[NFC] Manually optimize writes (#346)

This manually applies an optimization that Clang performs to see if I
can trip up the NV driver through DXC as well as Clang.

Author: llvm-beanz

test/Feature/WaveOps/WaveIsFirstLane.test

@@ -4,17 +4,21 @@ RWStructuredBuffer<uint> Out : register(u1);
 
 [numthreads(4, 1, 1)]
 void main(uint3 threadID : SV_DispatchThreadID) { 
+  uint tmp = 0xFF;
   switch (value[threadID.x]) {
     case 0:
-      Out[threadID.x] = WaveIsFirstLane(); // threads 0 and 1; 0 is first for both
+      tmp = WaveIsFirstLane(); // threads 0 and 1; 0 is first for both
       break;
     case 2:
-      Out[threadID.x] = WaveIsFirstLane(); // thread 3; 3 is first
+      tmp = WaveIsFirstLane(); // thread 3; 3 is first
       break;
     default:
-      Out[threadID.x] = WaveIsFirstLane(); // thread 2; 2 is first
+      tmp = WaveIsFirstLane(); // thread 2; 2 is first
       break;
   }
+  // Using a temporary value that is wave-divergent here seems to hit a driver
+  // bug in the NV 50-series GPUs
+  Out[threadID.x] = tmp;
   Out[threadID.x + 4] = WaveIsFirstLane(); // 0 is first for all
 }
 
@@ -64,6 +68,8 @@ DescriptorSets:
 # https://github.com/llvm/llvm-project/issues/145513
 # XFAIL: Clang-Vulkan
 
+# XFAIL: NV-Reconvergence-Issue-320
+
 # RUN: split-file %s %t
 # RUN: %dxc_target -T cs_6_5 -Fo %t.o %t/source.hlsl
 # RUN: %offloader %t/pipeline.yaml %t.o

test/lit.cfg.py

@@ -63,6 +63,10 @@ def setDeviceFeatures(config, device, compiler):
             config.available_features.add("Intel-Memory-Coherence-Issue-226")
     if "NVIDIA" in device["Description"]:
         config.available_features.add("%s-NV" % API)
+        NV50SeriesRegex = re.compile("NVIDIA GeForce [A-Z]+ 50[0-9]+")
+        NV50SeriesMatch = NV50SeriesRegex.match(device["Description"])
+        if NV50SeriesMatch and API == "DirectX":
+            config.available_features.add("NV-Reconvergence-Issue-320")
     if "AMD" in device["Description"]:
         config.available_features.add("%s-AMD" % API)