From b0709800418eb296f048274d7492205401b7e98c Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Fran=C3=A7ois=20F=C3=A9votte?= <ff@triscale-innov.com>
Date: Sat, 22 Jun 2019 18:25:45 +0200
Subject: [PATCH] Detect intrinsic FP functions (instead of higher-level ones)

---
 src/overdub.jl | 38 +++++++++++++++++++++++++-------------
 1 file changed, 25 insertions(+), 13 deletions(-)

diff --git a/src/overdub.jl b/src/overdub.jl
index 849b0d9..51635cd 100644
--- a/src/overdub.jl
+++ b/src/overdub.jl
@@ -2,25 +2,37 @@ using Cassette
 
 Cassette.@context CounterCtx;
 
-ops = [
-    (:add32, +, Float32, Float32),
-    (:sub32, -, Float32, Float32),
-    (:mul32, *, Float32, Float32),
-    (:div32, /, Float32, Float32),
-    (:add64, +, Float64, Float64),
-    (:sub64, -, Float64, Float64),
-    (:mul64, *, Float64, Float64),
-    (:div64, /, Float64, Float64),
-]
+const ops = (
+    (:add32, Core.Intrinsics.add_float, Float32),
+    (:sub32, Core.Intrinsics.sub_float, Float32),
+    (:mul32, Core.Intrinsics.mul_float, Float32),
+    (:div32, Core.Intrinsics.div_float, Float32),
+    (:add64, Core.Intrinsics.add_float, Float64),
+    (:sub64, Core.Intrinsics.sub_float, Float64),
+    (:mul64, Core.Intrinsics.mul_float, Float64),
+    (:div64, Core.Intrinsics.div_float, Float64),
+)
 
 @eval mutable struct Counter
     $((:($(op[1]) ::Int) for op in ops)...)
     Counter() = new($((0 for _ in 1:length(ops))...))
 end
 
-for (name, op, typ1, typ2) in ops
-    @eval function Cassette.prehook(ctx::CounterCtx, op::typeof($op), a::$typ1, b::$typ2)
-        ctx.metadata.$name +=1
+for typ1 in (Float32, Float64)
+    @eval function Cassette.prehook(ctx::CounterCtx,
+                                    op::Core.IntrinsicFunction,
+                                    ::$typ1,
+                                    ::$typ1)
+        $(Expr(:block,
+               (map(ops) do (name, op, typ2)
+                  typ1 == typ2 || return :nothing
+                  quote
+                    if op == $op
+                       ctx.metadata.$name += 1
+                       return
+                    end
+                  end
+                end)...))
     end
 end