- Aug 23, 1990
-
-
ram authored
SC. Fixed SAVE-IF-NECESSARY to change :SPECIFIED-SAVE TNs to :SAVE.
-
- Aug 16, 1990
-
-
ram authored
like any other global TN. Changed register saving to save the registers before the ALLOCATE-FRAME VOP so that the new local call arg passing works. Added local save optimization: only save a value once in a block, and don't restore if the value isn't used before the next save. Also, added some cost determination stuff that tries to tell when it is more trouble than it is worth to allocate a value in a register due to saving costs. Added support for pre-packed save TNs (:SPECIFIED-SAVE). Changed the packing order: now instead of more-or-less random, we pack in the order of the TN's first mention in the emit order.
-
- Jul 13, 1990
-
-
wlott authored
(vop-info (ir2-block-last-vop 2block))).
-
- Jun 15, 1990
-
-
wlott authored
-
- Jun 06, 1990
-
-
ram authored
and added appropriate new error routines.
-
- May 30, 1990
-
-
cvs2git authored
-
- May 11, 1990
- May 09, 1990
-
-
wlott authored
-
- Apr 23, 1990
- Apr 22, 1990
-
-
wlott authored
single element of element size. Why bother wrapping, when it's never going to fit?
-
- Apr 21, 1990
-
-
ram authored
the TN-REF-LOAD-TN. Changed EMIT-SAVES to run before PACK-LOAD-TNs. This should have no effect on semantics, but lets the saved registers be used for load TNs without having to spill them again. Probably not a real efficiency issue, but saving the same register twice looks silly.
-
wlott authored
number of blocks. Even if its not NIL, it will be wrong. Use ir2-block-count instead.
-
- Apr 19, 1990
-
-
ram authored
-
ram authored
allocated (initial size of 0).
-
ram authored
detecting argument/result references while scanning the NEXT-REF thread had become untenable now that times can explicitly be specified for operands. Fortunately solving this problem simplified the code. Also, changed TARGET-IF-DESIRABLE to always target so that load-TN targeting will work better. Changed COMPUTE-LIVE-TNS to kill any result TNs before returning. Because I stopped advancing to the previous VOP before packing arg load TNs, results appeared to conflict with argument load TNs. We don't need this, since all conflicts for TNs used by the VOP are explicitly tested for in the other code.
-
- Apr 17, 1990
-
-
ram authored
-
- Apr 16, 1990
-
-
ram authored
revision 1.2.1.4 date: 90/04/08 21:05:04; author: ram; state: Exp; lines added/del: 12/14 Fixed some compiler errors. ---------------------------- revision 1.2.1.3 date: 90/04/02 15:18:54; author: ram; state: Exp; lines added/del: 42/32 Added support for :COMPONENT TNs. Deleted some old debugging code in SELECT-LOCATION. Changed COMPUTE-LIVE-TNS to use DO-LIVE-TNS instead of rolling our own. Load TN pack will now notice conflicts with :ENVIRONMENT and :COMPONENT TNs. We didn't lose before because all :ENVIRONMENT TNs were on the stack. Added some (INLINE MEMBER) declarations. ---------------------------- revision 1.2.1.2 date: 90/03/27 16:07:59; author: ram; state: Exp; lines added/del: 207/311 Changed stuff to use DO-IR2-ENVIRONMENT-TNS instead of IR2-ENVIRONMENT-BLOCKS. Flushed TN cost hackery, now that this is replaced by representation selection. Changed stuff to iterate over alternate SCs instead of primitive-type-SCs, and changed stuff to use the save-p/alternates instead of the *save-scs* info. Changed saving and spilling to work using load-functions via the LOAD-OPERAND VOP. Changed load-TN packing to just stick the TN in the TN-REF-LOAD-TN instead of inserting move VOPs. This allowed the load-TN conflict stuff to be simplified. Changed load-TN packing to just go for the recorded load-SC, rather than trying to select arbitrary load SCs. Made error messages somewhat snazzier, tho these should be happening less often now that representation selection either detects lossage or allows you to do losing things. Changed restricted packing to work by a flag passed into pack, now that the restriction isn't encoded in the TN-COSTS. ---------------------------- revision 1.2.1.1 date: 90/03/08 14:01:30; author: ram; state: Exp; lines added/del: 2/2 Start work on better non-descriptor support.
-
- Apr 13, 1990
-
-
ram authored
vops correctly: we must insert a block on the target path to restore the spilled value.
-
- Apr 04, 1990
-
-
wlott authored
the right time.
-
- Mar 05, 1990
- Feb 06, 1990
-
-
wlott authored
-