Test MO_MulMayOflo

Here we add a test exercising MO_MulMayOflo, addresssing #2 (closed).

Until now we made no attempt to test this machop since the backend is free to implement it conservatively, precluding treating it as we do most other non-callish machops. I briefly considered generalizing the reference implementation to allow sets of allowed results but this seemed to add more complexity than this case warrants and would make it harder to reason about the correctness of the testsuite.

Here I instead added a separate test which tests MO_MulMayOflo independently.

As GHC currently does not have surface syntax for this primop, this depends upon ghc!8331 (closed).

In service of fixing ghc#21624 (closed).

src/Expr.hs

@@ -231,8 +231,8 @@ genExpr' _width = sized gen
         [ binOp EAdd
         , binOp ESub
         , binOp EMul
-        , divOp EQuot
-        , divOp ERem
+        , quotOp
+        , remOp
         , ENegate <$> arbitrary
         ]
 
@@ -267,7 +267,20 @@ genExpr' _width = sized gen
         ELit <$> chooseNumber (0, fromIntegral $ widthBits (knownWidth @width) - 1)
     subexpr2 = scale (`div` 2) genExpr
     binOp f = f <$> subexpr2 <*> subexpr2
-    divOp f = f <$> arbitrary <*> subexpr2 <*> nonzero subexpr2
+    quotOp = do
+        signedness <- arbitrary
+        num <- subexpr2
+        let num' = interpret num
+            -- divisions where the result overflows the expression width
+            -- (e.g. (-128::W8) / (-1::W8)) are undefined.
+            -- See test-primops#1.
+            okay_denom denom
+              | Signed <- signedness =
+                  num' /= maxBound || interpret denom /= maxBound
+              | otherwise = True
+        denom <- suchThat (nonzero subexpr2) okay_denom
+        return $ EQuot signedness num denom
+    remOp = ERem <$> arbitrary <*> subexpr2 <*> nonzero subexpr2
     nonzero = flip suchThat $ \x -> interpret x /= 0
 
 -- * SomeExpr