continue

Author: wsmoses

src/llvm/transforms.jl

@@ -634,7 +634,7 @@ function nodecayed_phis!(mod::LLVM.Module)
                     push!(todo, inst)
                     nb = IRBuilder()
                     position!(nb, inst)
-                    el_ty = if addr == 11
+                    el_ty = if addr == 11 && !LLVM.is_opaque(ty)
                         eltype(ty)
                     else
                         LLVM.StructType(LLVM.LLVMType[])
@@ -656,7 +656,7 @@ function nodecayed_phis!(mod::LLVM.Module)
 
                 for inst in todo
                     ty = value_type(inst)
-                    el_ty = if addr == 11
+                    el_ty = if addr == 11 && !LLVM.is_opaque(ty)
                         eltype(ty)
                     else
                         LLVM.StructType(LLVM.LLVMType[])
@@ -893,15 +893,17 @@ function nodecayed_phis!(mod::LLVM.Module)
                                     offset,
                                     API.EnzymeComputeByteOffsetOfGEP(b, v, offty),
                                 )
-                                v2 = bitcast!(
-                                    b,
-                                    v2,
-                                    LLVM.PointerType(
-                                        eltype(value_type(v)),
-                                        addrspace(value_type(v2)),
-                                    ),
-                                )
-                                @assert eltype(value_type(v2)) == eltype(value_type(v))
+                                if !LLVM.is_opaque(value_type(v2))
+                                    v2 = bitcast!(
+                                        b,
+                                        v2,
+                                        LLVM.PointerType(
+                                            eltype(value_type(v)),
+                                            addrspace(value_type(v2)),
+                                        ),
+                                    )
+                                    @assert eltype(value_type(v2)) == eltype(value_type(v))
+                                end
                                 return v2, offset, skipload
                             end
 
@@ -1024,7 +1026,7 @@ function nodecayed_phis!(mod::LLVM.Module)
                             @assert hadload
                         end
 
-                        if eltype(value_type(v)) != el_ty
+                        if !LLVM.is_opaque(value_type(v)) && eltype(value_type(v)) != el_ty
                             v = bitcast!(
                                 b,
                                 v,
@@ -1772,14 +1774,23 @@ function propagate_returned!(mod::LLVM.Module)
                             eraseInst(LLVM.parent(c), c)
                         end
                         B = IRBuilder()
+
                         position!(B, first(instructions(first(blocks(fn)))))
 
-                        argeltype = sret_ty(fn, i)
-                        al = alloca!(B, argeltype)
-                        if value_type(al) != value_type(arg)
-                            al = addrspacecast!(B, al, value_type(arg))
+                        has_use = false
+                        for _ in LLVM.uses(arg)
+                            has_use = true
+                            break
+                        end
+
+                        if has_use
+                            argeltype = sret_ty(fn, i)
+                            al = alloca!(B, argeltype)
+                            if value_type(al) != value_type(arg)
+                                al = addrspacecast!(B, al, value_type(arg))
+                            end
+                            LLVM.replace_uses!(arg, al)
                         end
-                        LLVM.replace_uses!(arg, al)
                     end
                 end
 

src/rules/customrules.jl

@@ -301,7 +301,12 @@ function enzyme_custom_setup_args(
                         LLVM.ConstantInt(LLVM.IntType(32), 0),
                     ],
                 )
-                if value_type(val) != eltype(value_type(ptr))
+                
+                if !is_opaque(value_type(ptr))
+                    @assert eltype(value_type(ptr)) == arty
+                end
+
+                if value_type(val) != arty
                     val = load!(B, arty, val)
                 end
                 store!(B, val, ptr)
@@ -353,33 +358,20 @@ function enzyme_custom_setup_args(
                         )
                     end
 
-                    msg = sprint() do io
-                        print(io, "custom rule lower failure fwd\n")
-                        print(io, "Ty = $Ty\n")
-                        print(io, "llty = $llty\n")
-                        print(io, "arty = $arty\n")
-                        print(io, "al0 = $al0\n")
-                        print(io, "ptr = $ptr\n")
-                        print(io, "val = $val\n")
-                        print(io, "arg = $arg\n")
-                    end
-                    throw(OpaquePointerError(msg))
+                    if !LLVM.is_opaque(value_type(val))
+                        if arty != eltype(value_type(val))
+                            msg = sprint() do io
+                                println(io, "Enzyme: active by ref type $Ty is wrong type in application of custom rule for $mi val=$(string(val)) ptr=$(string(ptr)) arty=$arty")
+                            end
 
-                    if arty == eltype(value_type(val))
-                        val = load!(B, arty, val)
-                    else
-                        bt = GPUCompiler.backtrace(orig)
-                        msg2 = sprint(Base.Fix2(Base.show_backtrace, bt))
-                        val = LLVM.UndefValue(arty)
-                        emit_error(
-                            B,
-                            orig,
-                            "Enzyme: active by ref type $Ty is wrong type in application of custom rule for $mi val=$(string(val)) ptr=$(string(ptr))\n"*msg2,
-                        )
+                            EnzymeInternalError(msg, ir, bt)
+                        end
                     end
+
+                    val = load!(B, arty, val)
                 end
 
-                if eltype(value_type(ptr)) == value_type(val)
+                if arty == value_type(val)
                     store!(B, val, ptr)
                     if any_jltypes(llty)
                         emit_writebarrier!(B, get_julia_inner_types(B, al0, val))