annotate third_party/luajit/src/lj_opt_narrow.c @ 279:b3b547563ec7

Add Google connector service and agent wiki Implement the C/Seobeo Google Drive and Gmail connector with encrypted OAuth storage, Zenbu authentication, browser testing, AI tool discovery, chunked HTTP decoding, and Bazel coverage. Consolidate repository guidance into progressive wiki documentation and enforce arena-first allocation for new first-party C code. Co-authored-by: Copilot <[email protected]> Copilot-Session: 84c338fd-0939-4bb3-b7f3-1062eb213e5d
author MrJuneJune <me@mrjunejune.com>
date Mon, 17 Aug 2026 22:22:36 -0700
parents 94705b5986b3
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
178
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
1 /*
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
2 ** NARROW: Narrowing of numbers to integers (double to int32_t).
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
3 ** STRIPOV: Stripping of overflow checks.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
4 ** Copyright (C) 2005-2023 Mike Pall. See Copyright Notice in luajit.h
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
5 */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
6
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
7 #define lj_opt_narrow_c
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
8 #define LUA_CORE
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
9
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
10 #include "lj_obj.h"
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
11
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
12 #if LJ_HASJIT
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
13
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
14 #include "lj_bc.h"
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
15 #include "lj_ir.h"
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
16 #include "lj_jit.h"
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
17 #include "lj_iropt.h"
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
18 #include "lj_trace.h"
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
19 #include "lj_vm.h"
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
20 #include "lj_strscan.h"
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
21
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
22 /* Rationale for narrowing optimizations:
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
23 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
24 ** Lua has only a single number type and this is a FP double by default.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
25 ** Narrowing doubles to integers does not pay off for the interpreter on a
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
26 ** current-generation x86/x64 machine. Most FP operations need the same
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
27 ** amount of execution resources as their integer counterparts, except
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
28 ** with slightly longer latencies. Longer latencies are a non-issue for
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
29 ** the interpreter, since they are usually hidden by other overhead.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
30 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
31 ** The total CPU execution bandwidth is the sum of the bandwidth of the FP
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
32 ** and the integer units, because they execute in parallel. The FP units
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
33 ** have an equal or higher bandwidth than the integer units. Not using
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
34 ** them means losing execution bandwidth. Moving work away from them to
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
35 ** the already quite busy integer units is a losing proposition.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
36 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
37 ** The situation for JIT-compiled code is a bit different: the higher code
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
38 ** density makes the extra latencies much more visible. Tight loops expose
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
39 ** the latencies for updating the induction variables. Array indexing
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
40 ** requires narrowing conversions with high latencies and additional
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
41 ** guards (to check that the index is really an integer). And many common
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
42 ** optimizations only work on integers.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
43 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
44 ** One solution would be speculative, eager narrowing of all number loads.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
45 ** This causes many problems, like losing -0 or the need to resolve type
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
46 ** mismatches between traces. It also effectively forces the integer type
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
47 ** to have overflow-checking semantics. This impedes many basic
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
48 ** optimizations and requires adding overflow checks to all integer
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
49 ** arithmetic operations (whereas FP arithmetics can do without).
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
50 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
51 ** Always replacing an FP op with an integer op plus an overflow check is
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
52 ** counter-productive on a current-generation super-scalar CPU. Although
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
53 ** the overflow check branches are highly predictable, they will clog the
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
54 ** execution port for the branch unit and tie up reorder buffers. This is
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
55 ** turning a pure data-flow dependency into a different data-flow
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
56 ** dependency (with slightly lower latency) *plus* a control dependency.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
57 ** In general, you don't want to do this since latencies due to data-flow
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
58 ** dependencies can be well hidden by out-of-order execution.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
59 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
60 ** A better solution is to keep all numbers as FP values and only narrow
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
61 ** when it's beneficial to do so. LuaJIT uses predictive narrowing for
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
62 ** induction variables and demand-driven narrowing for index expressions,
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
63 ** integer arguments and bit operations. Additionally it can eliminate or
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
64 ** hoist most of the resulting overflow checks. Regular arithmetic
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
65 ** computations are never narrowed to integers.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
66 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
67 ** The integer type in the IR has convenient wrap-around semantics and
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
68 ** ignores overflow. Extra operations have been added for
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
69 ** overflow-checking arithmetic (ADDOV/SUBOV) instead of an extra type.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
70 ** Apart from reducing overall complexity of the compiler, this also
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
71 ** nicely solves the problem where you want to apply algebraic
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
72 ** simplifications to ADD, but not to ADDOV. And the x86/x64 assembler can
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
73 ** use lea instead of an add for integer ADD, but not for ADDOV (lea does
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
74 ** not affect the flags, but it helps to avoid register moves).
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
75 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
76 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
77 ** All of the above has to be reconsidered for architectures with slow FP
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
78 ** operations or without a hardware FPU. The dual-number mode of LuaJIT
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
79 ** addresses this issue. Arithmetic operations are performed on integers
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
80 ** as far as possible and overflow checks are added as needed.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
81 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
82 ** This implies that narrowing for integer arguments and bit operations
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
83 ** should also strip overflow checks, e.g. replace ADDOV with ADD. The
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
84 ** original overflow guards are weak and can be eliminated by DCE, if
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
85 ** there's no other use.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
86 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
87 ** A slight twist is that it's usually beneficial to use overflow-checked
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
88 ** integer arithmetics if all inputs are already integers. This is the only
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
89 ** change that affects the single-number mode, too.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
90 */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
91
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
92 /* Some local macros to save typing. Undef'd at the end. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
93 #define IR(ref) (&J->cur.ir[(ref)])
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
94 #define fins (&J->fold.ins)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
95
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
96 /* Pass IR on to next optimization in chain (FOLD). */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
97 #define emitir(ot, a, b) (lj_ir_set(J, (ot), (a), (b)), lj_opt_fold(J))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
98
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
99 #define emitir_raw(ot, a, b) (lj_ir_set(J, (ot), (a), (b)), lj_ir_emit(J))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
100
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
101 /* -- Elimination of narrowing type conversions --------------------------- */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
102
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
103 /* Narrowing of index expressions and bit operations is demand-driven. The
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
104 ** trace recorder emits a narrowing type conversion (CONV.int.num or TOBIT)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
105 ** in all of these cases (e.g. array indexing or string indexing). FOLD
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
106 ** already takes care of eliminating simple redundant conversions like
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
107 ** CONV.int.num(CONV.num.int(x)) ==> x.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
108 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
109 ** But the surrounding code is FP-heavy and arithmetic operations are
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
110 ** performed on FP numbers (for the single-number mode). Consider a common
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
111 ** example such as 'x=t[i+1]', with 'i' already an integer (due to induction
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
112 ** variable narrowing). The index expression would be recorded as
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
113 ** CONV.int.num(ADD(CONV.num.int(i), 1))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
114 ** which is clearly suboptimal.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
115 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
116 ** One can do better by recursively backpropagating the narrowing type
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
117 ** conversion across FP arithmetic operations. This turns FP ops into
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
118 ** their corresponding integer counterparts. Depending on the semantics of
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
119 ** the conversion they also need to check for overflow. Currently only ADD
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
120 ** and SUB are supported.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
121 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
122 ** The above example can be rewritten as
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
123 ** ADDOV(CONV.int.num(CONV.num.int(i)), 1)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
124 ** and then into ADDOV(i, 1) after folding of the conversions. The original
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
125 ** FP ops remain in the IR and are eliminated by DCE since all references to
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
126 ** them are gone.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
127 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
128 ** [In dual-number mode the trace recorder already emits ADDOV etc., but
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
129 ** this can be further reduced. See below.]
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
130 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
131 ** Special care has to be taken to avoid narrowing across an operation
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
132 ** which is potentially operating on non-integral operands. One obvious
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
133 ** case is when an expression contains a non-integral constant, but ends
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
134 ** up as an integer index at runtime (like t[x+1.5] with x=0.5).
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
135 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
136 ** Operations with two non-constant operands illustrate a similar problem
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
137 ** (like t[a+b] with a=1.5 and b=2.5). Backpropagation has to stop there,
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
138 ** unless it can be proven that either operand is integral (e.g. by CSEing
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
139 ** a previous conversion). As a not-so-obvious corollary this logic also
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
140 ** applies for a whole expression tree (e.g. t[(a+1)+(b+1)]).
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
141 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
142 ** Correctness of the transformation is guaranteed by avoiding to expand
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
143 ** the tree by adding more conversions than the one we would need to emit
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
144 ** if not backpropagating. TOBIT employs a more optimistic rule, because
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
145 ** the conversion has special semantics, designed to make the life of the
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
146 ** compiler writer easier. ;-)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
147 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
148 ** Using on-the-fly backpropagation of an expression tree doesn't work
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
149 ** because it's unknown whether the transform is correct until the end.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
150 ** This either requires IR rollback and cache invalidation for every
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
151 ** subtree or a two-pass algorithm. The former didn't work out too well,
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
152 ** so the code now combines a recursive collector with a stack-based
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
153 ** emitter.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
154 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
155 ** [A recursive backpropagation algorithm with backtracking, employing
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
156 ** skip-list lookup and round-robin caching, emitting stack operations
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
157 ** on-the-fly for a stack-based interpreter -- and all of that in a meager
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
158 ** kilobyte? Yep, compilers are a great treasure chest. Throw away your
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
159 ** textbooks and read the codebase of a compiler today!]
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
160 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
161 ** There's another optimization opportunity for array indexing: it's
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
162 ** always accompanied by an array bounds-check. The outermost overflow
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
163 ** check may be delegated to the ABC operation. This works because ABC is
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
164 ** an unsigned comparison and wrap-around due to overflow creates negative
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
165 ** numbers.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
166 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
167 ** But this optimization is only valid for constants that cannot overflow
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
168 ** an int32_t into the range of valid array indexes [0..2^27+1). A check
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
169 ** for +-2^30 is safe since -2^31 - 2^30 wraps to 2^30 and 2^31-1 + 2^30
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
170 ** wraps to -2^30-1.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
171 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
172 ** It's also good enough in practice, since e.g. t[i+1] or t[i-10] are
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
173 ** quite common. So the above example finally ends up as ADD(i, 1)!
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
174 **
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
175 ** Later on, the assembler is able to fuse the whole array reference and
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
176 ** the ADD into the memory operands of loads and other instructions. This
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
177 ** is why LuaJIT is able to generate very pretty (and fast) machine code
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
178 ** for array indexing. And that, my dear, concludes another story about
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
179 ** one of the hidden secrets of LuaJIT ...
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
180 */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
181
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
182 /* Maximum backpropagation depth and maximum stack size. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
183 #define NARROW_MAX_BACKPROP 100
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
184 #define NARROW_MAX_STACK 256
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
185
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
186 /* The stack machine has a 32 bit instruction format: [IROpT | IRRef1]
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
187 ** The lower 16 bits hold a reference (or 0). The upper 16 bits hold
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
188 ** the IR opcode + type or one of the following special opcodes:
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
189 */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
190 enum {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
191 NARROW_REF, /* Push ref. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
192 NARROW_CONV, /* Push conversion of ref. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
193 NARROW_SEXT, /* Push sign-extension of ref. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
194 NARROW_INT /* Push KINT ref. The next code holds an int32_t. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
195 };
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
196
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
197 typedef uint32_t NarrowIns;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
198
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
199 #define NARROWINS(op, ref) (((op) << 16) + (ref))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
200 #define narrow_op(ins) ((IROpT)((ins) >> 16))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
201 #define narrow_ref(ins) ((IRRef1)(ins))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
202
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
203 /* Context used for narrowing of type conversions. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
204 typedef struct NarrowConv {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
205 jit_State *J; /* JIT compiler state. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
206 NarrowIns *sp; /* Current stack pointer. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
207 NarrowIns *maxsp; /* Maximum stack pointer minus redzone. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
208 IRRef mode; /* Conversion mode (IRCONV_*). */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
209 IRType t; /* Destination type: IRT_INT or IRT_I64. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
210 NarrowIns stack[NARROW_MAX_STACK]; /* Stack holding stack-machine code. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
211 } NarrowConv;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
212
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
213 /* Lookup a reference in the backpropagation cache. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
214 static BPropEntry *narrow_bpc_get(jit_State *J, IRRef1 key, IRRef mode)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
215 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
216 ptrdiff_t i;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
217 for (i = 0; i < BPROP_SLOTS; i++) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
218 BPropEntry *bp = &J->bpropcache[i];
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
219 /* Stronger checks are ok, too. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
220 if (bp->key == key && bp->mode >= mode &&
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
221 ((bp->mode ^ mode) & IRCONV_MODEMASK) == 0)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
222 return bp;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
223 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
224 return NULL;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
225 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
226
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
227 /* Add an entry to the backpropagation cache. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
228 static void narrow_bpc_set(jit_State *J, IRRef1 key, IRRef1 val, IRRef mode)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
229 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
230 uint32_t slot = J->bpropslot;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
231 BPropEntry *bp = &J->bpropcache[slot];
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
232 J->bpropslot = (slot + 1) & (BPROP_SLOTS-1);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
233 bp->key = key;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
234 bp->val = val;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
235 bp->mode = mode;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
236 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
237
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
238 /* Backpropagate overflow stripping. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
239 static void narrow_stripov_backprop(NarrowConv *nc, IRRef ref, int depth)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
240 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
241 jit_State *J = nc->J;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
242 IRIns *ir = IR(ref);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
243 if (ir->o == IR_ADDOV || ir->o == IR_SUBOV ||
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
244 (ir->o == IR_MULOV && (nc->mode & IRCONV_CONVMASK) == IRCONV_ANY)) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
245 BPropEntry *bp = narrow_bpc_get(nc->J, ref, IRCONV_TOBIT);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
246 if (bp) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
247 ref = bp->val;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
248 } else if (++depth < NARROW_MAX_BACKPROP && nc->sp < nc->maxsp) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
249 NarrowIns *savesp = nc->sp;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
250 narrow_stripov_backprop(nc, ir->op1, depth);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
251 if (nc->sp < nc->maxsp) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
252 narrow_stripov_backprop(nc, ir->op2, depth);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
253 if (nc->sp < nc->maxsp) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
254 *nc->sp++ = NARROWINS(IRT(ir->o - IR_ADDOV + IR_ADD, IRT_INT), ref);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
255 return;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
256 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
257 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
258 nc->sp = savesp; /* Path too deep, need to backtrack. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
259 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
260 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
261 *nc->sp++ = NARROWINS(NARROW_REF, ref);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
262 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
263
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
264 /* Backpropagate narrowing conversion. Return number of needed conversions. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
265 static int narrow_conv_backprop(NarrowConv *nc, IRRef ref, int depth)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
266 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
267 jit_State *J = nc->J;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
268 IRIns *ir = IR(ref);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
269 IRRef cref;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
270
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
271 if (nc->sp >= nc->maxsp) return 10; /* Path too deep. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
272
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
273 /* Check the easy cases first. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
274 if (ir->o == IR_CONV && (ir->op2 & IRCONV_SRCMASK) == IRT_INT) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
275 if ((nc->mode & IRCONV_CONVMASK) <= IRCONV_ANY)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
276 narrow_stripov_backprop(nc, ir->op1, depth+1);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
277 else
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
278 *nc->sp++ = NARROWINS(NARROW_REF, ir->op1); /* Undo conversion. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
279 if (nc->t == IRT_I64)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
280 *nc->sp++ = NARROWINS(NARROW_SEXT, 0); /* Sign-extend integer. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
281 return 0;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
282 } else if (ir->o == IR_KNUM) { /* Narrow FP constant. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
283 lua_Number n = ir_knum(ir)->n;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
284 if ((nc->mode & IRCONV_CONVMASK) == IRCONV_TOBIT) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
285 /* Allows a wider range of constants. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
286 int64_t k64 = (int64_t)n;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
287 if (n == (lua_Number)k64) { /* Only if const doesn't lose precision. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
288 *nc->sp++ = NARROWINS(NARROW_INT, 0);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
289 *nc->sp++ = (NarrowIns)k64; /* But always truncate to 32 bits. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
290 return 0;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
291 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
292 } else {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
293 int32_t k = lj_num2int(n);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
294 /* Only if constant is a small integer. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
295 if (checki16(k) && n == (lua_Number)k) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
296 *nc->sp++ = NARROWINS(NARROW_INT, 0);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
297 *nc->sp++ = (NarrowIns)k;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
298 return 0;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
299 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
300 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
301 return 10; /* Never narrow other FP constants (this is rare). */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
302 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
303
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
304 /* Try to CSE the conversion. Stronger checks are ok, too. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
305 cref = J->chain[fins->o];
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
306 while (cref > ref) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
307 IRIns *cr = IR(cref);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
308 if (cr->op1 == ref &&
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
309 (fins->o == IR_TOBIT ||
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
310 ((cr->op2 & IRCONV_MODEMASK) == (nc->mode & IRCONV_MODEMASK) &&
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
311 irt_isguard(cr->t) >= irt_isguard(fins->t)))) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
312 *nc->sp++ = NARROWINS(NARROW_REF, cref);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
313 return 0; /* Already there, no additional conversion needed. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
314 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
315 cref = cr->prev;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
316 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
317
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
318 /* Backpropagate across ADD/SUB. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
319 if (ir->o == IR_ADD || ir->o == IR_SUB) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
320 /* Try cache lookup first. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
321 IRRef mode = nc->mode;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
322 BPropEntry *bp;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
323 /* Inner conversions need a stronger check. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
324 if ((mode & IRCONV_CONVMASK) == IRCONV_INDEX && depth > 0)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
325 mode += IRCONV_CHECK-IRCONV_INDEX;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
326 bp = narrow_bpc_get(nc->J, (IRRef1)ref, mode);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
327 if (bp) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
328 *nc->sp++ = NARROWINS(NARROW_REF, bp->val);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
329 return 0;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
330 } else if (nc->t == IRT_I64) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
331 /* Try sign-extending from an existing (checked) conversion to int. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
332 mode = (IRT_INT<<5)|IRT_NUM|IRCONV_INDEX;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
333 bp = narrow_bpc_get(nc->J, (IRRef1)ref, mode);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
334 if (bp) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
335 *nc->sp++ = NARROWINS(NARROW_REF, bp->val);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
336 *nc->sp++ = NARROWINS(NARROW_SEXT, 0);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
337 return 0;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
338 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
339 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
340 if (++depth < NARROW_MAX_BACKPROP && nc->sp < nc->maxsp) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
341 NarrowIns *savesp = nc->sp;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
342 int count = narrow_conv_backprop(nc, ir->op1, depth);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
343 count += narrow_conv_backprop(nc, ir->op2, depth);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
344 if (count <= 1) { /* Limit total number of conversions. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
345 *nc->sp++ = NARROWINS(IRT(ir->o, nc->t), ref);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
346 return count;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
347 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
348 nc->sp = savesp; /* Too many conversions, need to backtrack. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
349 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
350 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
351
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
352 /* Otherwise add a conversion. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
353 *nc->sp++ = NARROWINS(NARROW_CONV, ref);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
354 return 1;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
355 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
356
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
357 /* Emit the conversions collected during backpropagation. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
358 static IRRef narrow_conv_emit(jit_State *J, NarrowConv *nc)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
359 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
360 /* The fins fields must be saved now -- emitir() overwrites them. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
361 IROpT guardot = irt_isguard(fins->t) ? IRTG(IR_ADDOV-IR_ADD, 0) : 0;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
362 IROpT convot = fins->ot;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
363 IRRef1 convop2 = fins->op2;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
364 NarrowIns *next = nc->stack; /* List of instructions from backpropagation. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
365 NarrowIns *last = nc->sp;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
366 NarrowIns *sp = nc->stack; /* Recycle the stack to store operands. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
367 while (next < last) { /* Simple stack machine to process the ins. list. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
368 NarrowIns ref = *next++;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
369 IROpT op = narrow_op(ref);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
370 if (op == NARROW_REF) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
371 *sp++ = ref;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
372 } else if (op == NARROW_CONV) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
373 *sp++ = emitir_raw(convot, ref, convop2); /* Raw emit avoids a loop. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
374 } else if (op == NARROW_SEXT) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
375 lj_assertJ(sp >= nc->stack+1, "stack underflow");
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
376 sp[-1] = emitir(IRT(IR_CONV, IRT_I64), sp[-1],
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
377 (IRT_I64<<5)|IRT_INT|IRCONV_SEXT);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
378 } else if (op == NARROW_INT) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
379 lj_assertJ(next < last, "missing arg to NARROW_INT");
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
380 *sp++ = nc->t == IRT_I64 ?
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
381 lj_ir_kint64(J, (int64_t)(int32_t)*next++) :
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
382 lj_ir_kint(J, *next++);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
383 } else { /* Regular IROpT. Pops two operands and pushes one result. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
384 IRRef mode = nc->mode;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
385 lj_assertJ(sp >= nc->stack+2, "stack underflow");
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
386 sp--;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
387 /* Omit some overflow checks for array indexing. See comments above. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
388 if ((mode & IRCONV_CONVMASK) == IRCONV_INDEX) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
389 if (next == last && irref_isk(narrow_ref(sp[0])) &&
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
390 (uint32_t)IR(narrow_ref(sp[0]))->i + 0x40000000u < 0x80000000u)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
391 guardot = 0;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
392 else /* Otherwise cache a stronger check. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
393 mode += IRCONV_CHECK-IRCONV_INDEX;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
394 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
395 sp[-1] = emitir(op+guardot, sp[-1], sp[0]);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
396 /* Add to cache. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
397 if (narrow_ref(ref))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
398 narrow_bpc_set(J, narrow_ref(ref), narrow_ref(sp[-1]), mode);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
399 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
400 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
401 lj_assertJ(sp == nc->stack+1, "stack misalignment");
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
402 return nc->stack[0];
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
403 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
404
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
405 /* Narrow a type conversion of an arithmetic operation. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
406 TRef LJ_FASTCALL lj_opt_narrow_convert(jit_State *J)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
407 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
408 if ((J->flags & JIT_F_OPT_NARROW)) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
409 NarrowConv nc;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
410 nc.J = J;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
411 nc.sp = nc.stack;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
412 nc.maxsp = &nc.stack[NARROW_MAX_STACK-4];
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
413 nc.t = irt_type(fins->t);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
414 if (fins->o == IR_TOBIT) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
415 nc.mode = IRCONV_TOBIT; /* Used only in the backpropagation cache. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
416 } else {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
417 nc.mode = fins->op2;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
418 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
419 if (narrow_conv_backprop(&nc, fins->op1, 0) <= 1)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
420 return narrow_conv_emit(J, &nc);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
421 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
422 return NEXTFOLD;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
423 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
424
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
425 /* -- Narrowing of implicit conversions ----------------------------------- */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
426
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
427 /* Recursively strip overflow checks. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
428 static TRef narrow_stripov(jit_State *J, TRef tr, int lastop, IRRef mode)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
429 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
430 IRRef ref = tref_ref(tr);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
431 IRIns *ir = IR(ref);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
432 int op = ir->o;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
433 if (op >= IR_ADDOV && op <= lastop) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
434 BPropEntry *bp = narrow_bpc_get(J, ref, mode);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
435 if (bp) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
436 return TREF(bp->val, irt_t(IR(bp->val)->t));
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
437 } else {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
438 IRRef op1 = ir->op1, op2 = ir->op2; /* The IR may be reallocated. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
439 op1 = narrow_stripov(J, op1, lastop, mode);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
440 op2 = narrow_stripov(J, op2, lastop, mode);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
441 tr = emitir(IRT(op - IR_ADDOV + IR_ADD,
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
442 ((mode & IRCONV_DSTMASK) >> IRCONV_DSH)), op1, op2);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
443 narrow_bpc_set(J, ref, tref_ref(tr), mode);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
444 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
445 } else if (LJ_64 && (mode & IRCONV_SEXT) && !irt_is64(ir->t)) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
446 tr = emitir(IRT(IR_CONV, IRT_INTP), tr, mode);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
447 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
448 return tr;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
449 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
450
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
451 /* Narrow array index. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
452 TRef LJ_FASTCALL lj_opt_narrow_index(jit_State *J, TRef tr)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
453 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
454 IRIns *ir;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
455 lj_assertJ(tref_isnumber(tr), "expected number type");
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
456 if (tref_isnum(tr)) /* Conversion may be narrowed, too. See above. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
457 return emitir(IRTGI(IR_CONV), tr, IRCONV_INT_NUM|IRCONV_INDEX);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
458 /* Omit some overflow checks for array indexing. See comments above. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
459 ir = IR(tref_ref(tr));
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
460 if ((ir->o == IR_ADDOV || ir->o == IR_SUBOV) && irref_isk(ir->op2) &&
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
461 (uint32_t)IR(ir->op2)->i + 0x40000000u < 0x80000000u)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
462 return emitir(IRTI(ir->o - IR_ADDOV + IR_ADD), ir->op1, ir->op2);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
463 return tr;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
464 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
465
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
466 /* Narrow conversion to integer operand (overflow undefined). */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
467 TRef LJ_FASTCALL lj_opt_narrow_toint(jit_State *J, TRef tr)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
468 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
469 if (tref_isstr(tr))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
470 tr = emitir(IRTG(IR_STRTO, IRT_NUM), tr, 0);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
471 if (tref_isnum(tr)) /* Conversion may be narrowed, too. See above. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
472 return emitir(IRTI(IR_CONV), tr, IRCONV_INT_NUM|IRCONV_ANY);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
473 if (!tref_isinteger(tr))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
474 lj_trace_err(J, LJ_TRERR_BADTYPE);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
475 /*
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
476 ** Undefined overflow semantics allow stripping of ADDOV, SUBOV and MULOV.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
477 ** Use IRCONV_TOBIT for the cache entries, since the semantics are the same.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
478 */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
479 return narrow_stripov(J, tr, IR_MULOV, (IRT_INT<<5)|IRT_INT|IRCONV_TOBIT);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
480 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
481
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
482 /* Narrow conversion to bitop operand (overflow wrapped). */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
483 TRef LJ_FASTCALL lj_opt_narrow_tobit(jit_State *J, TRef tr)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
484 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
485 if (tref_isstr(tr))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
486 tr = emitir(IRTG(IR_STRTO, IRT_NUM), tr, 0);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
487 if (tref_isnum(tr)) /* Conversion may be narrowed, too. See above. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
488 return emitir(IRTI(IR_TOBIT), tr, lj_ir_knum_tobit(J));
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
489 if (!tref_isinteger(tr))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
490 lj_trace_err(J, LJ_TRERR_BADTYPE);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
491 /*
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
492 ** Wrapped overflow semantics allow stripping of ADDOV and SUBOV.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
493 ** MULOV cannot be stripped due to precision widening.
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
494 */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
495 return narrow_stripov(J, tr, IR_SUBOV, (IRT_INT<<5)|IRT_INT|IRCONV_TOBIT);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
496 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
497
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
498 #if LJ_HASFFI
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
499 /* Narrow C array index (overflow undefined). */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
500 TRef LJ_FASTCALL lj_opt_narrow_cindex(jit_State *J, TRef tr)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
501 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
502 lj_assertJ(tref_isnumber(tr), "expected number type");
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
503 if (tref_isnum(tr))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
504 return emitir(IRT(IR_CONV, IRT_INTP), tr, (IRT_INTP<<5)|IRT_NUM|IRCONV_ANY);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
505 /* Undefined overflow semantics allow stripping of ADDOV, SUBOV and MULOV. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
506 return narrow_stripov(J, tr, IR_MULOV,
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
507 LJ_64 ? ((IRT_INTP<<5)|IRT_INT|IRCONV_SEXT) :
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
508 ((IRT_INTP<<5)|IRT_INT|IRCONV_TOBIT));
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
509 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
510 #endif
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
511
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
512 /* -- Narrowing of arithmetic operators ----------------------------------- */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
513
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
514 /* Check whether a number fits into an int32_t (-0 is ok, too). */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
515 static int numisint(lua_Number n)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
516 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
517 return (n == (lua_Number)lj_num2int(n));
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
518 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
519
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
520 /* Convert string to number. Error out for non-numeric string values. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
521 static TRef conv_str_tonum(jit_State *J, TRef tr, TValue *o)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
522 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
523 if (tref_isstr(tr)) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
524 tr = emitir(IRTG(IR_STRTO, IRT_NUM), tr, 0);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
525 /* Would need an inverted STRTO for this rare and useless case. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
526 if (!lj_strscan_num(strV(o), o)) /* Convert in-place. Value used below. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
527 lj_trace_err(J, LJ_TRERR_BADTYPE); /* Punt if non-numeric. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
528 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
529 return tr;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
530 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
531
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
532 /* Narrowing of arithmetic operations. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
533 TRef lj_opt_narrow_arith(jit_State *J, TRef rb, TRef rc,
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
534 TValue *vb, TValue *vc, IROp op)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
535 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
536 rb = conv_str_tonum(J, rb, vb);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
537 rc = conv_str_tonum(J, rc, vc);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
538 /* Must not narrow MUL in non-DUALNUM variant, because it loses -0. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
539 if ((op >= IR_ADD && op <= (LJ_DUALNUM ? IR_MUL : IR_SUB)) &&
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
540 tref_isinteger(rb) && tref_isinteger(rc) &&
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
541 numisint(lj_vm_foldarith(numberVnum(vb), numberVnum(vc),
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
542 (int)op - (int)IR_ADD)))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
543 return emitir(IRTGI((int)op - (int)IR_ADD + (int)IR_ADDOV), rb, rc);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
544 if (!tref_isnum(rb)) rb = emitir(IRTN(IR_CONV), rb, IRCONV_NUM_INT);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
545 if (!tref_isnum(rc)) rc = emitir(IRTN(IR_CONV), rc, IRCONV_NUM_INT);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
546 return emitir(IRTN(op), rb, rc);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
547 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
548
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
549 /* Narrowing of unary minus operator. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
550 TRef lj_opt_narrow_unm(jit_State *J, TRef rc, TValue *vc)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
551 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
552 rc = conv_str_tonum(J, rc, vc);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
553 if (tref_isinteger(rc)) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
554 uint32_t k = (uint32_t)numberVint(vc);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
555 if ((LJ_DUALNUM || k != 0) && k != 0x80000000u) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
556 TRef zero = lj_ir_kint(J, 0);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
557 if (!LJ_DUALNUM)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
558 emitir(IRTGI(IR_NE), rc, zero);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
559 return emitir(IRTGI(IR_SUBOV), zero, rc);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
560 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
561 rc = emitir(IRTN(IR_CONV), rc, IRCONV_NUM_INT);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
562 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
563 return emitir(IRTN(IR_NEG), rc, lj_ir_ksimd(J, LJ_KSIMD_NEG));
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
564 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
565
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
566 /* Narrowing of modulo operator. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
567 TRef lj_opt_narrow_mod(jit_State *J, TRef rb, TRef rc, TValue *vb, TValue *vc)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
568 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
569 TRef tmp;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
570 rb = conv_str_tonum(J, rb, vb);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
571 rc = conv_str_tonum(J, rc, vc);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
572 if ((LJ_DUALNUM || (J->flags & JIT_F_OPT_NARROW)) &&
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
573 tref_isinteger(rb) && tref_isinteger(rc) &&
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
574 (tvisint(vc) ? intV(vc) != 0 : !tviszero(vc))) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
575 emitir(IRTGI(IR_NE), rc, lj_ir_kint(J, 0));
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
576 return emitir(IRTI(IR_MOD), rb, rc);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
577 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
578 /* b % c ==> b - floor(b/c)*c */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
579 rb = lj_ir_tonum(J, rb);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
580 rc = lj_ir_tonum(J, rc);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
581 tmp = emitir(IRTN(IR_DIV), rb, rc);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
582 tmp = emitir(IRTN(IR_FPMATH), tmp, IRFPM_FLOOR);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
583 tmp = emitir(IRTN(IR_MUL), tmp, rc);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
584 return emitir(IRTN(IR_SUB), rb, tmp);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
585 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
586
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
587 /* -- Predictive narrowing of induction variables ------------------------- */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
588
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
589 /* Narrow a single runtime value. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
590 static int narrow_forl(jit_State *J, cTValue *o)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
591 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
592 if (tvisint(o)) return 1;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
593 if (LJ_DUALNUM || (J->flags & JIT_F_OPT_NARROW)) return numisint(numV(o));
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
594 return 0;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
595 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
596
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
597 /* Narrow the FORL index type by looking at the runtime values. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
598 IRType lj_opt_narrow_forl(jit_State *J, cTValue *tv)
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
599 {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
600 lj_assertJ(tvisnumber(&tv[FORL_IDX]) &&
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
601 tvisnumber(&tv[FORL_STOP]) &&
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
602 tvisnumber(&tv[FORL_STEP]),
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
603 "expected number types");
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
604 /* Narrow only if the runtime values of start/stop/step are all integers. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
605 if (narrow_forl(J, &tv[FORL_IDX]) &&
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
606 narrow_forl(J, &tv[FORL_STOP]) &&
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
607 narrow_forl(J, &tv[FORL_STEP])) {
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
608 /* And if the loop index can't possibly overflow. */
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
609 lua_Number step = numberVnum(&tv[FORL_STEP]);
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
610 lua_Number sum = numberVnum(&tv[FORL_STOP]) + step;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
611 if (0 <= step ? (sum <= 2147483647.0) : (sum >= -2147483648.0))
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
612 return IRT_INT;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
613 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
614 return IRT_NUM;
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
615 }
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
616
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
617 #undef IR
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
618 #undef fins
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
619 #undef emitir
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
620 #undef emitir_raw
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
621
94705b5986b3 [ThirdParty] Added WRK and luajit for load testing.
MrJuneJune <me@mrjunejune.com>
parents:
diff changeset
622 #endif