Update GPUCompiler to fix unreachable control flow. (#1951)

Author: maleadt

Manifest.toml

@@ -111,9 +111,9 @@ version = "0.1.5"
 
 [[GPUCompiler]]
 deps = ["ExprTools", "InteractiveUtils", "LLVM", "Libdl", "Logging", "Scratch", "TimerOutputs", "UUIDs"]
-git-tree-sha1 = "cb090aea21c6ca78d59672a7e7d13bd56d09de64"
+git-tree-sha1 = "c47730aca2381f935a52fd732190e424c507230e"
 uuid = "61eb1bfa-7361-4325-ad38-22787b887f55"
-version = "0.20.3"
+version = "0.21.0"
 
 [[InteractiveUtils]]
 deps = ["Markdown"]

Project.toml

@@ -41,7 +41,7 @@ CUDA_Runtime_Discovery = "0.2"
 CUDA_Runtime_jll = "0.6"
 ExprTools = "0.1"
 GPUArrays = "8.6"
-GPUCompiler = "0.20"
+GPUCompiler = "0.21"
 KernelAbstractions = "0.9.2"
 LLVM = "5"
 Preferences = "1"

lib/cudadrv/state.jl

@@ -241,6 +241,13 @@ function context(dev::CuDevice)
               maxlog=1, _id=devidx)
     end
 
+    # warn about some known bugs
+    if runtime_version() < v"11.5" && capability(dev) < v"7"
+        @warn("""There are known codegen bugs on CUDA 11.4 and earlier for older GPUs like your $(name(dev)).
+                 Please use CUDA 11.5 or later, or switch to a different device.""",
+              maxlog=1, _id=devidx)
+    end
+
     # configure the primary context
     pctx = CuPrimaryContext(dev)
     ctx = CuContext(pctx)

src/compiler/compilation.jl

@@ -88,27 +88,11 @@ end
     # (we actually only need 6.2, but NVPTX doesn't support that)
     ptx = v"6.3"
 
-    # we need to take care emitting LLVM instructions like `unreachable`, which
-    # may result in thread-divergent control flow that older `ptxas` doesn't like.
-    # see e.g. JuliaGPU/CUDAnative.jl#4
-    unreachable = true
-    if cap < v"7" || runtime_version() < v"11.3"
-        unreachable = false
-    end
-
-    # there have been issues with emitting PTX `exit` instead of `trap` as well,
-    # see e.g. JuliaGPU/CUDA.jl#431 and NVIDIA bug #3231266 (but since switching
-    # to the toolkit's `ptxas` that specific machine/GPU now _requires_ exit...)
-    exitable = true
-    if cap < v"7"
-        exitable = false
-    end
-
     # NVIDIA bug #3600554: ptxas segfaults with our debug info, fixed in 11.7
     debuginfo = runtime_version() >= v"11.7"
 
     # create GPUCompiler objects
-    target = PTXCompilerTarget(; cap, ptx, debuginfo, unreachable, exitable, kwargs...)
+    target = PTXCompilerTarget(; cap, ptx, debuginfo, kwargs...)
     params = CUDACompilerParams()
     CompilerConfig(target, params; kernel, name, always_inline)
 end