X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FnativeGen%2FNOTES;h=9068a7fc2caf51f6d7012db40fd8a9ac58284f58;hb=423d477bfecd490de1449c59325c8776f91d7aac;hp=9927b6eeeb643fc848814035feff8055b557685f;hpb=a040ea4a643d8113a023ec7ed880b1e33726b4f4;p=ghc-hetmet.git

diff --git a/ghc/compiler/nativeGen/NOTES b/ghc/compiler/nativeGen/NOTES
index 9927b6e..9068a7f 100644
--- a/ghc/compiler/nativeGen/NOTES
+++ b/ghc/compiler/nativeGen/NOTES
@@ -1,34 +1,41 @@
+TODO in new NCG
+~~~~~~~~~~~~~~~
 
-Known bugs/issues in nativeGen, 000202 (JRS)
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+- Are we being careful enough about narrowing those out-of-range CmmInts?
 
-All these bugs are for x86; I don't know about sparc/alpha.
+- Register allocator:
+    - fixup code
+    - keep track of free stack slots
 
--- absC -> stix translation for GET_TAG and in fact anything
-   to do with the packed-halfword layout info itbl field is
-   pretty dubious.  I think I have it fixed for big and little
-   endian 32-bit, but it won't work at all on a 64 bit platform.
+    Optimisations:
 
--- Most of the x86 insn selector code in MachCode.lhs needs to
-   be checked against the Rules of the Game recorded in that file.
-   I think there are a lot of subtle violations.
+    - picking the assignment on entry to a block: better to defer this
+      until we know all the assignments.  In a loop, we should pick
+      the assignment from the looping jump (fixpointing?), so that any
+      fixup code ends up *outside* the loop.  Otherwise, we should
+      pick the assignment that results in the least fixup code.
 
--- When selecting spill regs, don't use %eax if there is a CALL insn
-   (perhaps excluding calls to newCAF, since it doesn't return a
-    result).
+- splitting?
 
--- Keep track of the stack offset so that correct spill code can
-   be generated even if %esp moves.  At the moment %esp doesn't
-   move, so the problem doesn't exist, but there is a different
-   problem: ccalls put args in memory below %esp and only move
-   %esp immediately prior to the call.  This is dangerous because
-   (1) writing below %esp can cause a segmentation fault (as deemed
-   by the OS), and (2) if a signal should be handled on that stack
-   during argument construction, the args will get silently trashed.
+-- -----------------------------------------------------------------------------
+-- x86 ToDos
 
--- nofib/real/hidden gets slightly different FP answers from the
-   via-C route; possibly due to exp/log not being done in-line.
+- x86 genCCall needs to tack on the @size for stdcalls (might not be in the
+  foreignlabel).
 
--- Possibly implement GLDZ and GLD1 as analogues of FLDZ and FLD1
-   (x86), to reduce number of constants emitted in f-p code.
+- x86: should really clean up that IMUL64 stuff, and tell the code gen about
+  Intel imul instructions.
+
+- x86: we're not careful enough about making sure that we only use
+  byte-addressable registers in byte instructions.  Should we do it this
+  way, or stick to using 32-bit registers everywhere?
+
+- Use SSE for floating point, optionally.
+
+------------------------------------------------------------------------------
+-- Further optimisations:
+
+- We might be able to extend the scope of the inlining phase so it can
+  skip over more statements that don't affect the value of the inlined
+  expr.