log

1 files changed, 53 insertions, 0 deletions

diff --git a/log.md b/log.md
index 69106d8..a91fe61 100644
--- a/log.md
+++ b/log.md
@@ -495,3 +495,56 @@ in cg `4290310`).
 Tried compiling the Clean compiler with a hacked clm. Had to `strip cocl`
 manually, then achieved a size of 5.1M (5325412 bytes) compared to 5.3M
 (5475152) for ARM. Segfaults, of course, when tried to use.
+
+### 2016-11-28
+
+Hardcoded registers in `cgthumb2was.c` (not in the `register_name` array):
+
+- `r12`, r. 601, 606, 613, 619 (`w_as_*scratch_register*`) (removed `a24f0e1`)
+- `sp`, r. 1186 (`dyadic_instruction_allowed`, goes through `register_name`),
+  2396 (`rts`), 2412 (`rtsi`)
+- `pc`, r. 1425, 1431 (`jmp`), 1541, 1544, 1555, 1561, 1566 (`jsr`), 2396
+  (`rts`), 2412 (`rtsi`)
+- `lr`, r. 1541, 1544, 1555, 1561, 1566 (`jsr`) (removed `a24f0e1`)
+- `r5`, r. 2750, 2762 (`garbage_collect_test`) (removed `a24f0e1`)
+
+The below table lists some options that can be looked at for the register
+allocation optimisations. The first column, ARM, is the allocation as is used
+in the ARM RTS/CG, the other columns are possible changes.
+
+| Option           | ARM   | +Aux, -B | +Aux, -A | +Aux, +A | Aux/B, -A | Aux/B, -Scratch | Aux/B, -Link | Aux/B, -Heap |
+|------------------|:-----:|:--------:|:--------:|:--------:|:---------:|:---------------:|:------------:|:------------:|
+| **B-stack**      | **4** | **4**    | **4**    |   8      | **4**     | **4**           | **4**        | **4**        |
+|                  | **3** |  14      | **3**    |  14      | **3**     | **3**           | **3**        | **3**        |
+|                  | **2** |  12      | **2**    |  12      |  12       |  14             |  12          |  14          |
+|                  | **1** |  10      |  10      |  10      |  10       |  10             |  10          |  12          |
+|                  | **0** |   9      |   9      |   9      |  14       |   9             |   9          |   9          |
+| **A-stack**      | **6** | **6**    |  14      | **6**    | **6**     | **6**           | **6**        | **6**        |
+|                  | **7** | **7**    |  12      | **7**    | **7**     | **7**           | **7**        | **7**        |
+|                  |   8   |   8      |   8      | **4**    |   8       |   8             |   8          |   8          |
+|                  |  11   |  11      |  11      |  11      |  11       |  11             |  11          |  11          |
+| **Free heap**    | **5** | **5**    | **5**    | **5**    | **5**     | **5**           | **5**        | **5**        |
+| **A ptr**        |   9   | **0**    | **0**    | **0**    |   9       | **0**           | **0**        | **0**        |
+| **Heap ptr**     |  10   | **1**    | **1**    | **1**    | **1**     | **1**           | **1**        |  10          |
+| **Scratch**      |  12   | **2**    | **7**    | **2**    | **2**     |  12             | **2**        | **1**        |
+| **B ptr**        |  13   |  13      |  13      |  13      |  13       |  13             |  13          |  13          |
+| **Link/Scratch** |  14   | **3**    | **6**    | **3**    | **0**     | **2**           |  14          | **2**        |
+
+However, before this can be tried out, there is an **issue** in the code
+generator which makes that `frontend/predef.icl` in the cocl cannot be
+compiled: in the long array updates there are many negative offset addresses,
+and in Thumb these can be at most 255.
+
+Made some changes to the garbage collector:
+
+- In the copying collector, modified `copy_arity_0_node2_` to hopefully store
+  the right LSB (rts `5a329af`)
+- In the compacting/marking collector, there was still an issue with an address
+  being loaded from an address off-by-one (e.g. `0xcc000002f6` instead of
+  `0x2f60e`); fixed this temporarily in rts `b8710c5` but should talk to John
+  about this. Also, the compacting compiler still breaks on the following with
+  `-h 1k`:
+
+  ```clean
+  Start = (length xs, hd xs) where xs = fromto 1 100
+  ```