tail call i32 @__shfl_scan_and_step_i32(i32 %s1, i32  2);
diff --git a/examples_ptx/ptxcc/ptxcc b/examples_ptx/ptxcc/ptxcc
index 0f7e384e..ae0fca91 100755
--- a/examples_ptx/ptxcc/ptxcc
+++ b/examples_ptx/ptxcc/ptxcc
@@ -10,7 +10,7 @@ DEPTX=dePTX
 NVCC=nvcc
 
 $DEPTX < $PTXSRC > $PTXCU &&
-$NVCC -arch=sm_35 -dc $NVCCPARM -dryrun $PTXCU 2>&1 | \
+$NVCC -arch=sm_35 -G -dc $NVCCPARM -dryrun $PTXCU 2>&1 | \
   sed 's/\#\$//g'| \
   awk '{ if ($1 == "LIBRARIES=") print $1$2; else if ($1 == "cicc") print "cp '$PTXSRC'", $NF; else print $0 }' > $PTXSH &&
 sh $PTXSH
diff --git a/tests/exclusive-scan-and-1.ispc b/tests/exclusive-scan-and-1.ispc
index 31347b47..525d0c2a 100644
--- a/tests/exclusive-scan-and-1.ispc
+++ b/tests/exclusive-scan-and-1.ispc
@@ -4,9 +4,17 @@ export uniform int width() { return programCount; }
 export void f_f(uniform float RET[], uniform float aFOO[]) {
     RET[programIndex] = -1;
     int32 a = (programIndex & 1) ? 0xff : 0;
+#if 0
     if (programIndex & 1) {
         RET[programIndex] = exclusive_scan_and(a);
     }
+#else
+    const bool mask = programIndex & 1;
+    const float res = exclusive_scan_and(mask ? a : -1);
+    if (mask) {
+        RET[programIndex] = res;
+    }
+#endif
 }
 
 
diff --git a/tests/exclusive-scan-and-2.ispc b/tests/exclusive-scan-and-2.ispc
index b742a91e..7fd69648 100644
--- a/tests/exclusive-scan-and-2.ispc
+++ b/tests/exclusive-scan-and-2.ispc
@@ -4,9 +4,19 @@ export uniform int width() { return programCount; }
 export void f_f(uniform float RET[], uniform float aFOO[]) {
     RET[programIndex] = -1;
     int32 a = ~(1ul << programIndex);
+#if 0
     if ((programIndex < 32) && (programIndex & 1) == 0) {
         RET[programIndex] = exclusive_scan_and(a);
     }
+#else
+    const bool mask = ((programIndex < 32) && (programIndex & 1) == 0);
+    const float res = exclusive_scan_and(mask ? a : -1);
+    if (mask)
+    {
+        RET[programIndex] = res;
+    }
+#endif
+
 }