Dmitry Stogov
25bd3024da
Codegeneration for VA_ARG nodes (Windows and MacOS are not supported yet)
2023-11-23 19:38:33 +03:00
Dmitry Stogov
68068c4f58
Split ir_ctx->flags into public and private (ir_ctx->flags2)
2023-11-17 14:24:33 +03:00
Dmitry Stogov
8668550427
Initial support for LLVM loader (incomplete) ( #53 )
...
Currently we may just load LLVM module and convert all functions into IR (print it).
For better LLVM support IR framework needs support for modules.
Some LLVM features are not supported yet (see TODOs in ir_load_llvm.c) .
2023-10-11 12:46:31 +03:00
Dmitry Stogov
9ea551a34f
LLVM support for fastcall and vararg
2023-09-29 11:30:53 +03:00
Dmitry Stogov
dd2ecad299
Allow reuse of spill slots for objecs of smaller size
2023-08-02 13:20:13 +03:00
Dmitry Stogov
1d49fe6cc4
Reduce cost of disabled IR_ASSERT()
2023-07-27 11:16:00 +03:00
Dmitry Stogov
f6cf9140da
Update ir_insn.inputs_count and use it after ir_build_def_use_lists()
2023-07-06 01:15:08 +03:00
Dmitry Stogov
ce2d6ceba6
Fixed non-boolean constant GUARD condition checks
2023-06-29 23:49:20 +03:00
Dmitry Stogov
d67c212916
Separate codegen info output into ir_dump_codegen()
2023-06-21 22:36:36 +03:00
Dmitry Stogov
ebaefd376a
Fix stack frame and assign all spill slots before code genearatin
2023-06-21 19:04:22 +03:00
Dmitry Stogov
4124ef5150
Allow printing IR annotated with register-allocation, spill-code-placement, de-SSA and code-generation information
2023-06-21 13:28:15 +03:00
Dmitry Stogov
6a98514bdc
Move stack size related metricks to ir_ctx
2023-06-15 19:28:54 +03:00
Dmitry Stogov
311267714e
Use macros insted of bit ops
2023-06-14 20:23:32 +03:00
Dmitry Stogov
c93abd79b2
Remove IR_OPND_VAR
2023-05-19 13:00:55 +03:00
Dmitry Stogov
5c2023fd7f
Avoid live range constrction for VARs
2023-05-18 21:00:57 +03:00
Dmitry Stogov
ef31712f26
Remove usused IR_LIVE_INTERVAL_REG_LOAD flag
2023-05-18 14:17:54 +03:00
Dmitry Stogov
b6517f2028
Split IR_LIVE_INTERVAL_HAS_HINTS into IR_LIVE_INTERVAL_HAS_HINT_REGS and
...
IR_LIVE_INTERVAL_HAS_HINT_REFS
2023-05-18 10:48:31 +03:00
Dmitry Stogov
c3fca224ff
Introduce IR_OP_FLAG_PINNED and IR_BB_HAS_PHI/PI/PARAM/VAR flags
2023-05-05 16:59:22 +03:00
Dmitry Stogov
76f44fa126
Implemented path-exploration based live ranges construction
...
The algorithm is based on "Computing Liveness Sets for SSA-Form Programs",
Florian Brandner, Benoit Boissinot, Alain Darte, Benoit Dupont de Dinechin,
Fabrice Rastello. TR Inria RR-7503, 2011
In comparison to the existing bitset based approach, the new algorithm
is more memory-space efficient and faster on big functions.
2023-05-05 02:01:53 +03:00
Dmitry Stogov
1bbee7b9da
Get rid of ir_live_interval.top
2023-04-28 09:49:12 +03:00
Dmitry Stogov
1749168078
Add ir_insn_len() and ir_insn_inputs_to_len() private helpers
2023-04-21 13:40:55 +03:00
Dmitry Stogov
e01c43a967
Simplify access to nodes with variable inputs count
2023-04-21 12:40:17 +03:00
Dmitry Stogov
44f048ecd9
Reorder fields to reduce the structure size
2023-04-13 21:37:39 +03:00
Dmitry Stogov
2a17b87219
LSRA optimization
2023-04-13 20:56:49 +03:00
Dmitry Stogov
e5c01495da
Use arena to allocate live_intervals and nested data structures
2023-04-13 13:47:16 +03:00
Dmitry Stogov
0b78a322f8
Cache deleted live ranges in ir_ctx.unused_ranges
2023-04-13 11:42:47 +03:00
Dmitry Stogov
d71cbd47d5
Disable LICM across an OSR ENTRY if the value can't be restored at OSR ENTRY point
2023-04-07 16:36:27 +03:00
Dmitry Stogov
efa8a83153
Fix spilling code for arguments passed theought stack and change RA to
...
prefer reusing the same register for splitted intervals
i# utils/
2023-04-06 00:16:49 +03:00
Dmitry Stogov
1e5e9e08ce
Re-implement instruction fusion and live-range construction
2023-04-05 19:20:43 +03:00
Dmitry Stogov
ba0fa44447
Add "const" modifiers
2023-03-28 13:18:12 +03:00
Dmitry Stogov
87dbdcea0d
Add necessary compensation loads for bounded nodes when enter into function through OSR entry-point
2023-03-21 13:45:37 +03:00
Dmitry Stogov
f5b7065b10
Refactor the ENTRY nodes
...
Now all ENTRY nodes have a "fake" input control edge.
Through this edge all of them are dominated by START node.
2023-03-17 09:02:37 +03:00
Dmitry Stogov
9b34731d16
Fix most MSVC compilation warnings
2023-02-28 02:11:09 +03:00
Anatol Belski
964f5a0191
build: MSVC compatibility
...
Signed-off-by: Anatol Belski <ab@php.net>
2023-02-25 00:04:57 +01:00
Dmitry Stogov
fd653528e9
JMP optimization. Lift constant IJMP targets into jmp_table(s).
2023-02-16 22:41:55 +03:00
Dmitry Stogov
9b6b6996c4
Move IR_ALWAYS_INLINE and IR_NEVER_INLINE definition into public ir.h
2023-02-07 23:14:10 +03:00
Dmitry Stogov
677c6cb2cb
Move declaration of some register alloation related macros to public API
...
Use RLOAD.op3 as a flag to avoid spill store
2023-01-30 16:33:57 +03:00
Dmitry Stogov
32ad3d1052
Use inline functions to avoid false positive address sanitaizer warnings
2023-01-20 15:35:02 +03:00
Dmitry Stogov
6a4e239773
Create a sparate pass to remove unreachableble CFG blocks.
...
SCCP pass removes unreachable blocks before CFG construction.
In case of -O0 or -O1 pipeline (without SCCP) it's simpler and faster
to unlink unreachable CFG blocks once, then check for reachability
in almost any compilation pass.
-O2 pipeline (with SCCP) don't need this pass.
2022-11-29 20:02:07 +03:00
Dmitry Stogov
7fd1ccf48b
Eliminte useless checks
2022-11-24 12:23:05 +03:00
Dmitry Stogov
7d07a4ac89
Strength reduction
2022-11-18 14:38:19 +03:00
Dmitry Stogov
00395f0a23
Cleanup: separate ir_phi_input_number() function
2022-11-18 10:11:16 +03:00
Dmitry Stogov
3e3746d5cb
Refactor API that expose target CPU register constraints for register allocator
2022-11-17 23:30:35 +03:00
Jim Huang
af75775bac
Fix build on macOS
...
Corresponding compilation errors:
ir_private.h:11:10: fatal error: 'malloc.h' file not found
ir_gdb.c:162:17: error: implicit declaration of function 'offsetof' is invalid in C99 [-Werror,-Wimplicit-function-declaration]
.shofs = offsetof(ir_gdbjit_obj, sect),
2022-11-12 01:24:33 +08:00
Jim Huang
c4017eab4b
Fix build with GCC prior to 10
2022-11-11 18:05:05 +08:00
Dmitry Stogov
b56f1f5298
Refactor CFG construction algorithm to mark all BB starts in two passes.
...
First we perform a backwad DFS search from "stop" nodes. This pass may
leak some CFG paths ended by infinite loops. To support these paths we
record their possible roots and perform forward DFS search starting from
them.
2022-11-11 10:25:59 +03:00
Dmitry Stogov
47ddea00ac
Use better conditions
2022-11-11 02:56:16 +03:00
Dmitry Stogov
a4b09dd3ce
Improve CFG builder
2022-11-11 00:19:17 +03:00
Dmitry Stogov
78cc14aca9
Fix ir_worklist_clear()
2022-11-10 22:44:29 +03:00
Dmitry Stogov
f484b5d16b
micro-optimization
2022-11-09 13:23:30 +03:00