Compiler performance is much worse in for loopy givens

This is a deliberate choice, to simplify code, invariants, and I think
performance in typical cases.  The "loopy givens" case is situations like

   [G]  a ~ TF (a, Int)

where TF is a type function with TF (a,b) = (TF a, TF b).

See Note [An alternative story for the inert substitution] in TcFlatten.

testsuite/tests/perf/compiler/T5837.hs

@@ -10,14 +10,44 @@ t = undefined
 
 {-
 
-    [G] a ~ TF (a,Int)
+         [G] a ~ TF (a,Int)               -- a = a_am1
 -->
-    TF (a,Int) ~ fsk
-    fsk ~ a
+         [G] TF (a,Int) ~ fsk             -- fsk = fsk_am8
+inert    [G] fsk ~ a
+
 --->
-    fsk ~ (TF a, TF Int)
-    fsk ~ a
+    [G] fsk ~ (TF a, TF Int)
+inert    [G] fsk ~ a
+
 --->
     a ~ (TF a, TF Int)
+inert    [G] fsk ~ a
+
+---> (attempting to flatten (TF a) so that it does not mention a
+         TF a ~ fsk2
+inert    a ~ (fsk2, TF Int)
+inert    fsk ~ (fsk2, TF Int)
+
+---> (substitute for a)
+         TF (fsk2, TF Int) ~ fsk2
+inert    a ~ (fsk2, TF Int)
+inert    fsk ~ (fsk2, TF Int)
+
+---> (top-level reduction, re-orient)
+         fsk2 ~ (TF fsk2, TF Int)
+inert    a ~ (fsk2, TF Int)
+inert    fsk ~ (fsk2, TF Int)
+
+---> (attempt to flatten (TF fsk2) to get rid of fsk2
+         TF fsk2 ~ fsk3
+         fsk2 ~ (fsk3, TF Int)
+inert    a   ~ (fsk2, TF Int)
+inert    fsk ~ (fsk2, TF Int)
+
+--->
+         TF fsk2 ~ fsk3
+inert    fsk2 ~ (fsk3, TF Int)
+inert    a   ~ ((fsk3, TF Int), TF Int)
+inert    fsk ~ ((fsk3, TF Int), TF Int)
 
 -}
\ No newline at end of file

testsuite/tests/perf/compiler/all.T

@@ -455,7 +455,7 @@ test('T5837',
              # 40000000 (x86/Linux)
              # 2013-11-13: 45520936 (x86/Windows, 64bit machine)
              # 2041-09-03: 37096484 (Windows laptop, w/w for INLINABLE things
-           (wordsize(64), 75765728, 10)])
+           (wordsize(64), 651924880, 10)])
              # sample: 3926235424 (amd64/Linux, 15/2/2012)
              # 2012-10-02 81879216
              # 2012-09-20 87254264 amd64/Linux
@@ -464,6 +464,8 @@ test('T5837',
              #                                  for constraints solving
              # 2014-08-29 73639840 amd64/Linux, w/w for INLINABLE things
              # 2014-10-08 73639840 amd64/Linux, Burning Bridges and other small changes
+             # 2014-11-02 651924880      Linux, Accept big regression;
+             #   See Note [An alternative story for the inert substitution] in TcFlatten
       ],
       compile_fail,['-ftype-function-depth=50'])