CUDA: try more float alwaysinline (#1265)

* CUDA: try more float alwaysinline

* f64

Author: wsmoses

ext/ReactantCUDAExt.jl

@@ -106,41 +106,45 @@ end
 
 @inline function Base.convert(CT::Type{CuTracedRNumber{Float64,1}}, x::Number)
     return CT(
+	   Base.reinterpret(Core.LLVMPtr{Float64,1},
         Base.llvmcall(
             (
-                """define double addrspace(1)* @entry(double %d) alwaysinline {
+                """define i8 addrspace(1)* @entry(double %d) alwaysinline {
           %a = alloca double
           store atomic double %d, double* %a release, align 8
-          %ac = addrspacecast double* %a to double addrspace(1)*
-          ret double addrspace(1)* %ac
+	  %bc = bitcast double* %a to i8*
+          %ac = addrspacecast i8* %bc to i8 addrspace(1)*
+          ret i8 addrspace(1)* %ac
                     }
       """,
                 "entry",
             ),
-            Core.LLVMPtr{Float64,1},
+            Core.LLVMPtr{UInt8,1},
             Tuple{Float64},
             Base.convert(Float64, x),
-        ),
+        ))
     )
 end
 
 @inline function Base.convert(CT::Type{CuTracedRNumber{Float32,1}}, x::Number)
     return CT(
+	   Base.reinterpret(Core.LLVMPtr{Float32,1},
         Base.llvmcall(
             (
-                """define float addrspace(1)* @entry(float %d) alwaysinline {
+                """define i8 addrspace(1)* @entry(float %d) alwaysinline {
           %a = alloca float
           store atomic float %d, float* %a release, align 4
-          %ac = addrspacecast float* %a to float addrspace(1)*
-          ret float addrspace(1)* %ac
+	  %bc = bitcast float* %a to i8*
+          %ac = addrspacecast i8* %bc to i8 addrspace(1)*
+          ret i8 addrspace(1)* %ac
                     }
       """,
                 "entry",
             ),
-            Core.LLVMPtr{Float32,1},
+            Core.LLVMPtr{UInt8,1},
             Tuple{Float32},
             Base.convert(Float32, x),
-        ),
+	   ))
     )
 end
 
@@ -908,6 +912,14 @@ function compile(job)
         if Reactant.Compiler.DUMP_LLVMIR[]
             println("cuda.jl pre vendor IR\n", string(mod))
         end
+
+	LLVM.@dispose pb = LLVM.NewPMPassBuilder() begin
+	   LLVM.add!(pb, LLVM.NewPMModulePassManager()) do mpm
+	     LLVM.add!(mpm, LLVM.AlwaysInlinerPass())
+	   end
+	   LLVM.run!(pb, mod, tm)
+	end
+
         vendored_optimize_module!(job, mod)
         if Reactant.Compiler.DUMP_LLVMIR[]
             println("cuda.jl post vendor IR\n", string(mod))