Calling conv part 1 (#2782)

* Calling conv part 1

* fix

* more

* fix

* fix

* gopt

* more things

* gen mem unchecked

* fix

* recombine sret

* more fixes

* fix

* more improvements

* continuing

* fix

* fix

* 1.10 compile

* fix

* fix

* fix

* fix

* Shorten errors

* fix

* fix

* fix

* fix

* fix

Author: wsmoses

Project.toml

@@ -42,7 +42,7 @@ BFloat16s = "0.2, 0.3, 0.4, 0.5, 0.6"
 CEnum = "0.4, 0.5"
 ChainRulesCore = "1"
 EnzymeCore = "0.8.16"
-Enzyme_jll = "0.0.214, 0.0.215"
+Enzyme_jll = "0.0.216"
 GPUArraysCore = "0.1.6, 0.2"
 GPUCompiler = "1.6.2"
 LLVM = "9.1"

src/absint.jl

@@ -1,8 +1,8 @@
 # Abstractly interpret julia from LLVM
 
 # Return (bool if could interpret, julia object interpreted to)
-function absint(@nospecialize(arg::LLVM.Value), partial::Bool = false)::Tuple{Bool, Any}
-    if (value_type(arg) == LLVM.PointerType(LLVM.StructType(LLVMType[]), Tracked)) || (value_type(arg) == LLVM.PointerType(LLVM.StructType(LLVMType[]), Derived))
+function absint(@nospecialize(arg::LLVM.Value), partial::Bool = false, istracked::Bool=false)::Tuple{Bool, Any}
+    if (value_type(arg) == LLVM.PointerType(LLVM.StructType(LLVMType[]), Tracked)) || (value_type(arg) == LLVM.PointerType(LLVM.StructType(LLVMType[]), Derived)) || istracked
         ce, _ = get_base_and_offset(arg; offsetAllowed = false, inttoptr = true)
         if isa(ce, GlobalVariable)
             gname = LLVM.name(ce)
@@ -441,6 +441,12 @@ function abs_typeof(
             vals = absint(operands(arg)[3], partial)
             return (vals[1], vals[2], vals[1] ? GPUCompiler.BITS_REF : nothing)
         end
+        # Type tag is arg 3
+        if nm == "jl_alloc_genericmemory_unchecked" ||
+		nm == "ijl_alloc_genericmemory_unchecked"
+	    vals = absint(operands(arg)[3], partial, true)
+            return (vals[1], vals[2], vals[1] ? GPUCompiler.MUT_REF : nothing)
+        end
         # Type tag is arg 1
         if nm == "jl_alloc_array_1d" ||
                 nm == "ijl_alloc_array_1d" ||