@asmlift/core 0.9.0 → 0.10.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (101) hide show
  1. package/README.md +25 -24
  2. package/package.json +1 -1
  3. package/src/aggregate.ts +149 -0
  4. package/src/backend/c.ts +1 -0
  5. package/src/backend/cfamily.ts +165 -50
  6. package/src/backend/cpp.ts +2 -1
  7. package/src/backend/pascal.ts +31 -12
  8. package/src/contracts.ts +173 -16
  9. package/src/declare.ts +69 -20
  10. package/src/frontend/asmdata.ts +7 -1
  11. package/src/frontend/call-plan.ts +964 -60
  12. package/src/frontend/dead-reads.ts +628 -0
  13. package/src/frontend/device-pins.ts +23 -4
  14. package/src/frontend/emit.ts +4 -3
  15. package/src/frontend/fpu.ts +224 -17
  16. package/src/frontend/frame-objects.ts +1232 -192
  17. package/src/frontend/frontend.ts +64 -3
  18. package/src/frontend/high-half.ts +53 -4
  19. package/src/frontend/jump-table.ts +530 -0
  20. package/src/frontend/local-object.ts +158 -32
  21. package/src/frontend/mips.ts +2076 -215
  22. package/src/frontend/own-struct-return.ts +500 -0
  23. package/src/frontend/ppc.ts +2711 -232
  24. package/src/frontend/register-struct-call.ts +484 -0
  25. package/src/frontend/reloc-symbol.ts +22 -7
  26. package/src/frontend/ssa.ts +1460 -44
  27. package/src/frontend/stackargs.ts +106 -57
  28. package/src/frontend/thumb.ts +2044 -318
  29. package/src/ir/alias.ts +42 -0
  30. package/src/ir/bits.ts +681 -1
  31. package/src/ir/core.ts +100 -5
  32. package/src/ir/discipline.ts +86 -21
  33. package/src/ir/entry.ts +159 -0
  34. package/src/{structure → ir}/loops.ts +6 -2
  35. package/src/ir/opcodes.ts +109 -9
  36. package/src/ir/types.ts +86 -2
  37. package/src/ir/verify.ts +80 -11
  38. package/src/l3/ast.ts +290 -18
  39. package/src/l3/basecse.ts +33 -5
  40. package/src/l3/constinit.ts +382 -0
  41. package/src/l3/dce.ts +2 -2
  42. package/src/l3/initfirst.ts +42 -11
  43. package/src/l3/intcell.ts +2 -1
  44. package/src/l3/narrowdecl.ts +77 -36
  45. package/src/l3/pollguard.ts +8 -2
  46. package/src/l3/ptrcell.ts +31 -34
  47. package/src/l3/ptrfield.ts +123 -8
  48. package/src/l3/regspell.ts +9 -4
  49. package/src/l3/reindex.ts +283 -96
  50. package/src/l3/storage.ts +98 -1
  51. package/src/l3/symbol-refs.ts +84 -6
  52. package/src/l3/tailmerge.ts +4 -2
  53. package/src/l3/typing.ts +108 -15
  54. package/src/l3/unreduce.ts +6 -5
  55. package/src/l3/volatileptr.ts +17 -54
  56. package/src/mangle.ts +30 -0
  57. package/src/pattern/engine.ts +361 -73
  58. package/src/pipeline.ts +541 -24
  59. package/src/proto-context.ts +905 -45
  60. package/src/proto.ts +555 -40
  61. package/src/raise/aggmembers.ts +290 -0
  62. package/src/raise/arrays.ts +5 -1
  63. package/src/raise/declared-volatile.ts +207 -31
  64. package/src/raise/floathelpers.ts +254 -33
  65. package/src/raise/gvn.ts +3 -3
  66. package/src/raise/latch.ts +162 -6
  67. package/src/raise/memberarrays.ts +10 -4
  68. package/src/raise/offsetnames.ts +10 -26
  69. package/src/raise/pairparams.ts +33 -0
  70. package/src/raise/paramwidth.ts +196 -6
  71. package/src/raise/pre-recovery.ts +102 -19
  72. package/src/raise/pre-structure.ts +54 -0
  73. package/src/raise/recover.ts +12 -5
  74. package/src/raise/reroll.ts +694 -0
  75. package/src/raise/shortcircuit.ts +26 -12
  76. package/src/raise/softdiv.ts +1 -1
  77. package/src/raise/storesign.ts +153 -0
  78. package/src/raise/struct-arrays.ts +30 -11
  79. package/src/raise/structs.ts +396 -33
  80. package/src/raise/truncload.ts +4 -3
  81. package/src/raise/unthread.ts +400 -0
  82. package/src/rank-declare.ts +118 -12
  83. package/src/rank-variations.ts +89 -31
  84. package/src/rank.ts +307 -54
  85. package/src/runtime-helpers.ts +140 -22
  86. package/src/structure/analysis.ts +165 -2
  87. package/src/structure/arith-ops.ts +23 -0
  88. package/src/structure/globalaccess.ts +111 -9
  89. package/src/structure/hazards.ts +16 -64
  90. package/src/structure/local-statics.ts +145 -9
  91. package/src/structure/namecoalesce.ts +1 -1
  92. package/src/structure/pointer-spelling.ts +1106 -0
  93. package/src/structure/redundant-test.ts +1 -1
  94. package/src/structure/structure.ts +2180 -1054
  95. package/src/structure/switch-recover.ts +244 -100
  96. package/src/symbols.ts +6 -0
  97. package/src/target.ts +765 -52
  98. package/src/trace.ts +26 -8
  99. package/src/variation-definitions.ts +126 -14
  100. package/src/variation-gates.ts +2 -0
  101. package/src/variation-tokens.ts +5 -1
@@ -23,7 +23,9 @@
23
23
  // prototype (falling back to argument-register liveness). The frame bookkeeping — `stwu r1`,
24
24
  // the link-register save, callee-saved saves — is transparent to dataflow, so a value in a
25
25
  // callee-saved register survives the call; a word stored to the frame with a value is a
26
- // stack-slot variable (`frameStore`).
26
+ // stack-slot variable (`frameStore`). Inside the local area of a function that takes a frame
27
+ // address (`measurePpcFrame`), every access is memory through a frame object the shared audit
28
+ // proves instead.
27
29
  // • RECORD FORM (`.` = the Rc bit, e.g. `andi.`/`add.`) also sets cr0 from a signed compare of
28
30
  // the result against 0; that implicit compare is wired so a following `beq`/`bne` fuses.
29
31
  //
@@ -37,13 +39,45 @@
37
39
  // conditional-CTR forms (`bdz`/`bdnzt`/…), an unrecovered `bctr`, and a `bdnz` with no reaching
38
40
  // `mtctr`. NOTE mwcc at -O4 aggressively UNROLLS loops into a `bdnz` main loop + a remainder
39
41
  // loop; an unrolled loop recovers as the unrolled form (sound, rarely a match).
40
- import { Fn, Op, Successor, Value, mkOp, mkValue } from '../ir/core';
41
- import type { Opcode } from '../ir/opcodes';
42
+ import { ownArgumentWordsOn } from '../aggregate';
43
+ import {
44
+ Block,
45
+ Fn,
46
+ Op,
47
+ REGISTER_STRUCT,
48
+ STRUCT_HOMED,
49
+ Successor,
50
+ Value,
51
+ mkOp,
52
+ mkValue,
53
+ replaceAllUsesWith,
54
+ } from '../ir/core';
55
+ import { deletableWhenDead } from '../ir/discipline';
56
+ import { FCMP_TRUTH, NEGATED_FCMP, type Opcode, fcmpOfTruth } from '../ir/opcodes';
42
57
  import { T } from '../ir/types';
43
- import type { Prototypes } from '../proto';
58
+ import {
59
+ type FnProto,
60
+ type Prototypes,
61
+ declaredWidth,
62
+ declaresVoidReturn,
63
+ isFloatingSpelling,
64
+ ownFirstArgumentPassed,
65
+ wordsOf,
66
+ } from '../proto';
67
+ import type { ContextGlobals } from '../proto-context';
68
+ import { registerSaveRoutineOf } from '../runtime-helpers';
69
+ import type { SymbolMap } from '../symbols';
44
70
  import type { TargetDescription } from '../target';
45
71
  import { type AsmData, readJumpTable } from './asmdata';
46
- import { callDeclarations } from './call-plan';
72
+ import { type CallDeclarations, type ParamClass, callDeclarations } from './call-plan';
73
+ import {
74
+ type Declarations,
75
+ computedOnlyToReturn,
76
+ leftInDiscardRegister,
77
+ pinDeadReads,
78
+ unreadPairHalf,
79
+ } from './dead-reads';
80
+ import { pinDeviceAccesses } from './device-pins';
47
81
  import {
48
82
  type DisasmInstr,
49
83
  parseImm,
@@ -55,17 +89,64 @@ import { mkEmitKit, pushSwitchBr } from './emit';
55
89
  import { type FrontendRefusal, FrontendUnsupportedError } from './errors';
56
90
  import { inheritFlags } from './flags-edge';
57
91
  import { assertInputFormat } from './format';
58
- import { fpPrecision, fpuArgSlots, refuseBothPrecisions, writesFloatReturn } from './fpu';
59
- import type { Frontend } from './frontend';
92
+ import {
93
+ type ReturnHomeRule,
94
+ carriesFloatHome,
95
+ declaredPrecision,
96
+ floatHomeAt,
97
+ fpPrecision,
98
+ fpuArgSlots,
99
+ refuseBothPrecisions,
100
+ refuseFpWordCrossings,
101
+ returnHomeRule,
102
+ settleReturnHome,
103
+ writesFloatReturn,
104
+ } from './fpu';
105
+ import {
106
+ FRAME_OBJECT_AUDIT,
107
+ type FrameObjectRelift,
108
+ type FrameRange,
109
+ areaAboveOutgoing,
110
+ landsOnArea,
111
+ liftWithRelift,
112
+ } from './frame-objects';
113
+ import { type Frontend, type LiftOptions, type OwnDeclaration, ownDeclarationOf } from './frontend';
60
114
  import { makeHighHalves } from './high-half';
115
+ import {
116
+ type HoistedDispatch,
117
+ type IdiomLeft,
118
+ type StaleIdiomRead,
119
+ type TableKey,
120
+ agreedTable,
121
+ caseEntries,
122
+ enterCaseTarget,
123
+ entersBetween,
124
+ idiomLeaves,
125
+ makeTableBases,
126
+ parkedIn,
127
+ refuseReachableDispatch,
128
+ refuseStaleIdiomReads,
129
+ scaledIndexIn,
130
+ } from './jump-table';
61
131
  import { makeLocalStatics, readObjectLocalObject } from './local-object';
62
132
  import { opaqueDest } from './opaque';
63
133
  import { classifyRelocSymbol, unspellableReason } from './reloc-symbol';
64
- import { abiSortEntryParams, mintArgSlotHoles, stackSlotKey } from './ssa';
65
- import { makeSsaBuilder } from './ssa';
134
+ import {
135
+ type ArgSlots,
136
+ abiSortEntryParams,
137
+ incomingStackArgKey,
138
+ mintArgSlotHoles,
139
+ stackArgSlots,
140
+ stackSlotKey,
141
+ } from './ssa';
142
+ import { makeSsaBuilder, refuseSavedRegisterLeaks, registerLiveIns } from './ssa';
66
143
 
67
144
  type Instr = DisasmInstr;
68
145
 
146
+ /** The instructions that READ the carry bit (each also writes it), and every other writer of it. */
147
+ const CARRY_READERS = /^(adde|addme|addze|subfe|subfme|subfze)o?$/;
148
+ const CARRY_WRITERS = /^(addc|addco|addic|subfc|subfco|subfic|srawi|sraw|mtxer|mcrxr)$/;
149
+
69
150
  // Branch-condition mnemonic → the icmp for its TAKEN edge, split by compare signedness. PowerPC's
70
151
  // branch already names the relation (no negation fold needed). `signed` picks the row.
71
152
  const COND_SIGNED: Record<string, Opcode> = {
@@ -125,9 +206,11 @@ const isReg = (s: string | undefined): s is string => /^r\d+$/.test(s ?? '');
125
206
  * in; `fmr` is a copy and has no opcode. `fneg` and `fmr` carry no precision at all — an FPR holds a
126
207
  * double either way — so they take the function's (`frontend/fpu.ts` `fpPrecision`); in a function
127
208
  * of only those, the floats state none (`T.fUnstated`), because the float and double spellings of a
128
- * negation or a copy compile to one object. Everything else in the file keeps the register-file refusal `opaqueDest` gives it: `frsp`, the
129
- * loads and stores, the conversions, the compares, the fused multiply-adds, paired singles — and a
130
- * record form (`fadds.`), which sets `cr1`. */
209
+ * negation or a copy compile to one object. The compares and `fctiwz` (`FP_CROSSINGS`) are decoded
210
+ * in their own cases. Everything else in the file keeps the register-file refusal `opaqueDest`
211
+ * gives it: `frsp`, the loads and stores but an integer word's `stfd` to the frame, the other
212
+ * conversions, the fused multiply-adds, paired singles — and a record form (`fadds.`), which sets
213
+ * `cr1`. */
131
214
  const FP_ARITH: Readonly<Record<string, { op: Opcode | 'copy'; width: 32 | 64 | null }>> = {
132
215
  fadds: { op: 'fadd', width: 32 },
133
216
  fsubs: { op: 'fsub', width: 32 },
@@ -141,6 +224,10 @@ const FP_ARITH: Readonly<Record<string, { op: Opcode | 'copy'; width: 32 | 64 |
141
224
  fmr: { op: 'copy', width: null },
142
225
  };
143
226
  const FP_ARITH_MNEMONICS: ReadonlySet<string> = new Set(Object.keys(FP_ARITH));
227
+ /** The FPU instructions that move a float's information into the integer file, none of which states
228
+ * a precision: the compares (an FPR holds a double either way, and `a > b` over two floats is the
229
+ * same `fcmpo` as over two doubles) and `fctiwz`, the truncation to an integer word. */
230
+ const FP_CROSSINGS: ReadonlySet<string> = new Set(['fcmpo', 'fcmpu', 'fctiwz']);
144
231
 
145
232
  /** An FPU register: the token, and the key the SSA builder holds it under. ONE predicate for the
146
233
  * decode above and the register-file refusal below (`fpReg`), so a token one of them reads as an
@@ -171,8 +258,30 @@ const rlwinmMask = (mb: number, me: number): number => {
171
258
  const parseDisasm = (disasm: string): Instr[] => parseSharedDisasm(disasm, { relocs: true, hintSuffixes: true });
172
259
 
173
260
  /** A compare a `cr` field holds: its operands and signedness, which the reading branch turns into a
174
- * condition. */
175
- type CrCompare = { lhs: Value; rhs: Value; signed: boolean };
261
+ * condition. A FLOAT compare (`fcmpo`/`fcmpu`) states instead which compare each of the field's
262
+ * three bits holds, since a `cror` can change one (`bits`); `quiet` is `fcmpu`'s, which C emits for
263
+ * `==` and `!=` alone, so an ordering bit of it is refused. */
264
+ type CrCompare = { lhs: Value; rhs: Value; signed: boolean; bits?: CrBits; quiet?: boolean };
265
+ type CrBits = { readonly lt: Opcode; readonly gt: Opcode; readonly eq: Opcode };
266
+ /** The bit of a field each branch reads, and whether it is taken when the bit is SET: `ble` is
267
+ * "branch if not greater than", which after a float compare is `!(a > b)`, true on a NaN — never
268
+ * `a <= b`. */
269
+ const BRANCH_BIT: Readonly<Record<string, { bit: keyof CrBits; set: boolean }>> = {
270
+ blt: { bit: 'lt', set: true },
271
+ bge: { bit: 'lt', set: false },
272
+ bgt: { bit: 'gt', set: true },
273
+ ble: { bit: 'gt', set: false },
274
+ beq: { bit: 'eq', set: true },
275
+ bne: { bit: 'eq', set: false },
276
+ };
277
+ /** A condition-register bit as `cror` and the `mfcr` extracts name it: its field and its place in
278
+ * the field, or null for the summary-overflow bit (`so`), which a float compare sets on a NaN and
279
+ * nothing here reads. `eq` names cr0's; `4*cr1+eq` cr1's. */
280
+ const crBitOf = (tok: string): { field: string; bit: keyof CrBits } | null => {
281
+ const m = /^(?:4\*cr([0-7])\+)?(lt|gt|eq)$/.exec(tok);
282
+ return m ? { field: `cr${m[1] ?? 0}`, bit: m[2] as keyof CrBits } : null;
283
+ };
284
+ const CR_BIT_NAMES: readonly (keyof CrBits | null)[] = ['lt', 'gt', 'eq', null];
176
285
 
177
286
  /** What a block leaves in the condition register for a successor to inherit. `fields` holds, per
178
287
  * field, the compare it holds, or the sentence saying what overwrote it. A field absent from the map
@@ -192,99 +301,214 @@ interface PpcBlock {
192
301
  }
193
302
 
194
303
  // A recovered dense-switch jump table (Regime B), keyed
195
- // by the BOUNDS branch (`bgt DEF`) whose block emits the `switch_br`. `caseAddrs[k]` is the `.text`
304
+ // by the BOUNDS branch (`bgt DEF`, or `bgtlr`) whose block emits the `switch_br`. `caseAddrs[k]` is the `.text`
196
305
  // address of case `k` (dense 0..N-1); `bctrAddr` is the elided dispatch's indirect jump. The mwcc
197
306
  // idiom:
198
307
  // cmplwi rS,N-1 ; bgt DEF (bounds)
199
308
  // lis rT,0 [ADDR16_HA @tbl] ; slwi rIdx,rS,2 ; addi rB,rT,0 [ADDR16_LO @tbl]
200
309
  // ; lwzx rV,rB,rIdx ; mtctr rV ; bctr (dispatch — table in .data)
310
+ // mwcc schedules other words into the bounds pair (`boundsCompare`), orders the dispatch's freely,
311
+ // and 2.3.3 forms the table base outside the dispatch (`hoistedTable`).
201
312
  interface PpcJT {
202
313
  scrutReg: string;
314
+ /** The bounds compare's address: nothing may enter after it, up to the `bctr`. */
315
+ cmpAddr: number;
203
316
  caseAddrs: number[];
204
- defaultAddr: number;
317
+ /** The default arm's address, or null for a `bgtlr` bound: out-of-range values return. */
318
+ defaultAddr: number | null;
205
319
  bctrAddr: number;
320
+ /** The base register and table of a dispatch whose block does not form its base (mwcc 2.3.3
321
+ * hoists the `lis`/`addi` pair out of a loop): `lift` proves the base reaches it
322
+ * (frontend/jump-table.ts). Null when the pruned block forms the base itself. */
323
+ hoisted: HoistedDispatch | null;
324
+ /** What the pruned dispatch leaves in each register it writes, `ctr` included (`idiomLeaves`). */
325
+ leaves: ReadonlyMap<string, IdiomLeft>;
326
+ }
327
+
328
+ /** The `cr` field a compare writes or a conditional branch reads: its first operand when it names
329
+ * one, else cr0. */
330
+ const crFieldOf = (ins: Instr): string => (ins.ops[0]?.startsWith('cr') ? ins.ops[0] : 'cr0');
331
+
332
+ /** Whether `ins` may write a CR field. A DENY-LIST, so an unknown writer stops the reader rather
333
+ * than being stepped past: the compares, the record forms (`.`), the CR-logical and move-to-CR
334
+ * forms, and every branch — a `bl` destroys the volatile fields. */
335
+ const mayWriteCr = (ins: Instr): boolean =>
336
+ /^(cmp|fcmp|cr|mcr|mtcr|b)/.test(ins.mnemonic) || ins.mnemonic.endsWith('.');
337
+
338
+ /** Whether `ins` may write GPR `r`. A DENY-LIST like mips.ts's `writtenReg`: an instruction writes
339
+ * its first operand unless it is a store, compare, trap, move-to-special or cache op; an update form
340
+ * (`lwzu`, `stwux`) also writes its base, `lmw rD` writes rD..r31, and any branch may (a call
341
+ * clobbers the volatile registers). */
342
+ const mayWriteGpr = (ins: Instr, r: string): boolean => {
343
+ const m = ins.mnemonic;
344
+ if (m.startsWith('b')) {
345
+ return true;
346
+ }
347
+ if (m === 'lmw') {
348
+ return /^r\d+$/.test(r) && Number(r.slice(1)) >= Number(ins.ops[0]?.slice(1));
349
+ }
350
+ if (/^(l|st)\w*ux?$/.test(m) && ins.ops.slice(1).some((o) => new RegExp(`\\b${r}\\b`).test(o))) {
351
+ return true;
352
+ }
353
+ return !/^(st|cmp|fcmp|tw|mt|dc|ic|sync|isync|eieio)/.test(m) && ins.ops[0] === r;
354
+ };
355
+
356
+ /** The bounds compare of a jump table, or null: the `cmplwi rS,N-1` the bounds branch at `j` reads,
357
+ * with its address and the case count `n` it states. Scanned back inside the branch's block, past
358
+ * instructions that write neither a CR field nor `rS` — mwcc schedules the default's value or the
359
+ * LR save between the two — and it must set the CR field the branch tests. The first CR writer met
360
+ * decides: any other one, or a write of `rS` after the compare, declines. So does a branch landing
361
+ * after the compare, up to the bounds branch itself: that path brings a CR field of its own, and the
362
+ * `switch_br` would bound it by a compare it never ran (`entersBetween`). */
363
+ function boundsCompare(instrs: Instr[], j: number, scrutReg: string): { addr: number; n: number } | null {
364
+ const crField = crFieldOf(instrs[j]);
365
+ for (let k = j - 1; k >= 0 && !isXfer(instrs[k]); k--) {
366
+ const c = instrs[k];
367
+ if (c.mnemonic === 'cmplwi' && c.ops[c.ops.length - 2] === scrutReg && crFieldOf(c) === crField) {
368
+ return entersBetween(instrs, c.addr, instrs[j].addr)
369
+ ? null
370
+ : { addr: c.addr, n: parseImm(c.ops[c.ops.length - 1]) + 1 };
371
+ }
372
+ if (mayWriteCr(c) || mayWriteGpr(c, scrutReg)) {
373
+ return null;
374
+ }
375
+ }
376
+ return null;
206
377
  }
207
378
 
379
+ /** The table a base formed OUTSIDE the dispatch block addresses, as far as the instructions say:
380
+ * every `addi rB,rX,SYM@l` in the function must name one table, or the guess is left undecided.
381
+ * mwcc 2.3.3 forms the base once at entry and dispatches from a loop. Only a guess — `lift` pairs
382
+ * the halves and proves the base reaches the dispatch (frontend/jump-table.ts). */
383
+ const hoistedTable = (instrs: Instr[], rB: string): TableKey | null =>
384
+ agreedTable(
385
+ instrs.flatMap((x) =>
386
+ x.mnemonic === 'addi' && x.ops[0] === rB && x.reloc?.type === 'R_PPC_ADDR16_LO' && parseImm(x.ops[2]) === 0
387
+ ? [{ sym: x.reloc.sym, addend: x.reloc.addend }]
388
+ : [],
389
+ ),
390
+ );
391
+
208
392
  // Recover mwcc jump tables from the dispatch idiom + the AsmData side-table. Fail-closed: any
209
- // deviation from the exact idiom, or a table that doesn't resolve to N in-function targets, declines
393
+ // deviation from the idiom, or a table that doesn't resolve to N in-function targets, declines
210
394
  // (→ the `bctr` loud-fail fires). Index IDENTITY guard: `slwi rIdx,rS,2` must be the bounds-checked
211
395
  // scrutinee scaled only by <<2 — no xor/neg/extra op.
212
396
  function recoverPpcJumpTables(instrs: Instr[], ad: AsmData): Map<number, PpcJT> {
213
397
  const out = new Map<number, PpcJT>();
214
- for (let i = 5; i < instrs.length; i++) {
398
+ for (let i = 0; i < instrs.length; i++) {
215
399
  if (instrs[i].mnemonic !== 'bctr') {
216
400
  continue;
217
401
  }
218
- const [lis, lwzx, mtctr] = [instrs[i - 5], instrs[i - 2], instrs[i - 1]];
219
- // The index is scaled and the table's low half formed in either order: mwcc 2.4.2 emits the
220
- // `slwi` first, GC/1.2.5n the `addi`. The two instructions are independent.
221
- const [slwi, addi] =
222
- instrs[i - 4].mnemonic === 'slwi' ? [instrs[i - 4], instrs[i - 3]] : [instrs[i - 3], instrs[i - 4]];
223
- if (
224
- lis.mnemonic !== 'lis' ||
225
- slwi.mnemonic !== 'slwi' ||
226
- addi.mnemonic !== 'addi' ||
227
- lwzx.mnemonic !== 'lwzx' ||
228
- mtctr.mnemonic !== 'mtctr'
229
- ) {
402
+ // THE DISPATCH BLOCK: after the bounds branch, the nearest transfer before the `bctr`.
403
+ let j = i - 1;
404
+ while (j >= 0 && !isXfer(instrs[j])) {
405
+ j--;
406
+ }
407
+ const bgt = instrs[j];
408
+ // `bgt DEF`, or `bgtlr` when the default arm is the return itself: mwcc bounds a switch whose
409
+ // default falls straight out of a leaf that way.
410
+ if (!bgt || !((bgt.mnemonic === 'bgt' && bgt.target !== undefined) || bgt.mnemonic === 'bgtlr')) {
230
411
  continue;
231
412
  }
232
- const rV = mtctr.ops[0];
233
- if (lwzx.ops[0] !== rV) {
413
+ // Each piece of the idiom is the LAST write of its register inside the block, so nothing between
414
+ // a def and its reader can have changed it. The order varies: mwcc 2.4.2 scales the index first,
415
+ // GC/1.2.5n forms the table's low half first.
416
+ const defIn = (r: string, before: number): number => {
417
+ for (let k = before - 1; k > j; k--) {
418
+ if (mayWriteGpr(instrs[k], r)) {
419
+ return k;
420
+ }
421
+ }
422
+ return -1;
423
+ };
424
+ const mi = i - 1;
425
+ const mtctr = instrs[mi];
426
+ if (mtctr?.mnemonic !== 'mtctr' || mi <= j) {
234
427
  continue;
235
- } // rV = mem[rB + rIdx]
428
+ }
429
+ const li = defIn(mtctr.ops[0], mi); // rV = mem[rB + rIdx]
430
+ const lwzx = instrs[li];
431
+ if (lwzx?.mnemonic !== 'lwzx') {
432
+ continue;
433
+ }
236
434
  const [rB, rIdx] = [lwzx.ops[1], lwzx.ops[2]];
237
- if (slwi.ops[0] !== rIdx || (slwi.ops[2] !== '2' && slwi.ops[2] !== '0x2')) {
435
+ const si = defIn(rIdx, li);
436
+ const slwi = instrs[si];
437
+ if (slwi?.mnemonic !== 'slwi' || (slwi.ops[2] !== '2' && slwi.ops[2] !== '0x2')) {
238
438
  continue;
239
439
  } // identity: rIdx = rS<<2
240
440
  const scrutReg = slwi.ops[1];
241
- if (addi.ops[0] !== rB || parseImm(addi.ops[2]) !== 0) {
441
+ const ai = defIn(rB, li);
442
+ const lisI = ai < 0 ? -1 : defIn(instrs[ai].ops[1], ai);
443
+ const hoisted = ai < 0 ? hoistedTable(instrs, rB) : null;
444
+ let table: TableKey | null = hoisted;
445
+ if (ai >= 0) {
446
+ const [lis, addi] = [instrs[lisI], instrs[ai]];
447
+ // The table's base is the same `@ha`/`@l` pair the address fold consumes and must show the
448
+ // same evidence — both relocation types, one symbol, one addend — so a dispatch is never
449
+ // recovered from two instructions that never addressed the table together. It stays a
450
+ // separate pairing: it runs before the blocks exist, and it wants the table's NAME, for
451
+ // `readJumpTable`, not its address. It skips the naming policy deliberately — a table base is
452
+ // read here as data, never written into the candidate as an identifier — so the `@N` and
453
+ // `.`-prefixed symbols the address fold must refuse (292 of the corpus's `ADDR16` relocations)
454
+ // are harmless here.
455
+ table =
456
+ addi.mnemonic === 'addi' &&
457
+ parseImm(addi.ops[2]) === 0 &&
458
+ lis?.mnemonic === 'lis' &&
459
+ lis.ops[0] === addi.ops[1] &&
460
+ lis.reloc?.type === 'R_PPC_ADDR16_HA' &&
461
+ addi.reloc?.type === 'R_PPC_ADDR16_LO' &&
462
+ addi.reloc.sym === lis.reloc.sym &&
463
+ addi.reloc.addend === lis.reloc.addend
464
+ ? { sym: lis.reloc.sym, addend: lis.reloc.addend }
465
+ : null;
466
+ }
467
+ if (table === null) {
242
468
  continue;
243
- } // rB = &table (lo) + 0
244
- const rT = addi.ops[1];
245
- // The table's base is the same `@ha`/`@l` pair the address fold consumes and must show the same
246
- // evidence — both relocation types, one symbol, one addend — so a dispatch is never recovered
247
- // from two instructions that never addressed the table together. It stays a separate pairing:
248
- // it runs before the blocks exist, and it wants the table's NAME, for `readJumpTable`, not its
249
- // address. It skips the naming policy deliberately — a table base is read here as data, never
250
- // written into the candidate as an identifier — so the `@N` and `.`-prefixed symbols the
251
- // address fold must refuse (292 of the corpus's `ADDR16` relocations) are harmless here.
252
- const tableSym = lis.reloc?.sym;
469
+ }
470
+ // The block is pruned unread once the table is recovered, so it may hold the idiom and nothing
471
+ // else, nothing may branch into it past the bounds check, and none of the idiom may rewrite the
472
+ // scrutinee before the `slwi` scales it — the `switch_br` reads it at the bounds branch.
473
+ const idiom = new Set(hoisted === null ? [lisI, si, ai, li, mi] : [si, li, mi]);
253
474
  if (
254
- lis.ops[0] !== rT ||
255
- !tableSym ||
256
- lis.reloc?.type !== 'R_PPC_ADDR16_HA' ||
257
- addi.reloc?.type !== 'R_PPC_ADDR16_LO' ||
258
- addi.reloc.sym !== tableSym ||
259
- addi.reloc.addend !== lis.reloc.addend
475
+ instrs.slice(j + 1, i).some((_, o) => !idiom.has(j + 1 + o)) ||
476
+ instrs.slice(j + 1, si).some((x) => mayWriteGpr(x, scrutReg)) ||
477
+ entersBetween(instrs, bgt.addr, instrs[i].addr)
260
478
  ) {
261
479
  continue;
262
480
  }
263
- // Bounds: the nearest preceding `cmplwi scrutReg,N-1 ; bgt DEF` guard. The `bgt` is itself a
264
- // transfer, so match the pair directly (the cmplwi sits one instruction behind it); a different
265
- // transfer before the guard ⇒ decline.
266
- let bounds: { addr: number; n: number; def: number } | null = null;
267
- for (let j = i - 6; j >= 1; j--) {
268
- const bgt = instrs[j];
269
- if (bgt.mnemonic === 'bgt' && bgt.target !== undefined) {
270
- const c = instrs[j - 1];
271
- if (c && c.mnemonic === 'cmplwi' && c.ops[c.ops.length - 2] === scrutReg) {
272
- bounds = { addr: bgt.addr, n: parseImm(c.ops[c.ops.length - 1]) + 1, def: bgt.target };
273
- }
274
- break;
275
- }
276
- if (isXfer(bgt)) {
277
- break;
278
- }
279
- }
280
- if (!bounds || bounds.n < 2) {
481
+ const bound = boundsCompare(instrs, j, scrutReg);
482
+ if (bound === null || bound.n < 2) {
281
483
  continue;
282
484
  }
283
- const caseAddrs = readJumpTable(ad, tableSym, 0, bounds.n);
485
+ const caseAddrs = readJumpTable(ad, table.sym, table.addend, bound.n);
284
486
  if (!caseAddrs) {
285
487
  continue;
286
488
  }
287
- out.set(bounds.addr, { scrutReg, caseAddrs, defaultAddr: bounds.def, bctrAddr: instrs[i].addr });
489
+ // In program order, `mtctr` writing `ctr`, the `slwi` the scaled scrutinee, every other word its
490
+ // first operand (`lis`, `addi`, `lwzx`).
491
+ const leaves = idiomLeaves(
492
+ [...idiom].sort((a, b) => a - b).map((k) => ({ reg: k === mi ? 'ctr' : instrs[k].ops[0], scaled: k === si })),
493
+ );
494
+ out.set(bgt.addr, {
495
+ scrutReg,
496
+ cmpAddr: bound.addr,
497
+ caseAddrs,
498
+ defaultAddr: bgt.target ?? null,
499
+ bctrAddr: instrs[i].addr,
500
+ hoisted: hoisted === null ? null : { reg: rB, table: hoisted },
501
+ leaves,
502
+ });
503
+ }
504
+ // A table's case is a branch target too, which `entersBetween` cannot see before the tables are
505
+ // read: one landing between a table's bounds compare and its indirect jump — its own table's or
506
+ // another's — enters that window as a branch would, and that table declines.
507
+ const cases = [...out.values()].flatMap((jt) => jt.caseAddrs);
508
+ for (const [at, jt] of out) {
509
+ if (cases.some((a) => a > jt.cmpAddr && a <= jt.bctrAddr)) {
510
+ out.delete(at);
511
+ }
288
512
  }
289
513
  return out;
290
514
  }
@@ -307,7 +531,9 @@ function toBlocks(instrs: Instr[], name: string, jts: Map<number, PpcJT>): { blo
307
531
  for (const a of jt.caseAddrs) {
308
532
  leaders.add(a);
309
533
  }
310
- leaders.add(jt.defaultAddr);
534
+ if (jt.defaultAddr !== null) {
535
+ leaders.add(jt.defaultAddr);
536
+ }
311
537
  }
312
538
 
313
539
  const blocks: PpcBlock[] = [];
@@ -328,6 +554,13 @@ function toBlocks(instrs: Instr[], name: string, jts: Map<number, PpcJT>): { blo
328
554
  const idxOf = new Map(blocks.map((b, i) => [b.startAddr, i]));
329
555
  const fallAddr = (b: PpcBlock) => (b.branch ?? b.body[b.body.length - 1]).addr + 4;
330
556
  const succIdx: number[][] = blocks.map(() => []);
557
+ /** A conditional return's taken edge: a block of its own that returns, appended past the real ones.
558
+ * Its index. */
559
+ const synthReturn = (): number => {
560
+ blocks.push({ startAddr: -1 - blocks.length, body: [], branch: null, synthReturn: true });
561
+ succIdx.push([]);
562
+ return blocks.length - 1;
563
+ };
331
564
  blocks.forEach((b, i) => {
332
565
  const br = b.branch;
333
566
  // Recovered switch: the bounds block dispatches to its case blocks + default (the `cmplwi`/`bgt`
@@ -335,7 +568,13 @@ function toBlocks(instrs: Instr[], name: string, jts: Map<number, PpcJT>): { blo
335
568
  // leader, so all resolve; a target that is not a block boundary makes it malformed → loud-fail.
336
569
  const jt = br ? jts.get(br.addr) : undefined;
337
570
  if (jt) {
338
- const succ = [...jt.caseAddrs, jt.defaultAddr].map((a) => idxOf.get(a));
571
+ const succ: Array<number | undefined> = jt.caseAddrs.map((a) => idxOf.get(a));
572
+ if (jt.defaultAddr !== null) {
573
+ succ.push(idxOf.get(jt.defaultAddr));
574
+ } else {
575
+ // A `bgtlr` bound: the default edge is a synthetic return, as for any conditional return.
576
+ succ.push(synthReturn());
577
+ }
339
578
  if (succ.some((x) => x === undefined)) {
340
579
  throw new PpcUnsupportedError(`cannot lift '${name}': jump-table target is not a block boundary`);
341
580
  }
@@ -359,17 +598,15 @@ function toBlocks(instrs: Instr[], name: string, jts: Map<number, PpcJT>): { blo
359
598
  return;
360
599
  }
361
600
  if (isCondReturn(br)) {
362
- const synth: PpcBlock = { startAddr: -1 - blocks.length, body: [], branch: null, synthReturn: true };
363
- const si = blocks.length;
364
- blocks.push(synth);
365
- succIdx.push([]);
601
+ const si = synthReturn();
366
602
  const fall = idxOf.get(br.addr + 4); // fall = instruction after the cond-return (no delay slot)
367
603
  succIdx[i] = fall !== undefined ? [si, fall] : [si];
368
604
  return;
369
605
  }
370
606
  // conditional branch to a label: [taken, fall]. Both must land on block boundaries; otherwise the
371
- // branch leaves the function (a tail branch) or targets unrecovered flow — fail LOUD and catchably
607
+ // branch falls off the function's end or targets unrecovered flow — fail LOUD and catchably
372
608
  // rather than silently dropping an edge (which surfaces as an opaque `verify` successor-count error).
609
+ // A branch out of the function, relocated or not, has already refused as a tail call (`liftPrelude`).
373
610
  const taken = idxOf.get(br.target!);
374
611
  const fall = idxOf.get(br.addr + 4);
375
612
  if (taken === undefined || fall === undefined) {
@@ -402,10 +639,181 @@ function toBlocks(instrs: Instr[], name: string, jts: Map<number, PpcJT>): { blo
402
639
  };
403
640
  }
404
641
 
642
+ /** The PPC EABI's small-data base registers: r13 points into `.sdata`/`.sbss` and r2 into
643
+ * `.sdata2`/`.sbss2`, both set once by the runtime's startup code and never by a compiled function.
644
+ * A compiled access names its symbol in an `R_PPC_EMB_SDA21` relocation, which this frontend reads
645
+ * instead of the base (`sdaAccess`). A base read as a VALUE — `lbzx r0,r3,r13`, which mwcc 2.4.7
646
+ * emits for `playerOrderNew[j - 1]` (MP4 `OrderPlayers`), or `lwz r3,-0x7ff0(r13)` in a linked image —
647
+ * has no C spelling here, and the SSA builder would otherwise read the def-less register as an
648
+ * invented parameter. It refuses, as MIPS refuses `gp` used as data. */
649
+ const SMALL_DATA_BASES: readonly string[] = ['r13', 'r2'];
650
+
405
651
  /** The link register's save word, as an offset from the entry r1: the second word of the caller's
406
652
  * frame header, after the back chain. mwcc 2.3.3 spells it `4(r1)` before the push and 2.4.x
407
653
  * `N+4(r1)` after it (`r1Displacements`). */
408
654
  const LR_SAVE_WORD = 4;
655
+ /** The caller's parameter area, as an offset from the entry r1: past the back chain and the link
656
+ * register's save word, the words a call takes past its eight argument registers. */
657
+ const PARAM_AREA_AT = 8;
658
+ /** Every load that takes an `off(rA)` operand: integer, multiple, update and floating-point
659
+ * (`outgoing-area-probes.test.ts` classifies every memory mnemonic of the committed probes against it). */
660
+ export const PPC_FRAME_LOAD = /^(lbz|lhz|lha|lwz|lmw|lbzu|lhzu|lhau|lwzu|lfs|lfd|lfsu|lfdu|psq_l|psq_lu)$/;
661
+
662
+ /** The register save or restore routine a `bl` calls (runtime-helpers.ts `RegisterSaveRoutines`),
663
+ * named by its `R_PPC_REL24`, or null for any other `bl`. Only the relocation names it: a `bl` with
664
+ * none — a linked image, or an objdump without `-r` — is no routine, and `case 'bl'` refuses it
665
+ * there as a call with no callee symbol. */
666
+ const saveRoutineOf = (target: TargetDescription, ins: Instr) =>
667
+ ins.mnemonic === 'bl' && ins.reloc?.type === 'R_PPC_REL24'
668
+ ? registerSaveRoutineOf(target.registerSaveRoutines, ins.reloc.sym)
669
+ : null;
670
+
671
+ /** A general register a callee must preserve: one the EABI neither passes an argument in, nor lets
672
+ * a callee destroy, nor reserves (r1, and the small-data bases). `prologueSavedRegs` counts the
673
+ * first store of ANY register in `regs`; only these are saves of the caller's value in `offsets`. */
674
+ const isCalleeSaved = (target: TargetDescription, r: string): boolean =>
675
+ isReg(r) &&
676
+ r !== 'r1' &&
677
+ !SMALL_DATA_BASES.includes(r) &&
678
+ !target.argRegs.includes(r) &&
679
+ !target.callerSaved.includes(r);
680
+
681
+ /** Whether any of `body` names a callee-saved general register (`isCalleeSaved`), as an operand or
682
+ * a memory base, that the prologue did not save — the mark of hand-written asm, since compiled
683
+ * code saves such a register before it touches it. A save routine or `stmw` names only its first
684
+ * register, and saves it, so either is in `saved`.
685
+ *
686
+ * A GLOBAL REGISTER VARIABLE'S register (`target.globalRegisterVariables`) does not count: it is
687
+ * the one callee-saved register COMPILED code names unsaved, and the live-in model reads it as that
688
+ * global (`registerLiveIns`). Counted, it would make the detector and the model disagree about one
689
+ * register, and drop the bound a compiled body's own declaration states: mwcc 2.4.2's
690
+ * `register int g __asm__("r14"); int f(int a) { int x, y; if (a) { x = a; y = a + 1; } … a * g`
691
+ * homes x and y in r4/r5, and `f(s32)` would lift as `f(s32 a0, s32 a1, s32 a2)` with no note. The
692
+ * blind side is a hand-written entry point handed inputs in r14..r28 and in argument registers
693
+ * past its declaration, which then reads those as `uninit_<reg>`; asm that names no other unsaved
694
+ * callee-saved register is byte-for-byte the compiled shape. No function of the ac, mp4 or Pikmin
695
+ * builds names an unsaved r14..r28 at all. */
696
+ const namesUnsavedCalleeSaved = (
697
+ body: readonly Instr[],
698
+ target: TargetDescription,
699
+ saved: ReadonlySet<string>,
700
+ ): boolean =>
701
+ body.some((ins) =>
702
+ ins.ops.some((op) => {
703
+ const reg = op.includes('(') ? parseMem(op).base : op;
704
+ return (
705
+ isCalleeSaved(target, reg) && !saved.has(reg) && !Object.hasOwn(target.globalRegisterVariables?.regs ?? {}, reg)
706
+ );
707
+ }),
708
+ );
709
+
710
+ /** WHAT THE PROLOGUE SAVED, AND WHERE. `regs` is the save half of the register partition
711
+ * (frontend/ssa.ts, `LiveInModel.uninitRegs`). A register the ABI passes no argument in is an
712
+ * uninitialised local only where the compiler could home a local there, which it may do only after
713
+ * saving it; without this, an uninitialised local in a callee-saved register lifts as an extra
714
+ * PARAMETER (`int x; if (s) x = k(a); return k(a) + x;` would take a third).
715
+ *
716
+ * Read off the ENTRY block, the one every path starts with: `stmw rS,D(r1)` saves rS..r31, a GPR
717
+ * save routine rNN..r31, and `stw rN,D(r1)` rN. A register counts as saved only where NO earlier
718
+ * instruction names it, as an operand or as a memory base, read or written. That is what a save
719
+ * is: the first thing the function does with the register. A write before the store makes the
720
+ * `stw` a spill; a READ before it is asm handed a live value there, which the save then preserves
721
+ * — the very case the save half exists to keep a parameter for, so it must not count.
722
+ *
723
+ * NOT Thumb's rule (`savedRegs` in frontend/thumb.ts), which ends the measurement at the end of the
724
+ * LEADING run of saves. mwcc interleaves its saves with the body's first instructions
725
+ * (`stw r31,12(r1); mr r31,r3; stw r30,8(r1)`), and the save routine's `addi r11,r1,N` need not sit
726
+ * next to its `bl`, so a leading run would drop saves the prologue really makes. The two rules
727
+ * agree on what they let through: no instruction before a counted save touches the register. The
728
+ * operand text can only over-name (a register spelt in a branch target names nothing), which
729
+ * under-states the set — the direction that keeps a parameter.
730
+ *
731
+ * `offsets` are where the saves of the caller's values lie, as entry-r1 offsets, which bound the
732
+ * local area from above (`measurePpcFrame`): each `stw` of a callee-saved register (`isCalleeSaved`)
733
+ * the same first-naming rule counts, each such word of a `stmw`, each word a save routine stores
734
+ * below its frame address (`addi r11,r1,N`), and each floating-point save (`stfd`, or the FPR
735
+ * routine), which this frontend refuses elsewhere but which still bounds where a local may lie.
736
+ * Null where a save's offset cannot be named: r1's depth at it, or the routine's frame address, is
737
+ * unknown.
738
+ *
739
+ * A SAVE IS RELOADED. In a function that returns, an `stw` of a callee-saved register that no
740
+ * reload of it from r1 answers anywhere — an `lwz` of it, an `lmw` or restore routine reaching it —
741
+ * saves nothing: the caller gets the register back as the body left it, so the word is a value.
742
+ * mwcc 2.4 stores a file-scope global register variable passed on the stack exactly so, in the
743
+ * entry block — `register int g __asm__("r14"); return k(1, …, 8, g) + g;` is `stw r14,8(r1); bl k`
744
+ * and no reload of r14 — and counting it as a save would turn the read of `g` after the call into
745
+ * `uninit_r14` and drop the ninth argument. Which word a reload reads is judged where it is
746
+ * lowered (`restoreFrom`). A function with no return keeps the entry-block reading. */
747
+ function prologueSavedRegs(
748
+ entry: readonly Instr[],
749
+ target: TargetDescription,
750
+ r1At: ReadonlyMap<Instr, number | string>,
751
+ blocks: readonly PpcBlock[],
752
+ ): { regs: ReadonlySet<string>; offsets: number[] | null } {
753
+ const code = blocks.flatMap((b) => [...b.body, ...(b.branch ? [b.branch] : [])]);
754
+ const returns = blocks.some((b) => b.synthReturn) || code.some((ins) => isReturn(ins) || isCondReturn(ins));
755
+ const reloaded = (r: string) =>
756
+ code.some((ins) => {
757
+ const routine = saveRoutineOf(target, ins);
758
+ return routine
759
+ ? routine.restore && !routine.fpr && Number(r.slice(1)) >= routine.first
760
+ : ins.ops[1]?.includes('(') === true &&
761
+ parseMem(ins.ops[1]).base === 'r1' &&
762
+ ((ins.mnemonic === 'lwz' && ins.ops[0] === r) ||
763
+ (ins.mnemonic === 'lmw' && Number(r.slice(1)) >= Number(ins.ops[0]?.slice(1))));
764
+ });
765
+ const regs = new Set<string>();
766
+ const named = new Set<string>();
767
+ const offsets: number[] = [];
768
+ let unnamed = false;
769
+ const word = (off: number | null) => {
770
+ if (off === null) {
771
+ unnamed = true;
772
+ } else {
773
+ offsets.push(off);
774
+ }
775
+ };
776
+ let frameAddress: number | null = null;
777
+ for (const ins of entry) {
778
+ const [d, s] = ins.ops;
779
+ const depth = r1At.get(ins);
780
+ const routine = saveRoutineOf(target, ins);
781
+ if (ins.mnemonic === 'addi' && d === target.registerSaveRoutines?.frameReg && s === 'r1') {
782
+ frameAddress = typeof depth === 'number' ? depth + parseImm(ins.ops[2]) : null;
783
+ } else if (routine) {
784
+ if (!routine.restore) {
785
+ const bytes = routine.fpr ? 8 : 4;
786
+ for (let k = routine.first; k <= 31; k++) {
787
+ if (!routine.fpr && !named.has(`r${k}`)) {
788
+ regs.add(`r${k}`);
789
+ }
790
+ word(frameAddress === null ? null : frameAddress - bytes * (32 - k));
791
+ }
792
+ }
793
+ } else if (/^(stw|stmw|stfd)$/.test(ins.mnemonic) && s?.includes('(') && parseMem(s).base === 'r1') {
794
+ const first = ins.mnemonic === 'stmw' ? Number(d.slice(1)) : 0;
795
+ const n = ins.mnemonic === 'stmw' ? 32 - first : 1;
796
+ // A single `stw` no reload answers, in a function that returns, saves nothing (see below).
797
+ const unanswered = ins.mnemonic === 'stw' && isCalleeSaved(target, d) && returns && !reloaded(d);
798
+ for (let k = 0; k < (unanswered ? 0 : n); k++) {
799
+ const reg = ins.mnemonic === 'stmw' ? `r${first + k}` : d;
800
+ if (ins.mnemonic !== 'stfd' && !named.has(reg)) {
801
+ regs.add(reg);
802
+ }
803
+ if (ins.mnemonic === 'stfd' || (isCalleeSaved(target, reg) && !named.has(reg))) {
804
+ word(typeof depth === 'number' ? parseMem(s).off + depth + 4 * k : null);
805
+ }
806
+ }
807
+ }
808
+ for (const op of ins.ops) {
809
+ const reg = op.includes('(') ? parseMem(op).base : op;
810
+ if (isReg(reg)) {
811
+ named.add(reg);
812
+ }
813
+ }
814
+ }
815
+ return { regs, offsets: unnamed ? null : offsets };
816
+ }
409
817
 
410
818
  /** Where r1 stands, relative to its value at entry, just BEFORE an instruction runs: a byte count
411
819
  * (0 before the frame push, -N after `stwu r1,-N(r1)`), or why it is not known there.
@@ -467,16 +875,568 @@ function r1Displacements(blocks: PpcBlock[], succIdx: number[][]): Map<Instr, nu
467
875
  return at;
468
876
  }
469
877
 
878
+ /** What the link-register proof knows at one point: what the link register holds (`lr`: null where
879
+ * it is still the caller's return address on every path, else the instruction that put something
880
+ * else there on some path), which general registers hold a copy of the return address on every
881
+ * path (`regs`), and whether the link register's save word does (`slot`). */
882
+ interface LrState {
883
+ readonly lr: Instr | null;
884
+ readonly regs: ReadonlySet<string>;
885
+ readonly slot: boolean;
886
+ }
887
+
888
+ /** EVERY RETURN OF THIS FUNCTION JUMPS THROUGH THE CALLER'S RETURN ADDRESS, or the lift declines —
889
+ * PowerPC's half of what `refuseReturnsNotThroughCallersLr` asks of Thumb (frontend/thumb.ts).
890
+ *
891
+ * `blr` and a conditional return (`beqlr`) read as returns from the instruction alone, and each is
892
+ * one only where the link register holds the address this function was called with. A forward
893
+ * dataflow over the blocks tracks where that address is:
894
+ * • the link register holds it at entry; a `bl` writes it (a call, and a register save or restore
895
+ * routine too), and an `mtlr rS` puts back whatever rS holds;
896
+ * • `mflr rD` copies it into rD, `mr` copies it on, and any other write to a register loses it —
897
+ * an ordinary call loses every register the EABI lets a callee destroy, a save routine none, and
898
+ * a GPR restore routine the r{first}..r31 it reloads, as `lmw` does;
899
+ * • a `stw` of a register holding it into the link register's save word (`LR_SAVE_WORD`, named
900
+ * from the entry r1 through `r1At`) puts it in that word, and a `lwz` from that word copies it
901
+ * back out. Any other store reaching those bytes, a store through r1 at an unknown depth, a call
902
+ * made before the frame push (the callee saves ITS return address in that very word), and a
903
+ * save routine whose frame address could reach it all overwrite it.
904
+ * Paths meet by agreement: the link register, a register or the word holds the address only if it
905
+ * does on every path in.
906
+ *
907
+ * So `li r3,7; mtlr r4; blr`, `lwz r0,0(r3); mtlr r0; blr`, `mflr r0; addi r0,r0,4; mtlr r0; blr`, an
908
+ * `mtlr` on one path only, and a `bl` with no saved address put back before the `blr` all decline,
909
+ * where mwcc's `mflr r0; stw r0,N+4(r1)` … `lwz r0,N+4(r1); mtlr r0; blr` (2.3.3 saves at `4(r1)`
910
+ * before the push) and a leaf that never names the link register are returns. No compiler emits a
911
+ * violation; hand-written asm can, and it would otherwise lift as a C `return`.
912
+ *
913
+ * Returned per instruction that reads the link register as the return address where it may not be
914
+ * that, with the instruction that put something else there on some path: a return, and an `mflr`
915
+ * after a `bl` (its copy would be taken for the return address, and is the address after the `bl`
916
+ * instead). Absent everywhere else. Blocks no path from the entry reaches are not asked: they never
917
+ * run.
918
+ *
919
+ * NOT ASKED: a store through a register holding a frame address (`addi r3,r1,N`) is not traced to
920
+ * the save word. A compiler addresses that word only through r1. */
921
+ function linkRegisterProof(
922
+ blocks: PpcBlock[],
923
+ succIdx: number[][],
924
+ r1At: ReadonlyMap<Instr, number | string>,
925
+ target: TargetDescription,
926
+ ): Map<Instr, Instr> {
927
+ const at = new Map<Instr, Instr>();
928
+ const destroyed = new Set(['r0', ...target.argRegs, ...target.callerSaved]);
929
+ const frameReg = target.registerSaveRoutines?.frameReg;
930
+ const join = (a: LrState, b: LrState): LrState => ({
931
+ lr: a.lr ?? b.lr,
932
+ regs: new Set([...a.regs].filter((r) => b.regs.has(r))),
933
+ slot: a.slot && b.slot,
934
+ });
935
+ const same = (a: LrState, b: LrState) =>
936
+ a.lr === b.lr && a.slot === b.slot && a.regs.size === b.regs.size && [...a.regs].every((r) => b.regs.has(r));
937
+ // The bytes `[lo, hi)` a store through r1 writes, as entry-r1 offsets, or null where r1's depth is
938
+ // not known there or the address takes an index register.
939
+ const storedRange = (ins: Instr): [number, number] | null => {
940
+ const d = r1At.get(ins);
941
+ const mem = ins.ops[1];
942
+ if (typeof d !== 'number' || !mem?.includes('(')) {
943
+ return null;
944
+ }
945
+ const mn = ins.mnemonic;
946
+ const width =
947
+ mn === 'stmw'
948
+ ? 4 * (32 - Number(ins.ops[0].slice(1)))
949
+ : /^stb/.test(mn)
950
+ ? 1
951
+ : /^sth/.test(mn)
952
+ ? 2
953
+ : /^stfd/.test(mn)
954
+ ? 8
955
+ : 4;
956
+ const lo = parseMem(mem).off + d;
957
+ return [lo, lo + width];
958
+ };
959
+ const namesR1 = (ins: Instr) => ins.ops.some((op) => (op.includes('(') ? parseMem(op).base : op) === 'r1');
960
+
961
+ const run = (bi: number, inState: LrState, record: boolean): LrState => {
962
+ let { lr, slot } = inState;
963
+ const regs = new Set(inState.regs);
964
+ let frameAddress: number | null = null;
965
+ for (const ins of [...blocks[bi].body, ...(blocks[bi].branch ? [blocks[bi].branch] : [])]) {
966
+ const mn = ins.mnemonic;
967
+ const [d, s] = ins.ops;
968
+ if (
969
+ record &&
970
+ ((mn === 'mflr' && lr?.mnemonic === 'bl') || ((isReturn(ins) || isCondReturn(ins)) && lr !== null))
971
+ ) {
972
+ at.set(ins, lr!);
973
+ }
974
+ if (mn === 'bl') {
975
+ const routine = saveRoutineOf(target, ins);
976
+ lr = ins;
977
+ if (routine) {
978
+ // A save routine stores below its frame address; a restore routine writes no memory.
979
+ const bytes = routine.fpr ? 8 : 4;
980
+ if (
981
+ !routine.restore &&
982
+ (frameAddress === null ||
983
+ (frameAddress - bytes * (32 - routine.first) < LR_SAVE_WORD + 4 && frameAddress > LR_SAVE_WORD))
984
+ ) {
985
+ slot = false;
986
+ }
987
+ // A GPR restore routine reloads r{first}..r31 from their save words, as `lmw` does: none of
988
+ // them holds the return address after it, whatever was copied into them before.
989
+ if (routine.restore && !routine.fpr) {
990
+ for (const r of [...regs]) {
991
+ if (Number(r.slice(1)) >= routine.first) {
992
+ regs.delete(r);
993
+ }
994
+ }
995
+ }
996
+ } else {
997
+ for (const r of destroyed) {
998
+ regs.delete(r);
999
+ }
1000
+ if (r1At.get(ins) === 0 || typeof r1At.get(ins) !== 'number') {
1001
+ slot = false;
1002
+ }
1003
+ }
1004
+ continue;
1005
+ }
1006
+ if (mn === 'mtlr') {
1007
+ lr = regs.has(d) ? null : ins;
1008
+ continue;
1009
+ }
1010
+ if (mn === 'mflr') {
1011
+ if (lr === null) {
1012
+ regs.add(d);
1013
+ } else {
1014
+ regs.delete(d);
1015
+ }
1016
+ continue;
1017
+ }
1018
+ if (/^st/.test(mn)) {
1019
+ if (namesR1(ins)) {
1020
+ const range = storedRange(ins);
1021
+ if (mn === 'stw' && range !== null && range[0] === LR_SAVE_WORD && regs.has(d)) {
1022
+ slot = true;
1023
+ } else if (range === null || (range[0] < LR_SAVE_WORD + 4 && range[1] > LR_SAVE_WORD)) {
1024
+ slot = false;
1025
+ }
1026
+ }
1027
+ // An update form writes its base register.
1028
+ if (/^st\w*ux?$/.test(mn)) {
1029
+ regs.delete(s?.includes('(') ? parseMem(s).base : s);
1030
+ }
1031
+ continue;
1032
+ }
1033
+ if (/^(cmp|tw|mt|dc|ic|b|sync|isync|eieio)/.test(mn)) {
1034
+ continue;
1035
+ }
1036
+ if (mn === 'lmw' || /^lsw/.test(mn)) {
1037
+ const first = mn === 'lmw' ? Number(d.slice(1)) : 0;
1038
+ for (const r of [...regs]) {
1039
+ if (Number(r.slice(1)) >= first) {
1040
+ regs.delete(r);
1041
+ }
1042
+ }
1043
+ continue;
1044
+ }
1045
+ const copies =
1046
+ (mn === 'mr' && regs.has(s)) ||
1047
+ (mn === 'or' && s === ins.ops[2] && regs.has(s)) ||
1048
+ (mn === 'lwz' &&
1049
+ slot &&
1050
+ s?.includes('(') &&
1051
+ parseMem(s).base === 'r1' &&
1052
+ typeof r1At.get(ins) === 'number' &&
1053
+ parseMem(s).off + (r1At.get(ins) as number) === LR_SAVE_WORD);
1054
+ if (frameReg !== undefined && mn === 'addi' && d === frameReg && s === 'r1' && !ins.reloc) {
1055
+ const depth = r1At.get(ins);
1056
+ frameAddress = typeof depth === 'number' ? depth + parseImm(ins.ops[2]) : null;
1057
+ } else if (d === frameReg) {
1058
+ frameAddress = null;
1059
+ }
1060
+ if (isReg(d)) {
1061
+ regs.delete(d);
1062
+ }
1063
+ // A load with update writes its base register too.
1064
+ if (/^l\w*ux?$/.test(mn) && mn !== 'lux') {
1065
+ regs.delete(s?.includes('(') ? parseMem(s).base : s);
1066
+ }
1067
+ if (copies && isReg(d)) {
1068
+ regs.add(d);
1069
+ }
1070
+ }
1071
+ return { lr, regs, slot };
1072
+ };
1073
+
1074
+ const into: Array<LrState | undefined> = blocks.map(() => undefined);
1075
+ into[0] = { lr: null, regs: new Set(), slot: false };
1076
+ const work = [0];
1077
+ while (work.length) {
1078
+ const bi = work.pop()!;
1079
+ const out = run(bi, into[bi]!, false);
1080
+ for (const s of succIdx[bi]) {
1081
+ const prev = into[s];
1082
+ const next = prev === undefined ? out : join(prev, out);
1083
+ if (prev === undefined || !same(prev, next)) {
1084
+ into[s] = next;
1085
+ work.push(s);
1086
+ }
1087
+ }
1088
+ }
1089
+ // Ask each reader once the states are settled, so a reason names the fixed point's writer.
1090
+ into.forEach((st, bi) => {
1091
+ if (st !== undefined) {
1092
+ run(bi, st, true);
1093
+ }
1094
+ });
1095
+ return at;
1096
+ }
1097
+
1098
+ /** The sub-word load and store mnemonics this frontend lifts. */
1099
+ const SUBWORD_MEM: ReadonlySet<string> = new Set(['lbz', 'lha', 'lhz', 'stb', 'sth']);
1100
+
1101
+ /** An instruction that takes a frame address as data: `addi rD,r1,k`. The register save routines'
1102
+ * `addi r11,r1,N` is frame bookkeeping (`frame-address` in `nonValues`), not a capture. */
1103
+ const takesFrameAddress = (ins: Instr, frameReg: string | undefined): boolean =>
1104
+ ins.mnemonic === 'addi' && ins.ops[1] === 'r1' && ins.ops[0] !== 'r1' && ins.ops[0] !== frameReg && !ins.reloc;
1105
+
1106
+ /** THE LOADED-WORD BOUND (`compilerBehaviors.outgoingAreaNeverLoadedFrom`, 8 on mwcc), per r1
1107
+ * depth: the lowest frame word, as an offset from the entry r1, that this function LOADS at that
1108
+ * depth from the parameter area up, past the callee's back chain and link-register words. mwcc
1109
+ * loads no word of its outgoing area, before a call or after, so a word it loads is a local's,
1110
+ * and since the area is one region at the frame bottom every word at or above it is no call's
1111
+ * stack argument. Only an instruction the entry reaches has a depth, so every load counted can
1112
+ * run; one at another depth names another frame's words. Measured once, in `liftPrelude`: the
1113
+ * undeclared-call guess (through `measurePpcFrame`'s `outgoingEnd`), the unread-store decline and
1114
+ * the frame-object audit (`loadedWords`) take this number, and it counts what the IR does not
1115
+ * show the audit — the slot reloads and the epilogue's restores. The local area's partition does
1116
+ * not: it starts at the parameter area whatever the bound, and rests on the guess and the audit.
1117
+ * `Infinity` at every depth where the target states no such offset. */
1118
+ function ppcLoadedWordBound(
1119
+ blocks: readonly PpcBlock[],
1120
+ r1At: ReadonlyMap<Instr, number | string>,
1121
+ target: TargetDescription,
1122
+ ): (depth: number) => number {
1123
+ const loadedFrom = target.compilerBehaviors.outgoingAreaNeverLoadedFrom;
1124
+ if (loadedFrom === undefined) {
1125
+ return () => Infinity;
1126
+ }
1127
+ const bounds = new Map<number, number>();
1128
+ for (const b of blocks) {
1129
+ for (const ins of b.body) {
1130
+ const depth = r1At.get(ins);
1131
+ const mem = ins.ops.find((o) => /^-?\d+\(r1\)$/.test(o));
1132
+ if (typeof depth !== 'number' || mem === undefined || !PPC_FRAME_LOAD.test(ins.mnemonic)) {
1133
+ continue;
1134
+ }
1135
+ const off = parseMem(mem).off;
1136
+ const word = off - (((off % 4) + 4) % 4);
1137
+ if (word >= loadedFrom) {
1138
+ bounds.set(depth, Math.min(bounds.get(depth) ?? Infinity, depth + word));
1139
+ }
1140
+ }
1141
+ }
1142
+ return (depth: number) => bounds.get(depth) ?? Infinity;
1143
+ }
1144
+
1145
+ /** THE LOCAL AREA of a function that takes a frame address, reaches its frame below a word, or
1146
+ * keeps a dead local there, and the size of the one frame it pushes — or null, which leaves every
1147
+ * such access the refusal it had without one. Null where none of the three happens; where the
1148
+ * function does not push exactly one frame, with `stwu r1,-N(r1)` at entry depth; where a capture
1149
+ * runs at another depth; where the target states no layout, a save's offset cannot be named, or
1150
+ * no area is left; and where an r1 access at another depth lands on the area's bytes
1151
+ * (`landsOnArea`), which the routing by post-push offset would then misname.
1152
+ *
1153
+ * The area starts where the caller's parameter area does (`PARAM_AREA_AT`), past the back chain
1154
+ * and the link-register save word, and runs to the lowest save (`compilerBehaviors.linkageLocalArea`,
1155
+ * `areaAboveOutgoing`). The measured rule puts it past the widest call's stack words too; no call
1156
+ * this frontend lowers takes any, so a call whose declaration may take some refuses here, where the
1157
+ * area is drawn: one declared with more parameter words than argument registers, with a 64-bit parameter (the EABI
1158
+ * aligns a pair to an odd register, so a packed word count does not say where the stack words
1159
+ * start), or with more floating parameters than f1..f8. So does a declared parameter that is a
1160
+ * struct or union passed by value (`ParamClass` `by-value`), which mwcc passes as the address of a
1161
+ * copy in this area (`tb(1, b)` copies `b` to r1+8 and passes `addi r4,r1,8` at -O4,p) — a copy no
1162
+ * source declares, whose address the lift would pass where the callee takes the struct — and one
1163
+ * nothing sizes (`unsized`), which may be that, a pair or a stack word.
1164
+ *
1165
+ * A CALL NOTHING DECLARES IS PROVEN CLEAR OF THE AREA ONLY BY THE LOADED-WORD BOUND
1166
+ * (`ppcLoadedWordBound`): mwcc loads no word of its outgoing area from the parameter area up, so
1167
+ * the call's stack words, if any, lie below the lowest pushed-frame word this function loads.
1168
+ * `outgoingEnd` is that word, or the parameter area's start where every call is declared: where it
1169
+ * is past the area's start, the area may hold such a word. Its lowering's guess reads it, refusing
1170
+ * a call whose guessed arity reaches seven argument registers unless `outgoingEnd` is at or below
1171
+ * the area's start (`lift`, the guess's `refuse`) — but the guess counts only the registers this function
1172
+ * sets, none it passes straight through, so `g9(a, …, h, 77)` with r3..r10 untouched passes it
1173
+ * and the word at r1+8 is g9's ninth argument, inside the area. What refuses that call is the
1174
+ * shared audit (frame-objects.ts `storedNeverRead`): the store to r1+8 reaches the call unread,
1175
+ * below the bound it takes (`loadedWords`, the same measurement), so no load proves it a local's.
1176
+ * That audit is the backstop this partition rests on for an undeclared callee; a change to the
1177
+ * guess or to the audit's survivor rule moves it.
1178
+ *
1179
+ * A word stored to the frame that no word load reads back, in a function that makes no call, is a
1180
+ * DEAD LOCAL: no callee can take it as a stack argument. A tail branch, which would hand this
1181
+ * frame's caller's parameter area to another function, refuses before this runs (`liftPrelude`).
1182
+ * It lifts as a store to a frame object, and what reaches the C is then the shared lift's, as for
1183
+ * the same store on Thumb and MIPS: a fixed-offset scalar becomes a local, which the dead-store
1184
+ * pass drops (`l3/dce.ts`) unless its value is a read it keeps (`sp8 = *a0`), so `int v; v = x;`
1185
+ * at -O0,p lifts to an empty body and a `volatile` local's stores are lost with it — no code can
1186
+ * observe that word — and only a store at a runtime index stays one, into the frame array. A dead
1187
+ * local alone draws the area only where one lies inside it. */
1188
+ function measurePpcFrame({
1189
+ name,
1190
+ blocks,
1191
+ saves,
1192
+ r1At,
1193
+ target,
1194
+ calls,
1195
+ loadedWordBound,
1196
+ fail,
1197
+ }: {
1198
+ name: string;
1199
+ blocks: readonly PpcBlock[];
1200
+ /** the prologue's save words, as entry-r1 offsets (`prologueSavedRegs`) */
1201
+ saves: readonly number[] | null;
1202
+ r1At: ReadonlyMap<Instr, number | string>;
1203
+ target: TargetDescription;
1204
+ calls: CallDeclarations;
1205
+ /** `ppcLoadedWordBound`, the one the lift reads */
1206
+ loadedWordBound: (depth: number) => number;
1207
+ fail: FrontendRefusal;
1208
+ }): { area: FrameRange; size: number; outgoingEnd: number } | null {
1209
+ const code = blocks.flatMap((b) => [...b.body, ...(b.branch ? [b.branch] : [])]);
1210
+ const bls = blocks.flatMap((b) => b.body).filter((ins) => ins.mnemonic === 'bl' && !saveRoutineOf(target, ins));
1211
+ const frameReg = target.registerSaveRoutines?.frameReg;
1212
+ const r1Mem = (ins: Instr) => {
1213
+ const last = ins.ops[ins.ops.length - 1] ?? '';
1214
+ return last.includes('(') && parseMem(last).base === 'r1' && ins.mnemonic !== 'stwu' ? parseMem(last) : undefined;
1215
+ };
1216
+ const captures = code.filter((ins) => takesFrameAddress(ins, frameReg));
1217
+ const subword = code.filter((ins) => SUBWORD_MEM.has(ins.mnemonic) && r1Mem(ins) !== undefined);
1218
+ const entryOff = (ins: Instr) => {
1219
+ const at = r1At.get(ins);
1220
+ return typeof at === 'number' ? r1Mem(ins)!.off + at : undefined;
1221
+ };
1222
+ const loadedWords = new Set(
1223
+ code.flatMap((ins) => {
1224
+ const at = r1Mem(ins) && /^(lwz|lmw)$/.test(ins.mnemonic) ? entryOff(ins) : undefined;
1225
+ return at === undefined
1226
+ ? []
1227
+ : ins.mnemonic === 'lmw'
1228
+ ? [...Array(32 - Number(ins.ops[0].slice(1))).keys()].map((k) => at + 4 * k)
1229
+ : [at];
1230
+ }),
1231
+ );
1232
+ const deadStores =
1233
+ bls.length > 0
1234
+ ? []
1235
+ : code.filter(
1236
+ (ins) => ins.mnemonic === 'stw' && r1Mem(ins) !== undefined && !loadedWords.has(entryOff(ins) ?? NaN),
1237
+ );
1238
+ if (captures.length === 0 && subword.length === 0 && deadStores.length === 0) {
1239
+ return null;
1240
+ }
1241
+ const pushes = code.filter((ins) => ins.mnemonic === 'stwu' && ins.ops[0] === 'r1');
1242
+ if (pushes.length !== 1 || r1At.get(pushes[0]) !== 0 || parseMem(pushes[0].ops[1]).base !== 'r1') {
1243
+ return null;
1244
+ }
1245
+ const size = -parseMem(pushes[0].ops[1]).off;
1246
+ if (!(size > 0) || captures.some((ins) => r1At.get(ins) !== -size) || saves === null) {
1247
+ return null;
1248
+ }
1249
+ const area = target.compilerBehaviors.linkageLocalArea
1250
+ ? areaAboveOutgoing({ frameSize: size, outgoingEnd: PARAM_AREA_AT, saves: saves.map((o) => o + size), align: 4 })
1251
+ : null;
1252
+ if (area === null) {
1253
+ return null;
1254
+ }
1255
+ for (const ins of code) {
1256
+ const off = r1Mem(ins) === undefined || r1At.get(ins) === -size ? undefined : entryOff(ins);
1257
+ if (off !== undefined && landsOnArea(area, off + size, 4)) {
1258
+ return null;
1259
+ }
1260
+ }
1261
+ // Drawn for dead stores alone, the area must hold one: a save the prologue made is no local.
1262
+ const inArea = (ins: Instr) => {
1263
+ const at = (entryOff(ins) ?? NaN) + size;
1264
+ return at >= area.from && at < area.to;
1265
+ };
1266
+ if (captures.length === 0 && subword.length === 0 && !deadStores.some(inArea)) {
1267
+ return null;
1268
+ }
1269
+ const fprArgs = target.fpu?.argRegs.length ?? 0;
1270
+ for (const ins of bls) {
1271
+ const callee = ins.reloc?.sym ?? '';
1272
+ const params = calls.declaredParams(callee);
1273
+ if (params === null) {
1274
+ continue;
1275
+ }
1276
+ const at = `'${callee}', called at 0x${ins.addr.toString(16)},`;
1277
+ const copied = params.findIndex((p) => p.kind === 'by-value');
1278
+ if (copied >= 0) {
1279
+ fail(
1280
+ `cannot lift '${name}': address-taken stack local — ${at} takes its parameter ${copied + 1} as ` +
1281
+ `'${params[copied].spelling}', which no word holds: a struct or union passed by value, which this compiler ` +
1282
+ "passes as the address of a copy in the caller's local area — not modelled",
1283
+ );
1284
+ }
1285
+ const count = (kind: ParamClass) => params.filter((p) => p.kind === kind).length;
1286
+ const unsized = params.findIndex((p) => p.kind === 'unsized');
1287
+ const why =
1288
+ count('wide') > 0
1289
+ ? 'a 64-bit parameter, which the EABI places in an aligned register pair or on the stack'
1290
+ : count('word') > target.argRegs.length
1291
+ ? `${count('word')} parameter words and the argument registers carry ${target.argRegs.length}`
1292
+ : count('float') > fprArgs
1293
+ ? `${count('float')} floating parameters and the floating-point argument registers carry ${fprArgs}`
1294
+ : unsized >= 0
1295
+ ? `its parameter ${unsized + 1} as '${params[unsized].spelling}', which nothing here sizes: a word, a ` +
1296
+ 'pair, a stack word or a by-value copy in the local area'
1297
+ : undefined;
1298
+ if (why !== undefined) {
1299
+ fail(
1300
+ `cannot lift '${name}': outgoing stack arguments not modelled — ${at} is declared with ${why}, and this ` +
1301
+ "compiler's local area starts past the stack words a call takes",
1302
+ );
1303
+ }
1304
+ }
1305
+ // Every call a declaration sizes takes no stack word, or refused above. One nothing declares
1306
+ // takes none past the loaded-word bound of the pushed frame, the most its area can reach.
1307
+ const undeclared = bls.some((ins) => calls.declaredParams(ins.reloc?.sym ?? '') === null);
1308
+ const outgoingEnd = undeclared ? Math.max(PARAM_AREA_AT, loadedWordBound(-size) + size) : PARAM_AREA_AT;
1309
+ return { area, size, outgoingEnd };
1310
+ }
1311
+
1312
+ /** What a call's lowering on this frontend carries out (`CallLowering`): no pair, no stack word, no
1313
+ * struct returned through memory. */
1314
+ const PPC_CALL_LOWERING = {
1315
+ pairs: false,
1316
+ memoryReturn: false,
1317
+ stackArgs: false,
1318
+ voidReturn: false,
1319
+ argRegisterBoundsArity: false,
1320
+ } as const;
1321
+
1322
+ /** This frontend's refusal, for the shared code that refuses on its behalf. */
1323
+ const ppcRefusal: FrontendRefusal = (message) => {
1324
+ throw new PpcUnsupportedError(message);
1325
+ };
1326
+
1327
+ /** WHAT `lift` READS OFF THE TEXT BEFORE IT DECODES AN INSTRUCTION: the control transfers it
1328
+ * refuses, the blocks (with the preheader), r1's depth at each instruction, the prologue's saves,
1329
+ * the callees' declarations, and the local area those draw (`measurePpcFrame`). `lift` and
1330
+ * `__testing.frameOf` both run it, so the area a test pins is the one the lift draws. */
1331
+ function liftPrelude(
1332
+ name: string,
1333
+ instrs: Instr[],
1334
+ target: TargetDescription,
1335
+ prototypes: Prototypes,
1336
+ jts: Map<number, PpcJT>,
1337
+ ownDeclaration?: () => OwnDeclaration | undefined,
1338
+ ) {
1339
+ const recoveredBctr = new Set([...jts.values()].map((j) => j.bctrAddr));
1340
+ // TRUSTWORTHINESS: fail loud on an unmodelled control transfer rather than dropping it (which
1341
+ // would silently miscompile the control flow). CTR-counted loops and indirect branches land here.
1342
+ for (const ins of instrs) {
1343
+ if (ins.mnemonic.startsWith('b') && !isModeledBranch(ins)) {
1344
+ if (ins.mnemonic === 'bctr' && recoveredBctr.has(ins.addr)) {
1345
+ continue;
1346
+ } // recovered switch dispatch
1347
+ throw new PpcUnsupportedError(
1348
+ `cannot lift '${name}': unmodelled control transfer '${ins.mnemonic}' at 0x${ins.addr.toString(16)} ` +
1349
+ (ins.mnemonic === 'blrl' || ins.mnemonic === 'bctrl'
1350
+ ? `(an indirect call — a virtual dispatch or a call through a pointer — is not yet supported)`
1351
+ : `(CTR-counted loop or indirect branch — mwcc -O4 loop unrolling is not yet supported)`),
1352
+ );
1353
+ }
1354
+ }
1355
+ // A BRANCH OUT OF THE FUNCTION is a tail call: the callee takes over this frame's caller's
1356
+ // parameter area as its own, so a word stored there may be its stack argument, and the call
1357
+ // itself would be dropped. objdump prints one either at an address outside the function or,
1358
+ // relocated, at its own address, which would read as a loop. A conditional branch out
1359
+ // (`beq- g`, `R_PPC_REL14`) and a `bdnz` out are tail calls too.
1360
+ const tail = instrs.find(
1361
+ (i) =>
1362
+ (isUncond(i) || isCond(i) || isCtrLoop(i)) &&
1363
+ (i.reloc !== undefined || (i.target !== undefined && !instrs.some((j) => j.addr === i.target))),
1364
+ );
1365
+ if (tail !== undefined) {
1366
+ throw new PpcUnsupportedError(
1367
+ `cannot lift '${name}': the '${tail.mnemonic}' at 0x${tail.addr.toString(16)} leaves the function` +
1368
+ (tail.reloc ? ` for '${tail.reloc.sym}'` : '') +
1369
+ ' — a tail call is not modelled',
1370
+ );
1371
+ }
1372
+ const { blocks, succIdx } = toBlocks(instrs, name, jts);
1373
+ refuseReachableDispatch(name, blocks, recoveredBctr, ppcRefusal);
1374
+ const entryBody = blocks[0].body;
1375
+ // PREHEADER. A function whose first block is a branch target (`do b = b * b; while (--n);` is a
1376
+ // loop at the entry) would take its parameters as that loop header's phis. An empty block ahead
1377
+ // of it becomes the entry, and its live-ins are the function's parameters (frontend/ssa.ts
1378
+ // `ArgSlots`). It starts where the function does, which is what its label says.
1379
+ if (succIdx.some((ss) => ss.includes(0))) {
1380
+ blocks.unshift({ startAddr: blocks[0].startAddr, body: [], branch: null });
1381
+ succIdx.forEach((ss, i) => (succIdx[i] = ss.map((j) => j + 1)));
1382
+ succIdx.unshift([1]);
1383
+ }
1384
+ const r1At = r1Displacements(blocks, succIdx);
1385
+ const saves = prologueSavedRegs(entryBody, target, r1At, blocks);
1386
+ const calls = callDeclarations(name, target, prototypes, PPC_CALL_LOWERING, ppcRefusal, ownDeclaration);
1387
+ // THE LOCAL AREA, where a function takes a frame address, reaches its frame below a word, or
1388
+ // keeps a dead local there (`measurePpcFrame`, `compilerBehaviors.linkageLocalArea`): every r1
1389
+ // access inside it, any width, is MEMORY through an `laddr` the shared frame-object audit then
1390
+ // proves, as on Thumb and MIPS, and every word access outside it keeps the slot model. Null where
1391
+ // neither happens, or the frame is one this cannot partition, and then every capture keeps its
1392
+ // r1-as-data refusal and every sub-word access its own.
1393
+ const loadedWordBound = ppcLoadedWordBound(blocks, r1At, target);
1394
+ const frame = measurePpcFrame({
1395
+ name,
1396
+ blocks,
1397
+ saves: saves.offsets,
1398
+ r1At,
1399
+ target,
1400
+ calls,
1401
+ loadedWordBound,
1402
+ fail: ppcRefusal,
1403
+ });
1404
+ return { blocks, succIdx, r1At, savedRegs: saves.regs, calls, frame, loadedWordBound };
1405
+ }
1406
+
470
1407
  /** Lift disassembled PowerPC text → an L1 Fn with block-argument SSA. `asmData` (optional) supplies
471
1408
  * the data-section jump table for dense-switch (Regime-B) recovery; absent ⇒ a `bctr` dispatch
472
- * loud-fails. */
1409
+ * loud-fails.
1410
+ *
1411
+ * AT MOST TWICE, as on Thumb and MIPS (`frontend/frame-objects.ts` `liftWithRelift`): where the
1412
+ * frame-object audit answers that a callee may write the local area without bound, or that nothing
1413
+ * escapes and the per-object model does not describe it, it asks for the area kept as one object,
1414
+ * and the second lift's audit judges it so and refuses rather than asking again. */
473
1415
  export function lift(
474
1416
  name: string,
475
1417
  asm: string,
476
1418
  target: TargetDescription,
477
1419
  prototypes: Prototypes = {},
478
1420
  asmData?: AsmData,
1421
+ symbols?: SymbolMap,
1422
+ context?: ContextGlobals,
1423
+ options: LiftOptions = {},
479
1424
  ): Fn {
1425
+ return liftWithRelift(name, (oneObject) =>
1426
+ liftOnce(name, asm, target, prototypes, asmData, { symbols, context }, oneObject, options),
1427
+ );
1428
+ }
1429
+
1430
+ function liftOnce(
1431
+ name: string,
1432
+ asm: string,
1433
+ target: TargetDescription,
1434
+ prototypes: Prototypes,
1435
+ asmData: AsmData | undefined,
1436
+ declarations: Declarations,
1437
+ oneObject: FrameRange | undefined,
1438
+ { contextOwnWords, contextOwnProto }: Pick<LiftOptions, 'contextOwnWords' | 'contextOwnProto'>,
1439
+ ): Fn | FrameObjectRelift {
480
1440
  assertInputFormat('ppc', 'objdump', asm);
481
1441
  const instrs = parseDisasm(sliceSymbol(asm, name)); // ONE function only — an absent symbol declines loud
482
1442
  if (instrs.length === 0) {
@@ -486,19 +1446,27 @@ export function lift(
486
1446
  // recovered dispatch's `bctr` is subsumed into a `switch_br` (emitted from its bounds block), so
487
1447
  // it is exempted from the loud-fail below; an UNrecovered `bctr` still fails loud.
488
1448
  const jts = asmData ? recoverPpcJumpTables(instrs, asmData) : new Map<number, PpcJT>();
489
- const recoveredBctr = new Set([...jts.values()].map((j) => j.bctrAddr));
490
1449
  // A data reloc on an instruction whose decode did not TAKE it means objdump printed a LINK-TIME
491
1450
  // placeholder (`lis r4,0` + R_PPC_ADDR16_HA sym): the real value is the symbol's half, and
492
1451
  // lifting the 0 silently reads the wrong address — plausible-but-wrong C, the forbidden class.
493
1452
  // The jump-table idiom's own @tbl pair never reaches this guard: a recovered dispatch block
494
- // is the bounds branch's replaced fall-through, pruned as unreachable before decode.
1453
+ // is the bounds branch's replaced fall-through, pruned as unreachable before decode, and a
1454
+ // hoisted pair is decoded into a table base (`tableBases`).
495
1455
  //
496
1456
  // TWO refusals, because they are different problems and the reader acts on them differently. A
497
1457
  // symbol the naming policy calls unspellable (frontend/reloc-symbol.ts) is a DEAD END: recovering
498
1458
  // the address perfectly would still leave a name no C source can write, so no amount of lifting
499
1459
  // opens the row. A spellable symbol is a CAPABILITY GAP: the address is recoverable and the row
500
1460
  // waits on the fold. Saying which one a row hit is the whole value of the message.
501
- const relocSite = (ins: Instr) => `cannot lift '${name}': '${ins.mnemonic}' at 0x${ins.addr.toString(16)}`;
1461
+ //
1462
+ // An update form is decoded as the `addi` that advances its base and then the plain access
1463
+ // (`withUpdate` in `decode`); a refusal of that `addi`'s relocation, or of either one's write
1464
+ // (`write`), names the instruction the listing holds, not the one it was decoded through.
1465
+ const decodedFrom = new WeakMap<Instr, Instr>();
1466
+ const relocSite = (at: Instr) => {
1467
+ const ins = decodedFrom.get(at) ?? at;
1468
+ return `cannot lift '${name}': '${ins.mnemonic}' at 0x${ins.addr.toString(16)}`;
1469
+ };
502
1470
  /** Whether the decode of the instruction in hand has TAKEN its relocation. `decode` asks once, of
503
1471
  * every modelled instruction, so a relocation no case consumed refuses instead of being dropped.
504
1472
  * A per-case guard would only cover the cases that remember to carry one, and every other
@@ -538,33 +1506,18 @@ export function lift(
538
1506
  `is a link-time placeholder, not the value`,
539
1507
  );
540
1508
  };
541
- // TRUSTWORTHINESS: fail loud on an unmodelled control transfer rather than dropping it (which
542
- // would silently miscompile the control flow). CTR-counted loops and indirect branches land here.
543
- for (const ins of instrs) {
544
- if (ins.mnemonic.startsWith('b') && !isModeledBranch(ins)) {
545
- if (ins.mnemonic === 'bctr' && recoveredBctr.has(ins.addr)) {
546
- continue;
547
- } // recovered switch dispatch
548
- throw new PpcUnsupportedError(
549
- `cannot lift '${name}': unmodelled control transfer '${ins.mnemonic}' at 0x${ins.addr.toString(16)} ` +
550
- (ins.mnemonic === 'blrl' || ins.mnemonic === 'bctrl'
551
- ? `(an indirect call — a virtual dispatch or a call through a pointer — is not yet supported)`
552
- : `(CTR-counted loop or indirect branch — mwcc -O4 loop unrolling is not yet supported)`),
553
- );
554
- }
555
- }
556
- const { blocks, succIdx } = toBlocks(instrs, name, jts);
557
- // PREHEADER. A function whose first block is a branch target (`do b = b * b; while (--n);` is a
558
- // loop at the entry) would take its parameters as that loop header's phis. An empty block ahead
559
- // of it becomes the entry, and its live-ins are the function's parameters (frontend/ssa.ts
560
- // `ArgSlots`). It starts where the function does, which is what its label says.
561
- if (succIdx.some((ss) => ss.includes(0))) {
562
- blocks.unshift({ startAddr: blocks[0].startAddr, body: [], branch: null });
563
- succIdx.forEach((ss, i) => (succIdx[i] = ss.map((j) => j + 1)));
564
- succIdx.unshift([1]);
565
- }
566
-
567
- const r1At = r1Displacements(blocks, succIdx);
1509
+ const fail = ppcRefusal;
1510
+ // the own declaration binds only a body a compiler wrote (`ownDeclared`, measured below, before
1511
+ // any call is lowered)
1512
+ const { blocks, succIdx, r1At, savedRegs, calls, frame, loadedWordBound } = liftPrelude(
1513
+ name,
1514
+ instrs,
1515
+ target,
1516
+ prototypes,
1517
+ jts,
1518
+ () => (!ownDeclared ? undefined : ownDeclarationOf(name, prototypes, contextOwnProto)),
1519
+ );
1520
+ const lrUnproven = linkRegisterProof(blocks, succIdx, r1At, target);
568
1521
 
569
1522
  const preds: number[][] = blocks.map(() => []);
570
1523
  blocks.forEach((_, i) => {
@@ -580,25 +1533,81 @@ export function lift(
580
1533
  // handed to one in f1 — would resolve to a value the callee overwrote or be dropped. Both scans
581
1534
  // read the blocks the entry reaches: a `bl` on no path makes no call.
582
1535
  const reached = blocks.flatMap((b) => b.body);
583
- const callInsn = reached.find((ins) => ins.mnemonic === 'bl');
584
- const floatReturn = writesFloatReturn(reached, FP_ARITH_MNEMONICS, (t) => (isFpKey(t) ? t : null), fpu);
585
- const fpType = fpPrecision(reached, (m) => FP_ARITH[m]?.width);
1536
+ // A register save routine is no call: it moves only the registers it saves (`saveRoutine`).
1537
+ const callInsn = reached.find((ins) => ins.mnemonic === 'bl' && !saveRoutineOf(target, ins));
586
1538
 
587
- const ssa = makeSsaBuilder(name, blocks.length, preds, undefined, (k) =>
588
- isFpKey(k) ? (fpType ?? T.f32()) : undefined,
1539
+ // THE LIVE-IN MODEL is the register half only (this frontend claims no frame partition): a
1540
+ // def-less read of a register the ABI passes no argument in, and that the prologue saved or the
1541
+ // ABI lets a callee use unsaved, is an uninitialised local. `registerLiveIns` (frontend/ssa.ts)
1542
+ // composes it, as it does for Thumb. The FPU's callee-saved half is not measured, so f14..f31 keep
1543
+ // the treatment a target claiming no partition gets. The function's own declaration bounds the
1544
+ // general argument registers (`ownArgumentWordsOn`): mwcc 2.3.3 homes locals in r5..r10 past a
1545
+ // two-parameter one. The FPRs are not bounded here.
1546
+ //
1547
+ // NOT WHERE THE BODY IS HAND-WRITTEN. A declaration states a compiled function's register
1548
+ // interface, and only a compiled one's: a CodeWarrior `asm` function's type — `ksNesInit03__Fv`,
1549
+ // declared `void ()` — says nothing of the registers it really takes (r4..r8 there, set by the
1550
+ // emulator core that calls it), and bounding by it turns each into `uninit_<reg>`. The cheap
1551
+ // tell is a callee-saved register the body names without the prologue having saved it
1552
+ // (`stw r7,5612(r31)` with no save of r31), which compiled code never does: it saves before it
1553
+ // touches one. Such a body keeps every argument register a parameter, as with no declaration.
1554
+ const ownDeclared = !namesUnsavedCalleeSaved(reached, target, savedRegs);
1555
+ // Which home the function returns in (frontend/fpu.ts `returnHomeRule`), by the own declaration
1556
+ // where the body is a compiler's.
1557
+ const homeRule: ReturnHomeRule = !fpu
1558
+ ? { home: 'integer', declared: false }
1559
+ : returnHomeRule(
1560
+ ownDeclared ? ownDeclarationOf(name, prototypes, contextOwnProto) : undefined,
1561
+ reached.some((ins) => FP_CROSSINGS.has(ins.mnemonic)),
1562
+ () => writesFloatReturn(reached, FP_ARITH_MNEMONICS, (t) => (isFpKey(t) ? t : null), fpu),
1563
+ );
1564
+ const floatReturn = homeRule.home === 'float';
1565
+ const fpType = declaredPrecision(
1566
+ fpPrecision(reached, (m) => FP_ARITH[m]?.width ?? (FP_CROSSINGS.has(m) ? null : undefined)),
1567
+ ownDeclared ? ownDeclarationOf(name, prototypes, contextOwnProto) : undefined,
589
1568
  );
590
- const { irBlocks, readVar, writeVar, paramReg } = ssa;
591
- /** This frontend's refusal, for the shared code that refuses on its behalf. */
592
- const fail: FrontendRefusal = (message) => {
593
- throw new PpcUnsupportedError(message);
594
- };
595
- const calls = callDeclarations(
596
- name,
1569
+ const liveIns = registerLiveIns(
597
1570
  target,
598
- prototypes,
599
- { pairs: false, memoryReturn: false, stackArgs: false, voidReturn: false, argRegisterBoundsArity: false },
600
- fail,
1571
+ savedRegs,
1572
+ !ownDeclared ? undefined : (contextOwnWords ?? ownArgumentWordsOn(name, prototypes[name], target)),
601
1573
  );
1574
+ const ssa = makeSsaBuilder(
1575
+ name,
1576
+ blocks.length,
1577
+ preds,
1578
+ () => ({ ...liveIns, saveSite: 'store or save routine in its entry block' }),
1579
+ (k) => (isFpKey(k) ? (fpType ?? T.f32()) : undefined),
1580
+ liveIns.globalRegs,
1581
+ );
1582
+ const { irBlocks, readVar, writeVar, paramReg } = ssa;
1583
+
1584
+ /** Where an r1 access of `width` bytes at `ins` lies in the local area: its post-push offset, or
1585
+ * undefined for one outside it. An access across the area's edge refuses: half memory and half
1586
+ * slot is neither. `measurePpcFrame` refuses a frame where an access at another depth lands in
1587
+ * the area, so one inside it runs at the frame's own depth. */
1588
+ const inLocalArea = (ins: Instr, mem: string, width: number): number | undefined => {
1589
+ const depth = r1At.get(ins);
1590
+ if (frame === null || parseMem(mem).base !== 'r1' || typeof depth !== 'number') {
1591
+ return undefined;
1592
+ }
1593
+ const at = parseMem(mem).off + depth + frame.size;
1594
+ if (at >= frame.area.to || at + width <= frame.area.from) {
1595
+ return undefined;
1596
+ }
1597
+ if (at < frame.area.from || at + width > frame.area.to) {
1598
+ fail(
1599
+ `cannot lift '${name}': address-taken stack local — '${ins.mnemonic}' at 0x${ins.addr.toString(16)} reaches ` +
1600
+ `across an edge of the local area, r1+${frame.area.from} up to r1+${frame.area.to} — part memory, part slot`,
1601
+ );
1602
+ }
1603
+ return at;
1604
+ };
1605
+ // The frame offset each `laddr` this lift emitted names — function-scoped, since a register
1606
+ // holding one crosses blocks — and every capture moved into another (`addi rD,rCapture,k`
1607
+ // re-minted as the `laddr` of the sum), handed to the audit so a base only the move read is
1608
+ // dropped rather than judged as a capture (`foldMovedCaptures`).
1609
+ const laddrAt = new Map<Value, number>();
1610
+ const movedCaptures = new Set<Value>();
602
1611
 
603
1612
  /** The HIGH half of a relocated address, per value standing for one. The invariant, and every
604
1613
  * way a placeholder could leak past it, lives in frontend/high-half.ts — shared with
@@ -619,17 +1628,175 @@ export function lift(
619
1628
  lo: '@l',
620
1629
  fail,
621
1630
  });
1631
+ /** The jump-table bases a recovered dispatch reads from outside its own block — the `@l` that
1632
+ * completes one defines no value but the base (frontend/jump-table.ts). */
1633
+ const tableBases = makeTableBases(
1634
+ [...jts.values()].flatMap((j) => (j.hoisted ? [j.hoisted] : [])),
1635
+ fail,
1636
+ );
1637
+ // What an elided dispatch left at each case target's entry (jump-table.ts `caseEntries`), and the
1638
+ // markers written there, by value.
1639
+ const entries = caseEntries(
1640
+ blocks.flatMap((b, at) => {
1641
+ const jt = b.branch ? jts.get(b.branch.addr) : undefined;
1642
+ if (jt === undefined) {
1643
+ return [];
1644
+ }
1645
+ const n = jt.caseAddrs.length;
1646
+ // the default is a `bgtlr` bound's synthetic return of its own, or the default block
1647
+ return [
1648
+ { at, scrutReg: jt.scrutReg, cases: succIdx[at].slice(0, n), defaultBlock: succIdx[at][n], leaves: jt.leaves },
1649
+ ];
1650
+ }),
1651
+ preds,
1652
+ );
1653
+ const staleIdiom = new Map<Value, StaleIdiomRead>();
1654
+ const parked = parkedIn(staleIdiom);
1655
+ const scaledIndex = scaledIndexIn(staleIdiom);
1656
+ /** What this frontend defines registers with that is no argument: a pending `@ha` half and a
1657
+ * table base. NOT a marker an elided dispatch left (`staleIdiom`): a case may copy one
1658
+ * (`mr r3,r4`), and a return or a call that skipped it would drop what the case handed on. */
1659
+ const placeholders = {
1660
+ has: (v: Value) => highHalves.has(v) || tableBases.has(v),
1661
+ guardRead: (n: string, r: string, v: Value) => tableBases.guardRead(n, r, highHalves.guardRead(n, r, v)),
1662
+ };
622
1663
  /** Reading a register AS A VALUE. A register holding a high half is not one, so this refuses loud
623
1664
  * rather than handing back a plausible number standing for an address. `foldLoHalf` is the only
624
1665
  * code entitled to look at a high half, which is why it reaches `readVar` directly. */
625
- const readReg = (r: string, at: number): Value => highHalves.guardRead(name, r, readVar(r, at));
626
- /** The return address `mflr rD` copies out of the link register, per value standing for one, with
627
- * the address of the `mflr`. Like a high half it is a definition and not a value: storing it to
628
- * the link register's save word is the LR save (`frameStore`), and moving it back with `mtlr` is
629
- * the return. Any other use is a function reading its own return address (MP4's
630
- * `asm { mflr retaddr }` does, to tag an allocation with its caller), which no C expression
631
- * spells, so that use refuses: a store at once, anything else once the function is built. */
632
- const returnAddresses = new Map<Value, number>();
1666
+ const readReg = (r: string, at: number): Value => {
1667
+ if (SMALL_DATA_BASES.includes(r)) {
1668
+ fail(
1669
+ `cannot lift '${name}': ${r} (the small-data base) used as data — an access the relocation ` +
1670
+ `does not name is not modelled`,
1671
+ );
1672
+ }
1673
+ const v = placeholders.guardRead(name, r, readVar(r, at));
1674
+ const nv = nonValues.get(v);
1675
+ if (nv !== undefined && NON_VALUE_RULES[nv.kind].refuseAtRead) {
1676
+ refuseNonValue(nv);
1677
+ }
1678
+ return v;
1679
+ };
1680
+ /** Reading r1 (the stack pointer) as a DATA operand means frame-pointer arithmetic or an
1681
+ * address-taken local (`addi r3,r1,8` = `&local`) — not modellable without a stack abstraction,
1682
+ * and fabricating a value for r1 silently miscompiles. */
1683
+ const r1AsData = (): never =>
1684
+ fail(
1685
+ `cannot lift '${name}': stack pointer r1 used as data (address-taken local / frame arithmetic) — not supported`,
1686
+ );
1687
+ /** An `mflr` reached by a `bl` with no `mtlr` since, or a return the link-register proof could
1688
+ * not show jumps through the caller's return address (`linkRegisterProof`). */
1689
+ const refuseOverwrittenLink = (ins: Instr) => {
1690
+ const by = lrUnproven.get(ins);
1691
+ const at = (i: Instr) => `'${i.mnemonic}' at 0x${i.addr.toString(16)}`;
1692
+ if (by === undefined) {
1693
+ return;
1694
+ }
1695
+ if (ins.mnemonic === 'mflr') {
1696
+ fail(
1697
+ `cannot lift '${name}': the ${at(ins)} reads the link register the ${at(by)} overwrote on some path, ` +
1698
+ `with no 'mtlr' since — the return address is lost there`,
1699
+ );
1700
+ }
1701
+ if (by.mnemonic === 'bl') {
1702
+ fail(
1703
+ `cannot lift '${name}': the ${at(ins)} may not return through the caller's link register — the ` +
1704
+ `${at(by)} overwrote it on some path, and no 'mtlr' put the saved return address back`,
1705
+ );
1706
+ }
1707
+ fail(
1708
+ `cannot lift '${name}': the ${at(ins)} may not return through the caller's link register — on some ` +
1709
+ `path the ${at(by)} put ${by.ops[0]} there, which does not hold the return address an 'mflr' copied ` +
1710
+ `out or its reload from the link register's save word`,
1711
+ );
1712
+ };
1713
+ /** DEFINITIONS THAT ARE NOT VALUES: the frame bookkeeping this frontend erases. Each kind is
1714
+ * written into a register as an ordinary SSA definition, so SSA answers whether one reaches a use,
1715
+ * and is keyed by that value, as the high halves are (frontend/high-half.ts). Its one legitimate
1716
+ * consumer looks it up here; every other use refuses through `refuseNonValue`,
1717
+ * checked at a read (`readReg`) where the kind says so and, always, on the finished function
1718
+ * (`assertNoNonValueEscaped`), because a use can reach through a join's incoming arguments, which
1719
+ * no read sees. A new kind is one more member of the union, one rule and one refusal, and the
1720
+ * three checks cover it. The high halves keep their own module only because MIPS shares it.
1721
+ *
1722
+ * - `frame-address`: what `addi r11,r1,N` forms — its offset from the entry r1 (or why that is not
1723
+ * known there) and the address of the `addi`. Its one consumer is a register save or restore
1724
+ * routine (`saveRoutineOf`), which takes it in the target's `registerSaveRoutines.frameReg`, so
1725
+ * that register is the only destination deferred; every other `addi rD,r1,N` refuses where it
1726
+ * stands. Any other use is the frame read as data, and refuses as the `addi` itself would. One
1727
+ * no routine took refuses too, even where nothing reads it (`assertNonValuesConsumed`). A `bl`
1728
+ * to the routine that carries no relocation (a linked image, or objdump without `-r`) never gets
1729
+ * that far: it is no routine, and `case 'bl'` refuses it as a call with no callee symbol.
1730
+ * - `restored`: the caller's value a restore puts back in a callee-saved register
1731
+ * (`restoreFrom`). No C value names it, so it has no consumer.
1732
+ * - `return-address`: what `mflr rD` copies out of the link register, with the address of the
1733
+ * `mflr`. Storing it to the link register's save word is the LR save (`frameStore`), and moving
1734
+ * it back with `mtlr` is the return. Any other use is a function reading its own return address
1735
+ * (MP4's `asm { mflr retaddr }` does, to tag an allocation with its caller), which no C
1736
+ * expression spells: a store elsewhere refuses at once, anything else once the function is
1737
+ * built. A read alone does not refuse, so a copy (`mr`) still carries it to its save.
1738
+ * - `fctiwz-high`: the HIGH word an `stfd` of an `fctiwz` result stores, with the address of the
1739
+ * `stfd`. `fctiwz` leaves that half undefined, so a reload of it is no value. No consumer. */
1740
+ type NonValue =
1741
+ | { kind: 'frame-address'; off: number | string; addr: number }
1742
+ | { kind: 'restored'; reg: string; addr: number }
1743
+ | { kind: 'return-address'; addr: number }
1744
+ | { kind: 'fctiwz-high'; addr: number };
1745
+ const NON_VALUE_RULES: Record<NonValue['kind'], { refuseAtRead: boolean; mustBeConsumed: boolean }> = {
1746
+ 'frame-address': { refuseAtRead: true, mustBeConsumed: true },
1747
+ restored: { refuseAtRead: true, mustBeConsumed: false },
1748
+ 'return-address': { refuseAtRead: false, mustBeConsumed: false },
1749
+ 'fctiwz-high': { refuseAtRead: true, mustBeConsumed: false },
1750
+ };
1751
+ const nonValues = new Map<Value, NonValue>();
1752
+ /** The non-values their consumer took (only `frame-address` has one that must). */
1753
+ const consumedNonValues = new Set<Value>();
1754
+ /** The non-value of kind `k` that `v` stands for, or undefined. */
1755
+ const nonValueOf = <K extends NonValue['kind']>(v: Value, k: K): Extract<NonValue, { kind: K }> | undefined => {
1756
+ const nv = nonValues.get(v);
1757
+ return nv?.kind === k ? (nv as Extract<NonValue, { kind: K }>) : undefined;
1758
+ };
1759
+ const refuseNonValue = (nv: NonValue): never => {
1760
+ switch (nv.kind) {
1761
+ case 'frame-address':
1762
+ return r1AsData();
1763
+ case 'restored':
1764
+ return fail(
1765
+ `cannot lift '${name}': ${nv.reg} is read after the restore at 0x${nv.addr.toString(16)} put the ` +
1766
+ `caller's ${nv.reg} back — no C value names it`,
1767
+ );
1768
+ case 'return-address':
1769
+ return fail(
1770
+ `cannot lift '${name}': the return address the 'mflr' at 0x${nv.addr.toString(16)} ` +
1771
+ `copies out is used as a value — a function reading its own return address has no C spelling`,
1772
+ );
1773
+ case 'fctiwz-high':
1774
+ return fail(
1775
+ `cannot lift '${name}': the high word the 'stfd' at 0x${nv.addr.toString(16)} stores is reloaded and ` +
1776
+ `used — 'fctiwz' leaves that half of its result undefined`,
1777
+ );
1778
+ }
1779
+ };
1780
+ /** Checked on the finished function, after `finish` pruned the joins nothing reads. */
1781
+ const assertNoNonValueEscaped = (fnBlocks: Block[]) => {
1782
+ for (const b of fnBlocks) {
1783
+ for (const op of b.ops) {
1784
+ for (const v of [...op.operands, ...op.successors.flatMap((s) => s.args)]) {
1785
+ const nv = nonValues.get(v);
1786
+ if (nv !== undefined) {
1787
+ refuseNonValue(nv);
1788
+ }
1789
+ }
1790
+ }
1791
+ }
1792
+ };
1793
+ const assertNonValuesConsumed = () => {
1794
+ for (const [v, nv] of nonValues) {
1795
+ if (NON_VALUE_RULES[nv.kind].mustBeConsumed && !consumedNonValues.has(v)) {
1796
+ refuseNonValue(nv);
1797
+ }
1798
+ }
1799
+ };
633
1800
  const RET = target.returnReg;
634
1801
  const ARG_REGS = target.argRegs;
635
1802
 
@@ -638,9 +1805,15 @@ export function lift(
638
1805
  //
639
1806
  // A SAVE holds an entry value that is no C value: a callee-saved register's, the link register's
640
1807
  // word, or the back chain the frame push stores (the caller's r1). Nothing but the caller reads it
641
- // back, so its restore — a load into the register it was saved from — is dropped, and a load into
642
- // any other register refuses. Function-scoped, so a save in the prologue matches a restore in any
643
- // epilogue.
1808
+ // back. Its restore — a load into the register it was saved from — puts the caller's value back,
1809
+ // which no C value names, so a read of the register after it refuses (`restored` in
1810
+ // `nonValues`); a load into any other register refuses at once. Function-scoped, so a save in the
1811
+ // prologue matches a restore in any epilogue.
1812
+ //
1813
+ // ONE SAVE AND ONE RESTORE, whichever instruction does it: `stmw`/`lmw`, a single `stw`/`lwz`
1814
+ // and a register save routine all go through `saveInto` and `restoreFrom`. A single `stw` is the
1815
+ // one save that may also be a spill: of a register this function already wrote, it stores a
1816
+ // value. A `stmw` or a routine has no such reading, and saving a written register refuses.
644
1817
  //
645
1818
  // Every other word store is a VALUE — an argument's home, or a live value spilled — and it is the
646
1819
  // shared stack-slot variable (`stackSlotKey`, frontend/ssa.ts) MIPS and Thumb use. A reload into
@@ -650,7 +1823,29 @@ export function lift(
650
1823
  // A slot is one kind for the whole function: a word both saved and stored with a value has no
651
1824
  // single reading, and refuses.
652
1825
  const saveSlots = new Map<number, string>();
653
- const valueSlots = new Map<number, { addr: number; mem: string; read: boolean }>();
1826
+ // Each reload (`lwz`, `lmw`, restore routine) that restores a save, and the registers it restores:
1827
+ // a save holds only while every path restores it after its last write (`refuseSavedRegisterLeaks`).
1828
+ const restoredBy = new Map<Instr, ReadonlySet<string>>();
1829
+ // AN INCOMING STACK WORD HAS ONE NAME, `@sarg<k>`, for its loads and its stores: the EABI passes
1830
+ // integer argument word k >= 8 at entry r1 + 8 + 4(k - 8), in the caller's parameter area past
1831
+ // its back chain and link-register save word (`corpus/probe-mwcc-frame.c`, `in9`/`in10`). A read
1832
+ // no store reaches is the caller's argument through the ordinary live-in path, and a store makes
1833
+ // the parameter an assigned one, as Thumb and MIPS name it. Words only: a sub-word incoming
1834
+ // argument (`lha 10(r1)` for a `short`) refuses as a sub-word frame access, a float past f8 is
1835
+ // loaded with the unmodelled `lfs`/`lfd`, and a function reading one beside a float argument or
1836
+ // declared with a 64-bit or floating parameter refuses (`ppcArgSlots`). Undefined for every
1837
+ // other word.
1838
+ const incomingStackKey = (entryOff: number): string | undefined =>
1839
+ incomingStackArgKey(entryOff, PARAM_AREA_AT, ARG_REGS.length);
1840
+ const valueSlots = new Map<
1841
+ number,
1842
+ { addr: number; mem: string; read: boolean; stored: { v: Value; at: number }[] }
1843
+ >();
1844
+ // What each `mfcr` copied: the fields the condition register held, by the value the copy stands
1845
+ // for. Nothing defines that value — it is no integer — and only an extract of one bit of a float
1846
+ // compare's field reads it (`crBitValue`); anything else that does refuses once the SSA is finished.
1847
+ const crCopies = new Map<Value, ReadonlyMap<string, CrCompare | string>>();
1848
+ const loadedFrom = target.compilerBehaviors.outgoingAreaNeverLoadedFrom;
654
1849
  const refuseMixedSlot = (ins: Instr, mem: string, off: number, kind: 'save' | 'value') => {
655
1850
  if (kind === 'save' ? valueSlots.has(off) : saveSlots.has(off)) {
656
1851
  throw new PpcUnsupportedError(
@@ -692,10 +1887,149 @@ export function lift(
692
1887
  ? `${name}+0x${blocks[bi].startAddr.toString(16)}`
693
1888
  : `the return block '${name}' synthesizes`;
694
1889
 
695
- // A FLOAT RETURN is decided for the whole function (frontend/fpu.ts `writesFloatReturn`), and then
696
- // r3 is scratch like any other integer register.
1890
+ // A FLOAT RETURN is decided for the whole function (frontend/fpu.ts `returnHomeRule`), and then
1891
+ // r3 is scratch like any other integer register. An undecided one returns r3 here, and
1892
+ // `settleReturnHome` may hand each return f1 instead. A table base left in r3 is no return value, as a
1893
+ // pending half is none (`placeholders`); reading one would refuse, and a merge of one with a value
1894
+ // still does.
697
1895
  const returnOperands = (bi: number): Value[] =>
698
- floatReturn ? [readReg(fpu!.returnReg, bi)] : ssa.hasReachingDef(RET, bi) ? [readReg(RET, bi)] : [];
1896
+ floatReturn
1897
+ ? [readReg(fpu!.returnReg, bi)]
1898
+ : ssa.hasReachingDef(RET, bi, (v) => !placeholders.has(v) && !parked(v, RET))
1899
+ ? [readReg(RET, bi)]
1900
+ : [];
1901
+ // A RETURN NOTHING WROTE r3 FOR STILL RETURNS r3 — the caller's value, which on PowerPC is the
1902
+ // first argument. `int f(int a, int x) { if (x) return a; … return 5; }` is `cmpwi r4,0 ; bnelr`:
1903
+ // that path neither reads nor writes r3, so `returnOperands` finds no definition and returns
1904
+ // nothing, while every other path returns a word — C that falls off the end of an `int` function,
1905
+ // and that loses `a`. Whether the function returns a word is known only once every block is
1906
+ // filled, so a bare return is recorded here and given r3's entry value after the fill
1907
+ // (`returnUntouchedArg`) when another return hands back a word. Only where r3 still holds the
1908
+ // entry value on every path into it: a path a call destroyed r3 on returns nothing a C expression
1909
+ // names, and stays bare.
1910
+ //
1911
+ // AND ONLY WHERE r3 IS AN ARGUMENT AT ALL (`r3IsAnArgument`). "r3 still holds what the caller
1912
+ // left" says nothing about whether the caller left anything: `void f(void) { if (gA & 2) return;
1913
+ // gB++; }` uses r3 as a temporary on the other path, so that path "returns a word", and filling
1914
+ // the bare return would mint a parameter the ABI never passed — `f(s32 a0)`, which contradicts a
1915
+ // declared `void f(void)` (the candidate does not compile) and a C++ `__Fv` mangling.
1916
+ const bareReturns: Array<{ op: Op; bi: number }> = [];
1917
+ const wordReturns: number[] = [];
1918
+ // What r4, the second register of an r3:r4 return, may hold at each word return where no
1919
+ // declaration pins the width (frontend/dead-reads.ts `unreadPairHalf`).
1920
+ const unreadHigh = new Set<Value>();
1921
+ // Each of those word returns: what r3 returns there and what r4 may hold (`computedPairHalf`).
1922
+ const pairReturns: PairReturn[] = [];
1923
+ // Every value a compare consumed, whether or not a branch reads the field it set: a compare
1924
+ // nothing reads leaves no op, so this is the only record that the value was used
1925
+ // (frontend/dead-reads.ts `compared`).
1926
+ const compared = new Set<Value>();
1927
+ // Every value a carry reader consumed by folding it away: the constant a `subfze` reads, and the
1928
+ // `rB - 1` a `subfe rD,rA,rB` cancels. The machine read them, and no op says so (`computedPairHalf`).
1929
+ const readWithoutOp = new Set<Value>();
1930
+ // Every `subfc`/`subfic` difference whose carry a reader turned into a 0/1 truth value: the
1931
+ // difference may be the by-product of a 32-bit compare, not the low word of a 64-bit result,
1932
+ // which `computedPairHalf` settles (`TruthDifference`). `carrySource` maps each carry compare to
1933
+ // the difference its writer computed, and `truthPending` each reader result one step short of
1934
+ // the truth value to it.
1935
+ const truthDifferences: TruthDifference[] = [];
1936
+ const carrySource = new Map<Value, Value>();
1937
+ const truthPending = new Map<Value, { diff: Value; via: 'neg' | 'and1' }>();
1938
+ // Every load into the register the compiler discards an unused read into at a level that keeps a
1939
+ // plain one (`compilerBehaviors.discardsUnusedReadsInto`, frontend/dead-reads.ts `discarded`).
1940
+ const discarded = new Set<Value>();
1941
+ const discardReg = target.compilerBehaviors.discardsUnusedReadsInto;
1942
+ // Every value written to a register other than the discard register: `leftInDiscardRegister`
1943
+ // takes out of `discarded` each load one of them derives from, which mwcc put in a home.
1944
+ const writtenElsewhere: Value[] = [];
1945
+ // A `subfze` over a register this block takes in from another block: modelled as `~rA + CA`, and
1946
+ // kept only where `ssa.finish` resolves rA to one constant on every path in (`subfze` case below).
1947
+ const joinedCarryConsts: { ins: Instr; not: Op; add: Op; ca: Value }[] = [];
1948
+ let wordReturn = false;
1949
+ /** Every return, as it is emitted. */
1950
+ const returns: { op: Op; bi: number }[] = [];
1951
+ const emitReturn = (ops: Op[], bi: number) => {
1952
+ const op = mkOp('ret', { operands: returnOperands(bi) });
1953
+ ops.push(op);
1954
+ returns.push({ op, bi });
1955
+ if (floatReturn) {
1956
+ return;
1957
+ }
1958
+ if (op.operands.length > 0) {
1959
+ wordReturn = true;
1960
+ wordReturns.push(bi);
1961
+ const high = new Set<Value>();
1962
+ unreadPairHalf(ssa, prototypes[name], ARG_REGS[1], bi, high);
1963
+ high.forEach((v) => unreadHigh.add(v));
1964
+ pairReturns.push({ word: op.operands[0], high });
1965
+ } else {
1966
+ bareReturns.push({ op, bi });
1967
+ }
1968
+ };
1969
+ /** Whether every word return hands back a value computed ONLY to be returned: on every path into
1970
+ * it, r3's last write is one (frontend/dead-reads.ts `computedOnlyToReturn`), so the function
1971
+ * returns a word. `int f(int a) { switch (gS) { case 0: return 3; … } return a; }` is `li r3,3 ;
1972
+ * blr` on every case and a bare `blr` out of range. A temporary is read again (`lwz r3,gB ; addi
1973
+ * r0,r3,1`), and a call's result and a load are excluded — a void function's call leaves its
1974
+ * callee's result in r3, and a discarded `volatile` read is a load nobody reads — so `void
1975
+ * f(void) { if (gA & 2) return; gB++; }` still answers no.
1976
+ *
1977
+ * What it cannot tell apart is an `int f(void)` with a path that falls off its end: that path
1978
+ * returns whatever the caller left in r3 too, and is lifted as returning it — the bytes' own
1979
+ * behaviour, at the cost of a parameter the source did not name. A CodeWarrior C++ symbol says
1980
+ * so where there is one (`__Fv`), and a declaration always decides first. */
1981
+ const r3WrittenToReturn = (): boolean => {
1982
+ if (wordReturns.length === 0) {
1983
+ return false;
1984
+ }
1985
+ const computed = computedOnlyToReturn(irBlocks);
1986
+ return wordReturns.every((bi) => ssa.everyReachingDef(RET, bi, computed));
1987
+ };
1988
+ /** Whether the caller passes r3, so a bare return of it hands back an argument. A declaration
1989
+ * decides when it says (`ownFirstArgumentPassed`, which reads it as the live-in bound does): a
1990
+ * non-void return and a parameter word or `this` first, or a nonzero register count — and a free
1991
+ * function of none (`g__Fv`) or a count of 0 is passed nothing, which no body evidence
1992
+ * overrides. A hand-written body (`ownDeclared` false) is read as having none. Without one, only
1993
+ * the body can show it: r3 is already an entry parameter, or a higher argument register is live
1994
+ * in, which puts an argument in r3 by position (the obligation `mintArgSlotHoles` discharges), or
1995
+ * every word return hands back a value computed only to be returned (`r3WrittenToReturn`) — then
1996
+ * the function returns a word, and a bare return of it hands back what the caller left in r3. */
1997
+ const r3IsAnArgument = (): boolean => {
1998
+ const own = prototypes[name];
1999
+ if (declaresVoidReturn(own)) {
2000
+ return false;
2001
+ }
2002
+ // a count bounds no register, but says whether any argument is passed at all
2003
+ if (typeof own?.params === 'number') {
2004
+ return own.params > 0;
2005
+ }
2006
+ const declared = ownDeclared ? ownFirstArgumentPassed(name, own, target) : undefined;
2007
+ if (declared !== undefined) {
2008
+ return declared;
2009
+ }
2010
+ const live = new Set(irBlocks[0].params.map((p) => paramReg.get(p)));
2011
+ return ARG_REGS.some((r) => live.has(r)) || r3WrittenToReturn();
2012
+ };
2013
+ /** Whether the function's own STATED declaration returns a word: then it returns one even where
2014
+ * no path writes r3, and a bare return hands back what the caller left there — `int f(int a) {
2015
+ * return a; }` is a bare `blr`. A context's own entry is not read for this (its return is a
2016
+ * header nothing checked), and neither is a hand-written body's (`ownDeclared` false). */
2017
+ const statesWordReturn = (): boolean => {
2018
+ const r = prototypes[name]?.returns;
2019
+ const width = r === undefined || isFloatingSpelling(r) ? undefined : declaredWidth(r);
2020
+ return ownDeclared && width !== undefined && width <= 32;
2021
+ };
2022
+ const returnUntouchedArg = () => {
2023
+ if (!(wordReturn || statesWordReturn()) || bareReturns.length === 0 || !r3IsAnArgument()) {
2024
+ return;
2025
+ }
2026
+ for (const { op, bi } of bareReturns) {
2027
+ const v = ssa.entryValue(RET, bi);
2028
+ if (v !== undefined) {
2029
+ op.operands = [v];
2030
+ }
2031
+ }
2032
+ };
699
2033
 
700
2034
  const fillBlock = (b: PpcBlock, bi: number) => {
701
2035
  const ops = irBlocks[bi].ops;
@@ -703,23 +2037,33 @@ export function lift(
703
2037
 
704
2038
  // A synthetic conditional-return block: just return the current return register.
705
2039
  if (b.synthReturn) {
706
- const retOps = returnOperands(bi);
707
- ops.push(mkOp('ret', { operands: retOps }));
2040
+ emitReturn(ops, bi);
708
2041
  return;
709
2042
  }
2043
+ // A case target of an elided dispatch is entered with what the idiom left in its registers.
2044
+ const entry = entries.get(bi);
2045
+ if (entry !== undefined) {
2046
+ enterCaseTarget(
2047
+ entry,
2048
+ ops,
2049
+ {
2050
+ key: (r) => r,
2051
+ scrutinee: (r) => readReg(r, bi),
2052
+ dispatchScrutinee: (r) => ssa.readAtEnd(r, entry.at),
2053
+ write: (r, v) => writeVar(r, bi, v),
2054
+ },
2055
+ staleIdiom,
2056
+ );
2057
+ }
710
2058
 
711
2059
  // `lastDef` is the value most recently written to a register in this instruction — used to
712
2060
  // wire a record-form op's implicit cr0 side effect (see the `rc` handling in `decode`).
713
2061
  let lastDef: Value | null = null;
714
- // Reading r1 (the stack pointer) as a DATA operand means frame-pointer arithmetic or an
715
- // address-taken local (`addi r3,r1,8` = `&local`) — not modellable without a stack abstraction,
716
- // and fabricating a value for r1 silently miscompiles. Fail LOUD. (Frame bookkeeping never
717
- // reaches here: stwu / addi r1 / mflr / mtlr / r1-relative spills are handled before `read`.)
2062
+ // r1 read as data fails LOUD (`r1AsData`). Frame bookkeeping never reaches here: stwu / addi r1
2063
+ // / mflr / mtlr / r1-relative spills / the save routines' frame address are handled before `read`.
718
2064
  const read = (r: string): Value => {
719
2065
  if (r === 'r1') {
720
- throw new PpcUnsupportedError(
721
- `cannot lift '${name}': stack pointer r1 used as data (address-taken local / frame arithmetic) — not supported`,
722
- );
2066
+ r1AsData();
723
2067
  }
724
2068
  return readReg(r, bi);
725
2069
  };
@@ -740,10 +2084,50 @@ export function lift(
740
2084
  );
741
2085
  }
742
2086
  };
2087
+ /** The instruction a save or restore refusal names. */
2088
+ const site = (ins: Instr) =>
2089
+ `the '${ins.mnemonic}'${ins.mnemonic === 'bl' && ins.reloc ? ` to '${ins.reloc.sym}'` : ''} at 0x${ins.addr.toString(16)}`;
2090
+ /** Save `reg`'s entry value into the save slot at entry offset `off`, named `mem` in `ins`. */
2091
+ const saveInto = (ins: Instr, reg: string, off: number, mem: string) => {
2092
+ if (ssa.hasReachingDef(reg, bi)) {
2093
+ // A read leaves a definition too: the entry value, as a parameter, since `prologueSavedRegs`
2094
+ // does not count a save after a read.
2095
+ throw new PpcUnsupportedError(
2096
+ ssa.hasReachingDef(reg, bi, (v) => !paramReg.has(v))
2097
+ ? `cannot lift '${name}': ${site(ins)} saves ${reg} after this function wrote it, so '${mem}' holds a ` +
2098
+ `value, not the caller's ${reg}`
2099
+ : `cannot lift '${name}': ${site(ins)} saves ${reg} after this function read it — a value handed ` +
2100
+ `over in a register no argument is passed in, which a compiled prologue saves before any use`,
2101
+ );
2102
+ }
2103
+ refuseMixedSlot(ins, mem, off, 'save');
2104
+ saveSlots.set(off, reg);
2105
+ };
2106
+ /** Restore `reg` from the save slot at `off`: it must be the slot `reg` was saved into, and `reg`
2107
+ * then holds the caller's value again. r1 is no SSA register: its reload is the back chain the
2108
+ * frame push stored, and `r1Displacements` tracks it instead. */
2109
+ const restoreFrom = (ins: Instr, reg: string, off: number, mem: string) => {
2110
+ const saved = saveSlots.get(off);
2111
+ if (saved !== reg) {
2112
+ throw new PpcUnsupportedError(
2113
+ `cannot lift '${name}': reload of '${mem}' into ${reg}, ` +
2114
+ (saved === undefined
2115
+ ? `a word nothing saved it into (${site(ins)})`
2116
+ : `a slot ${saved} was saved into — the word holds the caller's ${saved}, which no C value ` +
2117
+ `names (${site(ins)})`),
2118
+ );
2119
+ }
2120
+ if (reg !== 'r1') {
2121
+ restoredBy.set(ins, new Set([...(restoredBy.get(ins) ?? []), reg]));
2122
+ const caller = mkValue(T.unk(32));
2123
+ nonValues.set(caller, { kind: 'restored', reg, addr: ins.addr });
2124
+ writeVar(reg, bi, caller);
2125
+ }
2126
+ };
743
2127
  // A word store/load based on r1: a SAVE or a VALUE slot (see `saveSlots` above). Returns false
744
2128
  // for any other base, which is ordinary memory.
745
2129
  const frameStore = (ins: Instr, srcReg: string, mem: string): boolean => {
746
- if (parseMem(mem).base !== 'r1') {
2130
+ if (parseMem(mem).base !== 'r1' || inLocalArea(ins, mem, 4) !== undefined) {
747
2131
  return false;
748
2132
  }
749
2133
  const off = entryOffset(ins, mem);
@@ -752,8 +2136,9 @@ export function lift(
752
2136
  // save word, 4 bytes above the entry r1. Anywhere else the word is a value — past the push,
753
2137
  // 8 bytes up is the callee's parameter area, where it is a ninth argument — and a register
754
2138
  // holding the return address on some paths and another value on others is not a save at all.
755
- if (ssa.hasReachingDef(srcReg, bi, (v) => returnAddresses.has(v))) {
756
- const mixed = ssa.hasReachingDef(srcReg, bi, (v) => !returnAddresses.has(v));
2139
+ const isReturnAddress = (v: Value) => nonValueOf(v, 'return-address') !== undefined;
2140
+ if (ssa.hasReachingDef(srcReg, bi, isReturnAddress)) {
2141
+ const mixed = ssa.hasReachingDef(srcReg, bi, (v) => !isReturnAddress(v));
757
2142
  if (mixed || off !== LR_SAVE_WORD) {
758
2143
  throw new PpcUnsupportedError(
759
2144
  `cannot lift '${name}': the store to '${mem}' at 0x${ins.addr.toString(16)} is not the link ` +
@@ -771,34 +2156,44 @@ export function lift(
771
2156
  }
772
2157
  // Any other save stores what the register held at entry: nothing defined it on any path. An
773
2158
  // argument register nothing defined holds the argument, which is a value.
774
- if (!isArg && !ssa.hasReachingDef(srcReg, bi)) {
775
- refuseMixedSlot(ins, mem, off, 'save');
776
- saveSlots.set(off, srcReg);
2159
+ // …save a global register variable's: compiled code stores one only as a value, the outgoing
2160
+ // stack argument `k(1, …, 8, g)` is (`stw r14,8(r1)`), and the builder records no definition
2161
+ // for it (its writes are stores to the global), so "nothing defined it" is no evidence of a save.
2162
+ if (!isArg && !liveIns.globalRegs?.has(srcReg) && !ssa.hasReachingDef(srcReg, bi)) {
2163
+ saveInto(ins, srcReg, off, mem);
777
2164
  return true;
778
2165
  }
779
2166
  refuseMixedSlot(ins, mem, off, 'value');
780
2167
  // An argument moved to its home is not an argument set up for the next call.
781
2168
  const v = (isArg ? ssa.entryValue(srcReg, bi) : undefined) ?? readReg(srcReg, bi);
2169
+ // An assignment to an incoming stack parameter is that parameter's SSA variable, and no call
2170
+ // made here takes the word: it is the caller's (a tail branch, whose callee would read the
2171
+ // same words as its own stack arguments, refuses in `liftPrelude`).
2172
+ const incoming = incomingStackKey(off);
2173
+ if (incoming !== undefined) {
2174
+ writeVar(incoming, bi, v);
2175
+ return true;
2176
+ }
782
2177
  writeVar(stackSlotKey(off), bi, v);
783
2178
  if (!valueSlots.has(off)) {
784
- valueSlots.set(off, { addr: ins.addr, mem, read: false });
2179
+ valueSlots.set(off, { addr: ins.addr, mem, read: false, stored: [] });
785
2180
  }
2181
+ valueSlots.get(off)!.stored.push({ v, at: ins.addr });
786
2182
  return true;
787
2183
  };
788
2184
  const frameLoad = (ins: Instr, dstReg: string, mem: string): boolean => {
789
- if (parseMem(mem).base !== 'r1') {
2185
+ if (parseMem(mem).base !== 'r1' || inLocalArea(ins, mem, 4) !== undefined) {
790
2186
  return false;
791
2187
  }
792
2188
  const off = entryOffset(ins, mem);
793
- const saved = saveSlots.get(off);
794
- if (saved === dstReg) {
2189
+ if (saveSlots.has(off)) {
2190
+ restoreFrom(ins, dstReg, off, mem);
795
2191
  return true;
796
2192
  }
797
- if (saved !== undefined) {
798
- throw new PpcUnsupportedError(
799
- `cannot lift '${name}': reload of '${mem}' into ${dstReg}, a slot ${saved} was saved into — the ` +
800
- `word holds the caller's ${saved}, which no C value names`,
801
- );
2193
+ const incoming = incomingStackKey(off);
2194
+ if (incoming !== undefined) {
2195
+ write(dstReg, readVar(incoming, bi));
2196
+ return true;
802
2197
  }
803
2198
  const slot = valueSlots.get(off);
804
2199
  if (slot === undefined || !ssa.hasReachingDef(stackSlotKey(off), bi)) {
@@ -810,9 +2205,92 @@ export function lift(
810
2205
  write(dstReg, readVar(stackSlotKey(off), bi));
811
2206
  return true;
812
2207
  };
2208
+ /** The instruction `decode` is lifting, for a refusal raised below it. */
2209
+ let decoding: Instr | null = null;
2210
+ // THE STACK POINTER IS NO VALUE A LIFT WRITES. The frame push, `addi r1,r1,N` and the reload of
2211
+ // the back chain move r1, and `r1Displacements` follows them; none reaches here. Any other write
2212
+ // (`lwz r1,4(r3)` in a longjmp, `mr r1,r10` in a context switch) would land in an SSA
2213
+ // variable nothing reads, since a read of r1 is refused as data (`read`), and the lift
2214
+ // would drop the write: a load into r1 would become a discarded volatile read. So it refuses.
2215
+ // THE SMALL-DATA BASES ARE NONE EITHER, for the same reason one step removed: an access through
2216
+ // r13 or r2 is read off its `R_PPC_EMB_SDA21` relocation (`sdaAccess`), never off the register,
2217
+ // so a write to either (`lwzu r13,4(r3)`, `mr r2,r4`) would be dropped and every later small-data
2218
+ // access would still name its symbol as if the base had not moved.
813
2219
  const write = (r: string, v: Value) => {
2220
+ // the instruction the listing holds, not the plain access an update form is decoded through
2221
+ const by = () => {
2222
+ const listed = decoding === null ? null : (decodedFrom.get(decoding) ?? decoding);
2223
+ return listed === null
2224
+ ? 'an instruction'
2225
+ : `'${listed.mnemonic} ${listed.ops.join(',')}' at 0x${listed.addr.toString(16)}`;
2226
+ };
2227
+ if (r === 'r1') {
2228
+ throw new PpcUnsupportedError(
2229
+ `cannot lift '${name}': ${by()} writes the stack pointer r1 — only the frame push, 'addi r1,r1,N' ` +
2230
+ `and the back-chain reload move r1 here, and the value it would take is not one this lift tracks`,
2231
+ );
2232
+ }
2233
+ if (SMALL_DATA_BASES.includes(r)) {
2234
+ throw new PpcUnsupportedError(
2235
+ `cannot lift '${name}': ${by()} writes the small-data base ${r} — a small-data access is read ` +
2236
+ `off its relocation, not off the register, so a moved base is not one this lift tracks`,
2237
+ );
2238
+ }
814
2239
  writeVar(r, bi, v);
815
2240
  lastDef = v;
2241
+ if (discardReg !== undefined && r !== discardReg) {
2242
+ writtenElsewhere.push(v);
2243
+ }
2244
+ };
2245
+ // A `bl` into a register save or restore routine is frame bookkeeping, not a call: it writes no
2246
+ // register this function reads (only the link register, already copied out — which
2247
+ // `linkRegisterProof` checks), so nothing is destroyed and r0 keeps the return address
2248
+ // `mflr` put there. A SAVE records a save slot per word, at the routine's fixed offsets below the
2249
+ // frame address in r11; a RESTORE must find the same register in every word it reloads. It
2250
+ // refuses where:
2251
+ // • r11 does not hold one frame address on every path into the call, or a call destroyed it on
2252
+ // the way (`useOutsideOps`; and a frame address no routine takes refuses at the end of `lift`);
2253
+ // • r1's displacement at the `addi` is not known, so no slot can be named;
2254
+ // • the routine's number is outside r14..r31 (no such runtime routine);
2255
+ // • it moves floating-point registers (named, not modelled);
2256
+ // • a register it saves was written or read first, or a word it restores was not saved from
2257
+ // that register — `saveInto` and `restoreFrom`, the rules every save and restore keeps.
2258
+ const saveRoutine = (ins: Instr, r: NonNullable<ReturnType<typeof saveRoutineOf>>) => {
2259
+ const at = `the 'bl' to '${r.sym}' at 0x${ins.addr.toString(16)}`;
2260
+ if (r.first < 14 || r.first > 31 || ins.reloc!.addend !== 0) {
2261
+ throw new PpcUnsupportedError(`cannot lift '${name}': ${at} names no register save routine of the runtime`);
2262
+ }
2263
+ // The FPR half is named and refused. Every compiled function that calls it (doubles live
2264
+ // across a call, a 24-float leaf; 2.4.2 and 2.4.7 at -O4,p, 2.3.3 at -O4,s) already declines
2265
+ // on the unmodelled `stfd`/`lfd` or on a float live across a call, and no row calls it, so a
2266
+ // model of its 8-byte slots would be a model no lift has checked.
2267
+ if (r.fpr) {
2268
+ throw new PpcUnsupportedError(
2269
+ `cannot lift '${name}': ${at} saves or restores floating-point registers — not modelled`,
2270
+ );
2271
+ }
2272
+ const fr = target.registerSaveRoutines!.frameReg;
2273
+ const base = ssa.hasReachingDef(fr, bi, (v) => !nonValueOf(v, 'frame-address')) ? undefined : readVar(fr, bi);
2274
+ const frame = base === undefined ? undefined : nonValueOf(base, 'frame-address');
2275
+ if (frame === undefined) {
2276
+ throw new PpcUnsupportedError(
2277
+ `cannot lift '${name}': ${at} does not find one frame address ('addi ${fr},r1,N') in ${fr} on every path`,
2278
+ );
2279
+ }
2280
+ consumedNonValues.add(base!);
2281
+ // If a call between the `addi` and here destroyed r11, the routine would store through what the
2282
+ // callee left there: the builder refuses that read once it knows it was used.
2283
+ ssa.useOutsideOps(base!, bi);
2284
+ if (typeof frame.off !== 'number') {
2285
+ throw new PpcUnsupportedError(
2286
+ `cannot lift '${name}': ${at} saves below the frame address formed at 0x${frame.addr.toString(16)}, where ` +
2287
+ `the stack pointer's offset from its entry value is not known: ${frame.off}`,
2288
+ );
2289
+ }
2290
+ for (let k = r.first; k <= 31; k++) {
2291
+ const disp = -4 * (32 - k);
2292
+ (r.restore ? restoreFrom : saveInto)(ins, `r${k}`, frame.off + disp, `${disp}(${fr})`);
2293
+ }
816
2294
  };
817
2295
  // Shared emitter kit (frontend/emit.ts) — the ISA-specific readers/guards stay above.
818
2296
  const kit = mkEmitKit(ops, write);
@@ -822,6 +2300,31 @@ export function lift(
822
2300
  // emit an honest `opaque` instead, which fails LOUD at assertResolved whether or not anything
823
2301
  // reads that register (see frontend/opaque.ts for the policy).
824
2302
  const emitOpaqueDest = (ins: Instr) => {
2303
+ // AN OPAQUE STANDS FOR ONE DESTINATION, `ops[0]`. An unmodelled load that writes a second
2304
+ // GPR would leave that one holding its old value with nothing marking it, which annotate mode
2305
+ // (the CLI and bench default) then prints as ordinary C: `lwzux r0,r3,r4; lbz r3,0(r3)` read
2306
+ // `*a0` where the byte is at `a0 + a1`. So refuse, as a store refuses, every load that writes
2307
+ // more than its first operand: an indexed update load (`lbzux`, `lhzux`, `lhaux`, `lwzux`,
2308
+ // which also write rA, as `mayWriteGpr` knows), `lmw` below r31 and the string loads. The FP
2309
+ // update forms keep the floating-point refusal `opaqueDest` gives them. The test is on the
2310
+ // lower-cased mnemonic, as `opaqueDest`'s classes are case-blind: the decode's cases are
2311
+ // lower case, so an upper-case `LWZU` or `LBZUX` (pasted asm; objdump and dtk never print
2312
+ // one) reaches here undecoded, and the displacement update forms are refused with the indexed.
2313
+ const mn = ins.mnemonic.toLowerCase();
2314
+ const lmwFirst = mn === 'lmw' ? Number(ins.ops[0]?.slice(1)) : NaN;
2315
+ const second = /^l(?!f)\w*ux?$/.test(mn)
2316
+ ? `its base ${mn.endsWith('x') ? ins.ops[1] : parseMem(ins.ops[1] ?? '').base}`
2317
+ : lmwFirst < 31
2318
+ ? `r${lmwFirst + 1}${lmwFirst + 1 < 31 ? '..r31' : ''}`
2319
+ : /^lsw[ix]$/.test(mn)
2320
+ ? 'the registers after it'
2321
+ : null;
2322
+ if (second !== null) {
2323
+ throw new PpcUnsupportedError(
2324
+ `cannot lift '${name} @0x${ins.addr.toString(16)}': unmodelled instruction '${ins.mnemonic}' writes ${second} ` +
2325
+ `as well as ${ins.ops[0]}, and an opaque stands for one destination`,
2326
+ );
2327
+ }
825
2328
  // storeClass: every PPC store mnemonic is st* — an unmodelled one (`stwbrx`, `sthbrx`, …)
826
2329
  // must throw, never skip (its first token is the SOURCE register).
827
2330
  const od = opaqueDest(ins.mnemonic, ins.ops, {
@@ -877,6 +2380,15 @@ export function lift(
877
2380
  ops.push(mkOp('gaddr', { results: [g], attrs: { sym } }));
878
2381
  return g;
879
2382
  };
2383
+ // A FRAME ADDRESS AS DATA: post-push offset k of the local area, lowered to the `laddr` the
2384
+ // frame-object audit proves (`FRAME_OBJECT_AUDIT`, frontend/frame-objects.ts `frameObjects`),
2385
+ // which also judges where it lies against the area: below it or above it refuses.
2386
+ const emitLaddr = (k: number): Value => {
2387
+ const v = mkValue(T.unk(32));
2388
+ ops.push(mkOp('laddr', { results: [v], attrs: { off: k } }));
2389
+ laddrAt.set(v, k);
2390
+ return v;
2391
+ };
880
2392
  // `lwz rD,0(0)` under an `R_PPC_EMB_SDA21` relocation is a SMALL-DATA access. Both printed
881
2393
  // fields are link-time placeholders: the linker substitutes r13/r2 for the base register and a
882
2394
  // section-relative displacement for the offset, so the address is exactly `&SYM` plus the
@@ -898,11 +2410,15 @@ export function lift(
898
2410
  // The base value + byte offset of a displacement memory operand: a small-data global, or an
899
2411
  // ordinary register base. Read BEFORE a store's source register, because read order decides
900
2412
  // block-parameter layout.
901
- const memOperand = (ins: Instr, mem: string): { base: Value; off: number } => {
2413
+ const memOperand = (ins: Instr, mem: string, width: number): { base: Value; off: number } => {
902
2414
  const sda = sdaAccess(ins, mem);
903
2415
  if (sda) {
904
2416
  return sda;
905
2417
  }
2418
+ const local = inLocalArea(ins, mem, width);
2419
+ if (local !== undefined) {
2420
+ return { base: emitLaddr(local), off: 0 };
2421
+ }
906
2422
  assertOrdinaryMem(mem);
907
2423
  const { off, base } = parseMem(mem);
908
2424
  return { base: read(base), off };
@@ -940,12 +2456,41 @@ export function lift(
940
2456
  const hi = highHalves.pair(relocSite(ins), rHi, readVar(rHi, bi), lo.sym, (h) => h.addend === lo.addend);
941
2457
  return emitGaddr(hi.sym);
942
2458
  };
2459
+ /** `addi rD,rHi,SYM@l` completing the base of a jump table a dispatch reads from outside its
2460
+ * block: the pair is proven as `foldLoHalf` proves it, and rD holds the table base. */
2461
+ const foldTableBase = (ins: Instr, rHi: string): Value | null => {
2462
+ const lo = ins.reloc!;
2463
+ const key = { sym: lo.sym, addend: lo.addend };
2464
+ if (!tableBases.wants(key)) {
2465
+ return null;
2466
+ }
2467
+ relocTaken = true;
2468
+ highHalves.pair(relocSite(ins), rHi, readVar(rHi, bi), lo.sym, (h) => h.addend === lo.addend);
2469
+ const v = mkValue(T.unk(32));
2470
+ tableBases.record(v, key, ins.addr);
2471
+ return v;
2472
+ };
943
2473
  const emitLoad = (ins: Instr, d: string, mem: string, width: number, signed: boolean) => {
944
- const { base, off } = memOperand(ins, mem);
945
- emit('load', d, [base], { off, width, signed });
2474
+ const { base, off } = memOperand(ins, mem, width);
2475
+ const v = emit('load', d, [base], { off, width, signed });
2476
+ if (d === discardReg) {
2477
+ discarded.add(v);
2478
+ }
946
2479
  };
947
2480
  const emitStore = (ins: Instr, srcReg: string, mem: string, width: number) => {
948
- const { base, off } = memOperand(ins, mem);
2481
+ const { base, off } = memOperand(ins, mem, width);
2482
+ // A register the prologue saves is saved above the local area by construction, so its
2483
+ // entry value stored inside it is a frame this does not describe.
2484
+ if (
2485
+ inLocalArea(ins, mem, width) !== undefined &&
2486
+ isCalleeSaved(target, srcReg) &&
2487
+ !ssa.hasReachingDef(srcReg, bi)
2488
+ ) {
2489
+ fail(
2490
+ `cannot lift '${name}': address-taken stack local — '${ins.mnemonic}' at 0x${ins.addr.toString(16)} stores ` +
2491
+ `the caller's ${srcReg} inside the local area, a frame this does not describe`,
2492
+ );
2493
+ }
949
2494
  ops.push(mkOp('store', { operands: [base, read(srcReg)], attrs: { off, width } }));
950
2495
  };
951
2496
  // Register+register INDEXED addressing (`lwzx rD,rA,rB` = *(rA+rB), `stwx rS,rA,rB` = *(rA+rB)=rS).
@@ -959,7 +2504,10 @@ export function lift(
959
2504
  return addr;
960
2505
  };
961
2506
  const emitLoadX = (d: string, rA: string, rB: string, width: number, signed: boolean) => {
962
- emit('load', d, [addrX(rA, rB)], { off: 0, width, signed });
2507
+ const v = emit('load', d, [addrX(rA, rB)], { off: 0, width, signed });
2508
+ if (d === discardReg) {
2509
+ discarded.add(v);
2510
+ }
963
2511
  };
964
2512
  const emitStoreX = (srcReg: string, rA: string, rB: string, width: number) => {
965
2513
  ops.push(mkOp('store', { operands: [addrX(rA, rB), read(srcReg)], attrs: { off: 0, width } }));
@@ -973,11 +2521,12 @@ export function lift(
973
2521
  // mwcc's switch dispatch `cmpwi r0,1; beq- case1; bge- default` reads one compare from the
974
2522
  // `beq-` and then from the `bge-` on its fall-through, which begins a block of its own.
975
2523
  //
976
- // The clobber rule is this frontend's own: a compare writes its field, a record-form op writes
977
- // cr0, and a call destroys every volatile field (CR_VOLATILE). A field something else wrote holds
978
- // the sentence saying so rather than an entry, so the branch that reads it can name the writer.
979
- // Every other instruction that writes a CR field (`crxor`, `mtcrf`, `fcmpu`, …) names no
980
- // register destination and is refused outright by the opaque policy, so it never reaches here.
2524
+ // The clobber rule is this frontend's own: a compare writes its field (a float compare too), a
2525
+ // record-form op writes cr0, a `cror` rewrites one bit of a float compare's field, and a call
2526
+ // destroys every volatile field (CR_VOLATILE). A field something else wrote holds the sentence
2527
+ // saying so rather than an entry, so the branch that reads it can name the writer. Every other
2528
+ // instruction that writes a CR field (`crxor`, `mtcrf`, …) names no register destination and is
2529
+ // refused outright by the opaque policy, so it never reaches here.
981
2530
  const inherited = inheritFlags(bi, {
982
2531
  preds,
983
2532
  exit: exitCr,
@@ -998,11 +2547,100 @@ export function lift(
998
2547
  const args = hasCr ? cmpOps.slice(1) : cmpOps;
999
2548
  return { cr: hasCr ? cmpOps[0] : 'cr0', lhsTok: args[0], rhsTok: args[args.length - 1] };
1000
2549
  };
2550
+ // An FPU register holds a float or, after `fctiwz`, an integer word (frontend/fpu.ts
2551
+ // `refuseFpWordCrossings` checks the edges); a float read refuses the word.
2552
+ const readFloat = (ins: Instr, key: string): Value => {
2553
+ const v = readReg(key, bi);
2554
+ if (v.type.kind !== 'float') {
2555
+ throw new PpcUnsupportedError(
2556
+ `cannot lift '${name}': '${ins.mnemonic}' at 0x${ins.addr.toString(16)} reads ${key} as a float, and it ` +
2557
+ 'holds an integer word',
2558
+ );
2559
+ }
2560
+ return v;
2561
+ };
2562
+ // A compare or a conversion refuses beside a call (`callInsn`), and in a function its own
2563
+ // declaration puts in both precisions (`declaredPrecision`): neither states a precision itself.
2564
+ const refuseCrossingUnmodelled = (ins: Instr) => {
2565
+ if (callInsn) {
2566
+ throw new PpcUnsupportedError(
2567
+ `cannot lift '${name}': '${ins.mnemonic}' computes on a float and '${callInsn.mnemonic}' at ` +
2568
+ `0x${callInsn.addr.toString(16)} makes a call — the floating-point registers a call reads, returns ` +
2569
+ `in and destroys are not modelled`,
2570
+ );
2571
+ }
2572
+ if (fpType === null) {
2573
+ refuseBothPrecisions(name, ins.mnemonic);
2574
+ }
2575
+ };
2576
+ // ONE BIT of an `mfcr` copy, as the 0/1 it is: bit `n` counts from the most significant, four to
2577
+ // a field (`lt`, `gt`, `eq`, `so`). `rlwinm rD,r0,K,31,31` rotates bit K-1 to the bottom and
2578
+ // `srwi rD,r0,31` takes bit 0. Only a float compare's bit is read this way here, and it is the
2579
+ // compare that bit holds (`CrBits`); anything else refuses.
2580
+ const crBitValue = (ins: Instr, copy: Value, n: number): Value => {
2581
+ const fields = crCopies.get(copy)!;
2582
+ const field = `cr${n >> 2}`;
2583
+ const bit = n < 0 ? null : CR_BIT_NAMES[n & 3];
2584
+ const held = fields.get(field);
2585
+ if (bit === null || typeof held !== 'object' || !held.bits || (held.quiet && bit !== 'eq')) {
2586
+ throw new PpcUnsupportedError(
2587
+ `cannot lift '${name}': '${ins.mnemonic}' at 0x${ins.addr.toString(16)} extracts from a copy of the ` +
2588
+ 'condition register something other than one bit a float compare set',
2589
+ );
2590
+ }
2591
+ const v = mkValue(T.unk(32));
2592
+ ops.push(mkOp(held.bits[bit], { operands: [held.lhs, held.rhs], results: [v] }));
2593
+ return v;
2594
+ };
1001
2595
  const recordCmp = (cr: string, lhs: Value, rhs: Value, signed: boolean) => {
1002
2596
  cmpDef.set(cr, { lhs, rhs, signed });
2597
+ compared.add(lhs).add(rhs);
2598
+ };
2599
+
2600
+ // The carry bit, XER[CA], as a 0/1 value. mwcc computes a materialised compare branch-free
2601
+ // through it (`subfc` then `subfe`/`subfze`/`adde`/`addze`), so a reader needs the comparison
2602
+ // the writer left there. Three writers are modelled, each as the compare its carry IS:
2603
+ // subfc rD,rA,rB rD = rB - rA CA = rB >=u rA
2604
+ // subfic rD,rA,K rD = K - rA CA = K >=u rA (K == 0: rA == 0)
2605
+ // addic rD,rA,K rD = rA + K CA = carry of the unsigned add (K == -1: rA != 0)
2606
+ // The compare is built at its reader, so a carry nothing reads adds no op. A carry is
2607
+ // UNKNOWN, and its reader refuses naming why, when no instruction of this block set it (a carry
2608
+ // across a block edge is not modelled), a call clobbered it, or a writer this frontend does not
2609
+ // model set it: `srawi`/`sraw` (the signed `/2^k` round-up), `addc` (the low half of a 64-bit
2610
+ // add), and the extended forms, whose own carry-out is a high word's. An `addic` carry is
2611
+ // read only as a compare (`subfe`, `subfze`): `adde`/`addze` over it is the high word of
2612
+ // mwcc 2.4.x's 64-bit add of a constant, which is not built, so they refuse it as they refuse
2613
+ // an `addc` carry.
2614
+ type Carry = { make: () => Value; addic: boolean } | { why: string };
2615
+ let carry: Carry = { why: 'nothing in its own block set it (a carry across a block edge is not modelled)' };
2616
+ const setCarry = (make: () => Value, addic = false) => {
2617
+ carry = { make, addic };
2618
+ };
2619
+ /** The carry, as the compare that set it. `wide` marks an extended add (`adde`/`addze`), which
2620
+ * over an `addic` carry is the high word of a 64-bit add. */
2621
+ const readCarry = (ins: Instr, wide = false): Value => {
2622
+ const why =
2623
+ 'why' in carry
2624
+ ? carry.why
2625
+ : wide && carry.addic
2626
+ ? 'an addic set it: an extended add over that carry is the high word of a 64-bit add, which is not modelled'
2627
+ : undefined;
2628
+ if (why !== undefined) {
2629
+ throw new PpcUnsupportedError(
2630
+ `cannot lift '${name}': the '${ins.mnemonic}' at 0x${ins.addr.toString(16)} ` +
2631
+ `reads the carry bit (XER[CA]), and ${why}`,
2632
+ );
2633
+ }
2634
+ return (carry as { make: () => Value }).make();
2635
+ };
2636
+ /** The value of `v` when this block defines it as a `const`. */
2637
+ const constIn = (v: Value): number | undefined => {
2638
+ const def = ops.find((o) => o.results[0] === v);
2639
+ return def?.opcode === 'const' ? (def.attrs.value as number) : undefined;
1003
2640
  };
1004
2641
 
1005
2642
  const decode = (ins: Instr) => {
2643
+ decoding = ins;
1006
2644
  const [d, s, t] = ins.ops;
1007
2645
  // Record form: a trailing `.` (the Rc bit) means the op ALSO sets cr0 from a signed compare
1008
2646
  // of its result against 0 (e.g. `andi.`, `addic.`). Decode the base op, then record that
@@ -1012,6 +2650,75 @@ export function lift(
1012
2650
  const mnem = rc ? ins.mnemonic.slice(0, -1) : ins.mnemonic;
1013
2651
  lastDef = null;
1014
2652
  relocTaken = false;
2653
+ // Every carry writer and every call leaves the carry unknown unless its case below models it.
2654
+ const clobberCarry = (how: string) => {
2655
+ carry = { why: `the '${ins.mnemonic}' at 0x${ins.addr.toString(16)} ${how}` };
2656
+ };
2657
+ if (CARRY_WRITERS.test(mnem)) {
2658
+ clobberCarry('set it, a carry this frontend does not model');
2659
+ } else if (/^b.*l$/.test(mnem)) {
2660
+ clobberCarry('is a call, which clobbers it');
2661
+ }
2662
+ // A LOAD OR STORE WITH UPDATE (`lbzu`/`lhzu`/`lhau`/`lwzu`, `stbu`/`sthu`/`stwu`) addresses
2663
+ // `rA + D` and leaves that address in rA: `lbzu rD,1(r3)` is `r3 += 1; rD = *(u8 *)r3`, and
2664
+ // `stbu rS,1(r3)` is `r3 += 1; *(u8 *)r3 = rS`. mwcc emits them for a fully unrolled scan of a
2665
+ // table, a run of stores to a global array and a pre-increment pointer walk (`*++p`) at -O4,p,
2666
+ // and for the word loop of a struct copy at every level. Decoded as exactly that: the
2667
+ // `addi rA,rA,D` this frontend already lifts, then the plain access at `0(rA)`. So an `@l` displacement (`lhzu r0,gH@l(r4)`, the
2668
+ // FIRST access of a run over a `u16`/`u32` table or a store run) completes its `lis` the way
2669
+ // `addi rD,rHi,gH@l` does, and leaves `&gH` in rA; a base holding a frame address moves as a
2670
+ // capture moved by `addi` does, and the frame-object audit judges what it then reaches.
2671
+ //
2672
+ // It refuses where that reading is not the instruction's:
2673
+ // • a base field of 0 (`r0`) is the literal 0, which the ISA makes an invalid form for every
2674
+ // update form, and so is a load into its own base (rD == rA): neither has one meaning;
2675
+ // • a store of its own base (rS == rA) stores the value BEFORE the update, which an
2676
+ // `addi`-first decode would not;
2677
+ // • a base of r1 moves the stack pointer, which `r1Displacements` tracks only for the frame
2678
+ // push (`case 'stwu'`) and `addi r1,r1,N`, and so does a load into r1;
2679
+ // • a relocation other than `R_PPC_ADDR16_LO` (no compiler emits one).
2680
+ // Not decoded, and refused as unmodelled (`emitOpaqueDest` refuses the loads, which write a
2681
+ // second register an opaque cannot stand for; `opaqueDest` the stores): the INDEXED update forms
2682
+ // (`lwzux`, `stbux`, …), which in 2004 functions with an update form across pikmin,
2683
+ // ac-decomp and marioparty4 appear only in hand-written asm, and the floating-point ones
2684
+ // (`lfsu`, `stfdu`, …), whose plain forms this frontend does not lift either.
2685
+ const withUpdate = (plain: string, store: boolean) => {
2686
+ const [r, mem] = ins.ops;
2687
+ const { off, base } = parseMem(mem ?? '');
2688
+ const why =
2689
+ !isReg(base) || base === 'r0'
2690
+ ? `its base is ${base === 'r0' ? 'r0' : `'${mem}'`}, which the ISA makes an invalid form of an update`
2691
+ : base === 'r1'
2692
+ ? 'it moves the stack pointer, which only the frame push may do here'
2693
+ : !store && r === 'r1'
2694
+ ? 'it loads into the stack pointer, a value this lift does not track'
2695
+ : r === base
2696
+ ? store
2697
+ ? `it stores its own base ${r}, whose stored value is the one before the update`
2698
+ : `it loads into its own base ${r}, which the ISA makes an invalid form`
2699
+ : ins.reloc !== undefined && ins.reloc.type !== 'R_PPC_ADDR16_LO'
2700
+ ? `it carries a '${ins.reloc.type}' relocation, which no update form is decoded under`
2701
+ : undefined;
2702
+ if (why !== undefined) {
2703
+ throw new PpcUnsupportedError(
2704
+ `cannot lift '${name}': '${ins.mnemonic} ${ins.ops.join(',')}' at 0x${ins.addr.toString(16)} is a ` +
2705
+ `${store ? 'store' : 'load'} with update not modelled — ${why}`,
2706
+ );
2707
+ }
2708
+ const advance: Instr = { ...ins, mnemonic: 'addi', ops: [base, base, String(off)] };
2709
+ decodedFrom.set(advance, ins);
2710
+ decode(advance);
2711
+ // The advanced base is the update's side effect, and proves no argument (`ssa.ts`
2712
+ // `noteIncidentalWrite`).
2713
+ if (lastDef !== null) {
2714
+ ssa.noteIncidentalWrite(lastDef);
2715
+ }
2716
+ const access: Instr = { addr: ins.addr, mnemonic: plain, ops: [r, `0(${base})`] };
2717
+ decodedFrom.set(access, ins);
2718
+ decode(access);
2719
+ // The `addi` took the relocation or threw, at its own choke point.
2720
+ relocTaken = true;
2721
+ };
1015
2722
  switch (mnem) {
1016
2723
  case 'nop':
1017
2724
  break;
@@ -1022,24 +2729,41 @@ export function lift(
1022
2729
  // and it has a reaching definition all the same.
1023
2730
  case 'bl': {
1024
2731
  relocTaken = ins.reloc?.type === 'R_PPC_REL24';
1025
- const sym = ins.reloc?.sym ?? 'func';
2732
+ const routine = saveRoutineOf(target, ins);
2733
+ if (routine) {
2734
+ saveRoutine(ins, routine);
2735
+ break;
2736
+ }
2737
+ // A `bl` no relocation names is a call the assembler resolved to an offset in this
2738
+ // section — into this function's own body, or to a neighbour — and has no name to call.
2739
+ if (ins.reloc?.type !== 'R_PPC_REL24') {
2740
+ throw new PpcUnsupportedError(
2741
+ `cannot lift '${name}': the 'bl' at 0x${ins.addr.toString(16)} has no callee symbol — no ` +
2742
+ `R_PPC_REL24 relocation names it, so it calls an offset in this section, which has no name to call it by`,
2743
+ );
2744
+ }
2745
+ const sym = ins.reloc.sym;
1026
2746
  calls.lower({
1027
2747
  callee: sym,
1028
2748
  ssa,
1029
2749
  bi,
1030
2750
  read,
1031
2751
  write,
1032
- // a pending `@ha` half is not an argument
1033
- highHalves,
2752
+ // a pending `@ha` half and a table base are no argument
2753
+ placeholders,
2754
+ parked,
2755
+ scaledIndex,
1034
2756
  guess: {
1035
2757
  // a gap in the argument registers refuses, naming this address (`fallbackArgc`)
1036
2758
  at: ins.addr,
1037
2759
  // A GUESS THAT FILLS EVERY ARGUMENT REGISTER CANNOT SAY WHERE THE LIST ENDS. The ninth
1038
2760
  // argument travels in the callee's parameter area, 8 bytes above the pushed r1 (past the
1039
2761
  // back chain and the LR save word), and a value stored there that reaches the call is that
1040
- // argument or a local the compiler put in the same words. Reading the word back, before or
1041
- // after the call, does not decide it, so the end-of-lift check for a store nothing reloads
1042
- // is not enough here. Without a prototype this refuses; a declared arity says which.
2762
+ // argument or a local the compiler put in the same words. Reading THAT word back does not
2763
+ // decide it on a compiler that may load its outgoing area, so the end-of-lift check for a
2764
+ // store nothing reloads is not enough here. Without a prototype this refuses; a declared
2765
+ // arity says which — and so does the loaded-word bound: mwcc loads no word of its
2766
+ // parameter area, so a word at or above the lowest one it loads is a local's.
1043
2767
  //
1044
2768
  // A guess of seven is the same question for one shape. A 64-bit argument takes an aligned
1045
2769
  // register pair, so one that follows seven words skips r10 and fills the first two words
@@ -1050,15 +2774,32 @@ export function lift(
1050
2774
  if (argc < ARG_REGS.length - 1) {
1051
2775
  return;
1052
2776
  }
2777
+ // The same question decides where the local area starts: past the widest call's
2778
+ // stack words, so a guess that cannot say whether this call takes any cannot draw it
2779
+ // — save where the loaded-word bound reaches the area's start, so that no word of the
2780
+ // area is any call's stack argument (`measurePpcFrame`'s `outgoingEnd`).
2781
+ if (frame !== null && (r1At.get(ins) !== -frame.size || frame.outgoingEnd > frame.area.from)) {
2782
+ fail(
2783
+ `cannot lift '${name}': outgoing stack arguments not modelled — nothing declares '${sym}', called at ` +
2784
+ `0x${ins.addr.toString(16)} with ${argc} argument registers filled, so whether it takes stack words ` +
2785
+ `is unknown, and this compiler's local area starts past them`,
2786
+ );
2787
+ }
1053
2788
  const depth = r1At.get(ins);
1054
- const reaching = (off: number) => valueSlots.has(off) && ssa.hasReachingDef(stackSlotKey(off), bi);
1055
- const inParamArea = (off: number) => typeof depth !== 'number' || (off >= depth + 8 && off < 0);
2789
+ const reaching = (off: number) =>
2790
+ valueSlots.has(off) && ssa.hasReachingDef(stackSlotKey(off), bi, (v) => !nonValues.has(v));
2791
+ // …save a word at or above the loaded-word bound, which is a local's
2792
+ const inParamArea = (off: number) =>
2793
+ typeof depth !== 'number' ||
2794
+ (off >= depth + PARAM_AREA_AT && off < 0 && off + 4 <= loadedWordBound(depth));
1056
2795
  const word = [...valueSlots.keys()].find((off) => inParamArea(off) && reaching(off));
1057
2796
  const passed =
1058
2797
  argc === ARG_REGS.length || typeof depth !== 'number'
1059
2798
  ? word
1060
- : reaching(depth + 8) && reaching(depth + 12)
1061
- ? depth + 8
2799
+ : reaching(depth + PARAM_AREA_AT) &&
2800
+ reaching(depth + PARAM_AREA_AT + 4) &&
2801
+ inParamArea(depth + PARAM_AREA_AT + 4)
2802
+ ? depth + PARAM_AREA_AT
1062
2803
  : undefined;
1063
2804
  if (passed !== undefined) {
1064
2805
  const slot = valueSlots.get(passed)!;
@@ -1081,11 +2822,11 @@ export function lift(
1081
2822
  break;
1082
2823
  }
1083
2824
  // Stack-frame + link-register bookkeeping. `stwu r1,-N(r1)` / `addi r1,r1,N` adjust the frame
1084
- // pointer; `mflr` copies the return address out (`returnAddresses`) and `mtlr` puts it back.
2825
+ // pointer; `mflr` copies the return address out (a `return-address` in `nonValues`) and `mtlr`
2826
+ // puts it back.
1085
2827
  // A word store or load on r1 is a save or a value slot (`frameStore`/`frameLoad`); `stmw`
1086
2828
  // records save slots.
1087
- // The GENERAL form `stwu rS,D(rA)` (base ≠ r1) is a real store-with-BASE-UPDATE
1088
- // (`*(rA+D)=rS; rA+=D`) — neither effect is modelled here, so loud-fail rather than drop both.
2829
+ // The GENERAL form `stwu rS,D(rA)` (base ≠ r1) is a store with update (`withUpdate`).
1089
2830
  // The push itself must be the only one, taken from the entry r1: every slot is named by its
1090
2831
  // offset from there (`r1Displacements`), and a push from anywhere else moves that origin.
1091
2832
  case 'stwu':
@@ -1104,12 +2845,12 @@ export function lift(
1104
2845
  saveSlots.set(parseMem(s).off, 'r1');
1105
2846
  break;
1106
2847
  }
1107
- throw new PpcUnsupportedError(
1108
- `cannot lift '${name}': stwu with update on ${parseMem(s).base} (store-with-base-update) not modelled`,
1109
- );
2848
+ withUpdate('stw', true);
2849
+ break;
1110
2850
  case 'mflr': {
2851
+ refuseOverwrittenLink(ins);
1111
2852
  const lr = mkValue(T.unk(32));
1112
- returnAddresses.set(lr, ins.addr);
2853
+ nonValues.set(lr, { kind: 'return-address', addr: ins.addr });
1113
2854
  write(d, lr);
1114
2855
  break;
1115
2856
  }
@@ -1124,9 +2865,16 @@ export function lift(
1124
2865
  break;
1125
2866
  case 'stmw':
1126
2867
  if (parseMem(s).base === 'r1') {
1127
- for (const slot of frameRange(d, entryOffset(ins, s))) {
1128
- refuseMixedSlot(ins, s, slot.off, 'save');
1129
- saveSlots.set(slot.off, slot.reg);
2868
+ if (inLocalArea(ins, s, 4 * (32 - Number(d.slice(1)))) !== undefined) {
2869
+ fail(
2870
+ `cannot lift '${name}': address-taken stack local — the 'stmw' at 0x${ins.addr.toString(16)} saves ` +
2871
+ `into the local area, a frame this does not describe`,
2872
+ );
2873
+ }
2874
+ const off = entryOffset(ins, s);
2875
+ const depth = off - parseMem(s).off;
2876
+ for (const slot of frameRange(d, off)) {
2877
+ saveInto(ins, slot.reg, slot.off, `${slot.off - depth}(r1)`);
1130
2878
  }
1131
2879
  break;
1132
2880
  }
@@ -1137,7 +2885,14 @@ export function lift(
1137
2885
  case 'lmw':
1138
2886
  if (parseMem(s).base === 'r1') {
1139
2887
  for (const slot of frameRange(d, parseMem(s).off)) {
1140
- frameLoad(ins, slot.reg, `${slot.off}(r1)`);
2888
+ const word = `${slot.off}(r1)`;
2889
+ if (inLocalArea(ins, word, 4) !== undefined) {
2890
+ fail(
2891
+ `cannot lift '${name}': address-taken stack local — the 'lmw' at 0x${ins.addr.toString(16)} ` +
2892
+ `reloads ${slot.reg} out of '${word}', inside the local area, a frame this does not describe`,
2893
+ );
2894
+ }
2895
+ frameLoad(ins, slot.reg, word);
1141
2896
  }
1142
2897
  break;
1143
2898
  }
@@ -1179,9 +2934,14 @@ export function lift(
1179
2934
  `is '${s}', not the expected 0 placeholder`,
1180
2935
  );
1181
2936
  }
2937
+ // A jump table's own name is read as data and never written down, so the naming policy
2938
+ // does not ask it; RELA pairs a half only with a low half of the same symbol and addend,
2939
+ // and that one, with no immediate of its own, makes a table base (`foldTableBase`).
2940
+ const table = tableBases.wants({ sym: ins.reloc.sym, addend: ins.reloc.addend });
2941
+ relocTaken = true;
1182
2942
  const hi = mkValue(T.unk(32));
1183
2943
  highHalves.record(hi, {
1184
- sym: spellableSym(ins),
2944
+ sym: table ? ins.reloc.sym : spellableSym(ins),
1185
2945
  addend: ins.reloc.addend,
1186
2946
  addr: ins.addr,
1187
2947
  mnemonic: ins.mnemonic,
@@ -1201,7 +2961,9 @@ export function lift(
1201
2961
  // The STACK POINTER first, before any fold looks at the operands. `addi r1,r1,N` is frame
1202
2962
  // teardown and is skipped; a relocation on it is no known compiler's output, and folding
1203
2963
  // one would write an address into r1 and let the function lift as if the teardown had
1204
- // happened — silent, and about the frame rather than the value.
2964
+ // happened — silent, and about the frame rather than the value. Only `addi r1,r1,N` is the
2965
+ // teardown: `addi r1,r3,N` sets r1 to another register's value, and falls through to
2966
+ // `write`, which refuses it like any other write to the stack pointer.
1205
2967
  if (d === 'r1') {
1206
2968
  if (ins.reloc) {
1207
2969
  throw new PpcUnsupportedError(
@@ -1209,15 +2971,59 @@ export function lift(
1209
2971
  `adjust — an address built in r1 is not a shape this frontend models`,
1210
2972
  );
1211
2973
  }
2974
+ if (s === 'r1') {
2975
+ break;
2976
+ }
2977
+ }
2978
+ // A capture, `addi rD,r1,k`, is the `laddr` of the local area's offset k, where the frame
2979
+ // has one (`measurePpcFrame`); without one it reaches `read`'s r1-as-data refusal. A capture
2980
+ // moved by a constant is the capture at the sum.
2981
+ if (frame !== null && takesFrameAddress(ins, target.registerSaveRoutines?.frameReg)) {
2982
+ write(d, emitLaddr(parseImm(t) + (r1At.get(ins) as number) + frame.size));
2983
+ break;
2984
+ }
2985
+ if (
2986
+ frame !== null &&
2987
+ mnem === 'addi' &&
2988
+ !ins.reloc &&
2989
+ isReg(s) &&
2990
+ s !== 'r0' &&
2991
+ s !== 'r1' &&
2992
+ !SMALL_DATA_BASES.includes(s)
2993
+ ) {
2994
+ const raw = readVar(s, bi);
2995
+ const at = laddrAt.get(raw);
2996
+ if (at !== undefined) {
2997
+ movedCaptures.add(raw);
2998
+ write(d, emitLaddr(at + parseImm(t)));
2999
+ break;
3000
+ }
3001
+ }
3002
+ if (mnem === 'addi' && d === target.registerSaveRoutines?.frameReg && s === 'r1' && !ins.reloc) {
3003
+ const v = mkValue(T.unk(32));
3004
+ const depth = r1At.get(ins);
3005
+ nonValues.set(v, {
3006
+ kind: 'frame-address',
3007
+ off: typeof depth === 'number' ? depth + parseImm(t) : (depth ?? 'the instruction is unreachable'),
3008
+ addr: ins.addr,
3009
+ });
3010
+ write(d, v);
1212
3011
  break;
1213
3012
  }
1214
3013
  // `addi rD,rHi,SYM@l` completes the address a `lis` began: rD = &SYM (+ the relocation's
1215
- // addend). It is the only `@l` consumer modelled here. The others — `ori`, `addic`, and
1216
- // the `@l`-in-a-displacement form (1,935 integer load/store sites across the three
1217
- // checkouts, plus the float loads) — are left unbuilt per docs/level-tower.md's "earn the
1218
- // level": no row in this benchmark reaches them. Each refuses at its own guard, and a
1219
- // high half none of them consumed is caught by `readReg` or by the dangling-`@ha` check.
3014
+ // addend). It is the only `@l` consumer modelled here, and an update form's `@l`
3015
+ // displacement reaches it as the `addi` that advances its base (`withUpdate`). The others —
3016
+ // `ori`, `addic`, and the `@l`-in-a-displacement form of a PLAIN load or store (1,935
3017
+ // integer load/store sites across the three checkouts, plus the float loads) — are left
3018
+ // unbuilt per docs/level-tower.md's "earn the level": no row in this benchmark reaches
3019
+ // them. Each refuses at its own guard, and a high half none of them consumed is caught
3020
+ // by `readReg` or by the dangling-`@ha` check.
1220
3021
  if (mnem === 'addi' && ins.reloc?.type === 'R_PPC_ADDR16_LO') {
3022
+ const base = parseImm(t) === 0 ? foldTableBase(ins, s) : null;
3023
+ if (base !== null) {
3024
+ write(d, base);
3025
+ break;
3026
+ }
1221
3027
  const g = foldLoHalf(ins, s, t);
1222
3028
  if (ins.reloc.addend === 0) {
1223
3029
  write(d, g);
@@ -1240,7 +3046,23 @@ export function lift(
1240
3046
  write(d, read(s));
1241
3047
  break;
1242
3048
  }
1243
- emitBin('add', d, read(s), constVal(parseImm(t)));
3049
+ {
3050
+ const a = read(s);
3051
+ const k = parseImm(t);
3052
+ emitBin('add', d, a, constVal(k));
3053
+ if (mnem === 'addic') {
3054
+ // a + K carries out exactly when a >=u 2^32 - K; K == 0 never carries.
3055
+ setCarry(
3056
+ () =>
3057
+ k === 0
3058
+ ? constVal(0)
3059
+ : k === -1
3060
+ ? kit.tmp('icmp_ne', [a, constVal(0)])
3061
+ : kit.tmp('icmp_uge', [a, constVal((2 ** 32 - (k >>> 0)) | 0)]),
3062
+ true,
3063
+ );
3064
+ }
3065
+ }
1244
3066
  break;
1245
3067
  // add immediate SHIFTED — the register-based `%ha` anchor: mwcc derives an absolute base
1246
3068
  // from a scaled index (`addis r4,r3,-32736` = r3 + 0x80200000). The jump-table lis/addi
@@ -1255,16 +3077,107 @@ export function lift(
1255
3077
  emitBin('add', d, read(s), constVal((parseImm(t) << 16) >> 0));
1256
3078
  break;
1257
3079
  case 'subf':
1258
- case 'subfc':
1259
3080
  case 'subfo':
1260
- emitBin('sub', d, read(t), read(s));
3081
+ case 'subfc':
3082
+ case 'subfco': {
3083
+ const b = read(t);
3084
+ const a = read(s);
3085
+ const diff = emitBin('sub', d, b, a);
3086
+ if (mnem.startsWith('subfc')) {
3087
+ setCarry(() => {
3088
+ const c = kit.tmp('icmp_uge', [b, a]);
3089
+ carrySource.set(c, diff);
3090
+ return c;
3091
+ });
3092
+ }
1261
3093
  break; // rD = rB - rA (reversed)
1262
- case 'subfic':
1263
- emitBin('sub', d, constVal(parseImm(t)), read(s));
3094
+ }
3095
+ case 'subfic': {
3096
+ const k = constVal(parseImm(t));
3097
+ const a = read(s);
3098
+ const diff = emitBin('sub', d, k, a);
3099
+ setCarry(() => {
3100
+ const c = parseImm(t) === 0 ? kit.tmp('icmp_eq', [a, constVal(0)]) : kit.tmp('icmp_uge', [k, a]);
3101
+ carrySource.set(c, diff);
3102
+ return c;
3103
+ });
1264
3104
  break; // rD = imm - rA
1265
- case 'neg':
1266
- emitUn('neg', d, read(s));
3105
+ }
3106
+ // The carry READERS, each over the carry its block's writer left (`readCarry`):
3107
+ // subfe rD,rA,rB = ~rA + rB + CA adde rD,rA,rB = rA + rB + CA
3108
+ // subfze rD,rA = ~rA + CA addze rD,rA = rA + CA
3109
+ // Two `subfe` shapes are modelled, both of whose register halves cancel: `subfe rD,rY,rY`
3110
+ // (~y + y = -1, so CA - 1, reading no rY — mwcc writes it over a register holding anything,
3111
+ // the return address included) and `subfe rD,rA,rB` with rA = rB - 1 (~(x-1) + x = 0, so
3112
+ // CA). Any other `subfe` is the high half of a 64-bit subtract, and `subfze` over a register
3113
+ // that is no constant is a 64-bit negate: both stay unmodelled, as do `addme`/`subfme`.
3114
+ case 'subfe': {
3115
+ if (s === t) {
3116
+ const c = readCarry(ins);
3117
+ const r = emitBin('add', d, c, constVal(-1));
3118
+ const diff = carrySource.get(c);
3119
+ if (diff !== undefined) {
3120
+ truthPending.set(r, { diff, via: 'neg' }); // CA - 1; its negation is the 0/1 `rA <u rB`
3121
+ }
3122
+ break;
3123
+ }
3124
+ const b = read(t);
3125
+ const def = ops.find((o) => o.results[0] === read(s));
3126
+ if (def?.opcode === 'add' && def.operands[0] === b && constIn(def.operands[1]) === -1) {
3127
+ readWithoutOp.add(def.results[0]);
3128
+ write(d, readCarry(ins));
3129
+ break;
3130
+ }
3131
+ emitOpaqueDest(ins);
3132
+ break;
3133
+ }
3134
+ case 'subfze': {
3135
+ const x = read(s);
3136
+ const k = constIn(x);
3137
+ if (k !== undefined) {
3138
+ readWithoutOp.add(x);
3139
+ const ca = readCarry(ins);
3140
+ ~k === 0 ? write(d, ca) : emitBin('add', d, constVal(~k), ca);
3141
+ break;
3142
+ }
3143
+ // mwcc hoists the constant out of a loop (`li r3,-1` ahead of the loop, `subfze r0,r3` in
3144
+ // its body), so the register comes in through a join. Whether every path brings the same
3145
+ // constant is known only once every block is filled; the exact `~rA + CA` stands until then.
3146
+ // The value may also come down through blocks with one predecessor from the block that set
3147
+ // it, so only a value this block computes is settled here; an entry parameter is the
3148
+ // caller's value, never a constant, and the settling refuses it.
3149
+ if (bi === 0 || ops.some((o) => o.results.includes(x))) {
3150
+ emitOpaqueDest(ins);
3151
+ break;
3152
+ }
3153
+ const ca = readCarry(ins);
3154
+ const inv = mkValue(T.unk(32));
3155
+ const not = mkOp('not', { operands: [x], results: [inv] });
3156
+ ops.push(not);
3157
+ emitBin('add', d, inv, ca);
3158
+ joinedCarryConsts.push({ ins, not, add: ops[ops.length - 1], ca });
3159
+ break;
3160
+ }
3161
+ case 'adde':
3162
+ emitBin('add', d, binTmp('add', read(s), read(t)), readCarry(ins, true));
1267
3163
  break;
3164
+ case 'addze': {
3165
+ const c = readCarry(ins, true);
3166
+ const r = emitBin('add', d, read(s), c);
3167
+ const diff = carrySource.get(c);
3168
+ if (diff !== undefined) {
3169
+ truthPending.set(r, { diff, via: 'and1' }); // mwcc's signed compare keeps bit 0 of this
3170
+ }
3171
+ break;
3172
+ }
3173
+ case 'neg': {
3174
+ const x = read(s);
3175
+ const n = emitUn('neg', d, x);
3176
+ if (truthPending.get(x)?.via === 'neg') {
3177
+ truthDifferences.push({ diff: truthPending.get(x)!.diff, carry: x, truth: n });
3178
+ }
3179
+ break;
3180
+ }
1268
3181
  case 'mullw':
1269
3182
  case 'mullwo':
1270
3183
  emitBin('mul', d, read(s), read(t));
@@ -1358,6 +3271,10 @@ export function lift(
1358
3271
  emitShImm('shl', d, read(s), parseImm(t));
1359
3272
  break;
1360
3273
  case 'srwi':
3274
+ if (crCopies.has(read(s))) {
3275
+ write(d, crBitValue(ins, read(s), parseImm(t) === 31 ? 0 : -1));
3276
+ break;
3277
+ }
1361
3278
  emitShImm('shr_u', d, read(s), parseImm(t));
1362
3279
  break;
1363
3280
  case 'srawi':
@@ -1377,28 +3294,57 @@ export function lift(
1377
3294
  break;
1378
3295
  }
1379
3296
  // rotate-and-mask, rotate 0 only: `clrlwi rD,rS,n` = rS & (~0>>>n); `clrrwi` = rS & (~0<<n).
1380
- case 'clrlwi':
1381
- emitBin('and', d, read(s), constVal((0xffffffff >>> parseImm(t)) >>> 0));
3297
+ case 'clrlwi': {
3298
+ const x = read(s);
3299
+ const m = emitBin('and', d, x, constVal((0xffffffff >>> parseImm(t)) >>> 0));
3300
+ if (parseImm(t) === 31 && truthPending.get(x)?.via === 'and1') {
3301
+ truthDifferences.push({ diff: truthPending.get(x)!.diff, carry: x, truth: m });
3302
+ }
1382
3303
  break;
3304
+ }
1383
3305
  case 'clrrwi':
1384
3306
  emitBin('and', d, read(s), constVal((0xffffffff << parseImm(t)) >>> 0));
1385
3307
  break;
1386
3308
  case 'rlwinm': {
1387
3309
  const [sh, mb, me] = [parseImm(ins.ops[2]), parseImm(ins.ops[3]), parseImm(ins.ops[4])];
3310
+ if (crCopies.has(read(s))) {
3311
+ write(d, crBitValue(ins, read(s), mb === 31 && me === 31 ? (sh + 31) % 32 : -1));
3312
+ break;
3313
+ }
1388
3314
  // `rlwinm rD,rS,SH,MB,ME` = rotl(rS,SH) & mask(MB,ME). Three cases we can lower exactly:
1389
3315
  // • SH==0 — a pure masked AND.
1390
- // • right-shift EXTRACT `(x>>n)&m` — ME==31 and the field does not wrap (SH+MB>=32): the
1391
- // compiler's form for `(x>>n)&mask`, with n=32-SH and mask=mask(MB,31). Emit the shift
1392
- // then the AND — the faithful, idiomatic C that recompiles to this exact rlwinm.
3316
+ // • right-shift EXTRACT `(x>>n)&m` — the field lies in the SH bits the rotate brings round
3317
+ // from the top (MB >= 32-SH, MB <= ME): the compiler's form for `(x>>n)&mask`, with
3318
+ // n=32-SH and mask=mask(MB,ME). Emit the shift then the AND — the faithful, idiomatic C
3319
+ // that recompiles to this exact rlwinm. Where that field is the one bit x's sign bit
3320
+ // lands on (MB == ME == 32-SH) it is `(x >>u 31) << (SH-1)`: mwcc merges a shift of a
3321
+ // 0/1 value into its extract (`(a < b) << 3` ends `rlwinm r3,r0,4,28,28` on 2.4.x), and
3322
+ // the `srwi 31` the branch-free compare folds read is then whole (pattern/engine.ts
3323
+ // MWCC_COMPARE_PATTERNS).
3324
+ // • left-shift INSERT `(x&m)<<SH` — the field lies wholly in the bits the shift fills
3325
+ // (MB <= ME <= 31-SH): `(x & (mask >> SH)) << SH`, mwcc's form for a masked value moved
3326
+ // up (`(a < b) << 5` over 2.3.3's `& 1` ends `rlwinm r3,r0,5,26,26`).
1393
3327
  // • anything else (a genuine rotate / bitfield insert) stays an opaque — loud, not dropped.
1394
3328
  if (sh === 0) {
1395
3329
  emitBin('and', d, read(s), constVal(rlwinmMask(mb, me)));
1396
3330
  break;
1397
3331
  }
1398
- if (me === 31 && sh + mb >= 32) {
3332
+ if (mb === me && mb === 32 - sh && sh > 1) {
3333
+ const sign = mkValue(T.unk(32));
3334
+ ops.push(mkOp('shr_u', { operands: [read(s)], results: [sign], attrs: { imm: 31 } }));
3335
+ emitShImm('shl', d, sign, sh - 1);
3336
+ break;
3337
+ }
3338
+ if (mb <= me && me <= 31 - sh) {
3339
+ const masked = mkValue(T.unk(32));
3340
+ ops.push(mkOp('and', { operands: [read(s), constVal(rlwinmMask(mb, me) >>> sh)], results: [masked] }));
3341
+ emitShImm('shl', d, masked, sh);
3342
+ break;
3343
+ }
3344
+ if (mb <= me && sh + mb >= 32) {
1399
3345
  const shifted = mkValue(T.unk(32));
1400
3346
  ops.push(mkOp('shr_u', { operands: [read(s)], results: [shifted], attrs: { imm: (32 - sh) & 31 } }));
1401
- emitBin('and', d, shifted, constVal(rlwinmMask(mb, 31)));
3347
+ emitBin('and', d, shifted, constVal(rlwinmMask(mb, me)));
1402
3348
  break;
1403
3349
  }
1404
3350
  emitOpaqueDest(ins);
@@ -1433,6 +3379,17 @@ export function lift(
1433
3379
  case 'stb':
1434
3380
  emitStore(ins, d, s, 1);
1435
3381
  break;
3382
+ // The update forms (`withUpdate`); `stwu` is decoded with the frame push above.
3383
+ case 'lwzu':
3384
+ case 'lhzu':
3385
+ case 'lhau':
3386
+ case 'lbzu':
3387
+ withUpdate(mnem.slice(0, -1), false);
3388
+ break;
3389
+ case 'sthu':
3390
+ case 'stbu':
3391
+ withUpdate(mnem.slice(0, -1), true);
3392
+ break;
1436
3393
  // Register+register indexed forms (variable-index array access). Widths/signedness mirror the
1437
3394
  // displacement loads/stores above; `lhax` is the sign-extending halfword (algebraic).
1438
3395
  case 'lwzx':
@@ -1482,6 +3439,92 @@ export function lift(
1482
3439
  recordCmp(c.cr, read(c.lhsTok), rhs, false);
1483
3440
  break;
1484
3441
  }
3442
+ // A FLOAT COMPARE sets its field's three bits from the IEEE relation of its operands; a NaN
3443
+ // clears all three (and sets `so`). Which compare each bit holds is what a branch, a `cror`
3444
+ // or an extract then reads.
3445
+ case 'fcmpo':
3446
+ case 'fcmpu': {
3447
+ const [cr, a, bReg] = ins.ops;
3448
+ if (fpu === undefined || !/^cr[0-7]$/.test(cr ?? '') || ![a, bReg].every(isFpKey)) {
3449
+ emitOpaqueDest(ins);
3450
+ break;
3451
+ }
3452
+ refuseCrossingUnmodelled(ins);
3453
+ const lhs = readFloat(ins, a);
3454
+ const rhs = readFloat(ins, bReg);
3455
+ cmpDef.set(cr, {
3456
+ lhs,
3457
+ rhs,
3458
+ signed: true,
3459
+ bits: { lt: 'fcmp_olt', gt: 'fcmp_ogt', eq: 'fcmp_oeq' },
3460
+ quiet: mnem === 'fcmpu',
3461
+ });
3462
+ break;
3463
+ }
3464
+ // `cror bD,bA,bB` ORs two bits into a third. mwcc's `<=` and `>=` over floats set `eq` to
3465
+ // `lt | eq` or `gt | eq` of one field, which is the compare whose truth table is the OR of
3466
+ // the two (`FCMP_TRUTH`). Bits of two fields, or of an integer compare, are not modelled.
3467
+ case 'cror': {
3468
+ const [dst, x, y] = ins.ops.map(crBitOf);
3469
+ const field = dst?.field;
3470
+ const held = field === undefined ? undefined : cmpDef.get(field);
3471
+ const bits = typeof held === 'object' ? held.bits : undefined;
3472
+ if (!dst || !x || !y || x.field !== field || y.field !== field || held === undefined || !bits) {
3473
+ throw new PpcUnsupportedError(
3474
+ `cannot lift '${name}': 'cror' at 0x${ins.addr.toString(16)} — only an OR of two bits of the field a ` +
3475
+ 'float compare set, into that field, is modelled',
3476
+ );
3477
+ }
3478
+ const truth = FCMP_TRUTH[bits[x.bit]].map((t, i) => t || FCMP_TRUTH[bits[y.bit]][i]);
3479
+ const opcode = fcmpOfTruth(truth);
3480
+ if (opcode === undefined) {
3481
+ throw new PpcUnsupportedError(
3482
+ `cannot lift '${name}': 'cror' at 0x${ins.addr.toString(16)} ORs two bits into no compare of the operands`,
3483
+ );
3484
+ }
3485
+ cmpDef.set(field!, { ...(held as CrCompare), bits: { ...bits, [dst.bit]: opcode } });
3486
+ break;
3487
+ }
3488
+ case 'mfcr': {
3489
+ const copy = mkValue(T.unk(32));
3490
+ crCopies.set(copy, new Map(cmpDef));
3491
+ write(d, copy);
3492
+ break;
3493
+ }
3494
+ // `fctiwz fD,fS` truncates toward zero into the integer word of fD's low half, which reaches
3495
+ // the integer file only through memory: `stfd fD,N(r1)` and `lwz rT,N+4(r1)` (the low word,
3496
+ // big-endian). So fD holds that integer as a WORD, and its `stfd` stores it to the frame word
3497
+ // a reload reads; its high half is no value (`fctiwz-high`), so a use of its reload refuses.
3498
+ case 'fctiwz': {
3499
+ if (fpu === undefined || ins.ops.length !== 2 || !ins.ops.every(isFpKey)) {
3500
+ emitOpaqueDest(ins);
3501
+ break;
3502
+ }
3503
+ refuseCrossingUnmodelled(ins);
3504
+ const v = mkValue(T.unk(32));
3505
+ ops.push(mkOp('fptosi', { operands: [readFloat(ins, s)], results: [v] }));
3506
+ writeVar(d, bi, v);
3507
+ break;
3508
+ }
3509
+ case 'stfd': {
3510
+ const held = isFpKey(d) && fpu !== undefined ? readReg(d, bi) : undefined;
3511
+ if (held === undefined || held.type.kind === 'float' || parseMem(s).base !== 'r1') {
3512
+ emitOpaqueDest(ins);
3513
+ break;
3514
+ }
3515
+ const off = entryOffset(ins, s) + 4;
3516
+ refuseMixedSlot(ins, s, off - 4, 'value');
3517
+ refuseMixedSlot(ins, s, off, 'value');
3518
+ const high = mkValue(T.unk(32));
3519
+ nonValues.set(high, { kind: 'fctiwz-high', addr: ins.addr });
3520
+ writeVar(stackSlotKey(off - 4), bi, high);
3521
+ writeVar(stackSlotKey(off), bi, held);
3522
+ if (!valueSlots.has(off)) {
3523
+ valueSlots.set(off, { addr: ins.addr, mem: s, read: false, stored: [] });
3524
+ }
3525
+ valueSlots.get(off)!.stored.push({ v: held, at: ins.addr });
3526
+ break;
3527
+ }
1485
3528
  case 'fadds':
1486
3529
  case 'fsubs':
1487
3530
  case 'fmuls':
@@ -1509,11 +3552,11 @@ export function lift(
1509
3552
  }
1510
3553
  const [dst, ...srcs] = ins.ops;
1511
3554
  if (op === 'copy') {
1512
- writeVar(dst, bi, read(srcs[0]));
3555
+ writeVar(dst, bi, readFloat(ins, srcs[0]));
1513
3556
  break;
1514
3557
  }
1515
3558
  const v = mkValue(fpType);
1516
- ops.push(mkOp(op, { operands: srcs.map(read), results: [v] }));
3559
+ ops.push(mkOp(op, { operands: srcs.map((k) => readFloat(ins, k)), results: [v] }));
1517
3560
  writeVar(dst, bi, v);
1518
3561
  break;
1519
3562
  }
@@ -1521,10 +3564,14 @@ export function lift(
1521
3564
  emitOpaqueDest(ins);
1522
3565
  break; // unmodelled: an honest opaque, never a silent drop
1523
3566
  }
3567
+ if (CARRY_READERS.test(mnem)) {
3568
+ clobberCarry('set it as the carry out of an extended add or subtract, which is not modelled');
3569
+ }
1524
3570
  // A record-form op sets cr0 from a signed compare of its result against 0. Wire that so the
1525
3571
  // next branch reading cr0 fuses (`andi. r0,r3,1; beq …` → `if ((a0 & 1) == 0)`).
1526
3572
  if (rc && lastDef) {
1527
3573
  cmpDef.set('cr0', { lhs: lastDef, rhs: constVal(0), signed: true });
3574
+ compared.add(lastDef);
1528
3575
  } else if (rc) {
1529
3576
  cmpDef.set(
1530
3577
  'cr0',
@@ -1553,7 +3600,16 @@ export function lift(
1553
3600
  // blocks (dense 0..N-1) then the default (last successor). The `cmplwi`/`bgt` and the elided
1554
3601
  // dispatch (`lis…lwzx;mtctr;bctr`) are subsumed.
1555
3602
  const jt = br ? jts.get(br.addr) : undefined;
3603
+ // Asked before the jump-table lowering: a table whose bound is a conditional return
3604
+ // (`cmplwi rS,N; bgtlr`) lowers that return as the switch's default edge, and it jumps through
3605
+ // whatever the link register holds there exactly as a plain `bgtlr` does.
3606
+ if (br && (isReturn(br) || isCondReturn(br))) {
3607
+ refuseOverwrittenLink(br);
3608
+ }
1556
3609
  if (jt) {
3610
+ if (jt.hoisted) {
3611
+ tableBases.demand(jt.hoisted, bi, jt.bctrAddr);
3612
+ }
1557
3613
  pushSwitchBr(
1558
3614
  ops,
1559
3615
  readReg(jt.scrutReg, bi),
@@ -1624,7 +3680,16 @@ export function lift(
1624
3680
  );
1625
3681
  }
1626
3682
  const table = !cmp.signed ? COND_UNSIGNED : COND_SIGNED;
1627
- const cond = mkCmp(ops, table[base], cmp.lhs, cmp.rhs);
3683
+ const { bit, set } = BRANCH_BIT[base];
3684
+ if (cmp.bits && cmp.quiet && bit !== 'eq') {
3685
+ throw new PpcUnsupportedError(
3686
+ `cannot lift '${name}': conditional branch '${base}' orders the operands of an 'fcmpu', which C emits ` +
3687
+ 'for == and != alone',
3688
+ );
3689
+ }
3690
+ const cond = cmp.bits
3691
+ ? mkCmp(ops, set ? cmp.bits[bit] : NEGATED_FCMP[cmp.bits[bit]], cmp.lhs, cmp.rhs)
3692
+ : mkCmp(ops, table[base], cmp.lhs, cmp.rhs);
1628
3693
  ops.push(mkOp('cond_br', { operands: [cond], successors: succIdx[bi].map((j) => succ(j)) }));
1629
3694
  return;
1630
3695
  }
@@ -1633,7 +3698,17 @@ export function lift(
1633
3698
  ops.push(mkOp('br', { attrs: { fallthrough: true }, successors: [succ(succIdx[bi][0])] }));
1634
3699
  return;
1635
3700
  }
1636
- ops.push(mkOp('ret', { operands: returnOperands(bi) }));
3701
+ // A LAST block with no terminator runs off the end of the text: the input stops before the
3702
+ // function does (a dump split at a foreign label, a hand-supplied fragment), and what
3703
+ // follows is unknown. Lifting it as a return invents one. Thumb refuses the same way.
3704
+ if (!br) {
3705
+ const last = b.body[b.body.length - 1];
3706
+ fail(
3707
+ `cannot lift '${name}': control falls off the end (the block at 0x${b.startAddr.toString(16)} has no ` +
3708
+ `terminator and no successor${last ? `; its last instruction is at 0x${last.addr.toString(16)}` : ''})`,
3709
+ );
3710
+ }
3711
+ emitReturn(ops, bi);
1637
3712
  return;
1638
3713
  }
1639
3714
  // unconditional branch
@@ -1644,50 +3719,454 @@ export function lift(
1644
3719
  fillBlock(b, bi);
1645
3720
  ssa.markFilled(bi);
1646
3721
  });
3722
+ refuseSavedRegisterLeaks(
3723
+ name,
3724
+ blocks.map((b) => ({
3725
+ instrs: [...b.body, ...(b.branch ? [b.branch] : [])],
3726
+ returns: b.synthReturn === true || (b.branch !== null && (isReturn(b.branch) || isCondReturn(b.branch))),
3727
+ })),
3728
+ preds,
3729
+ new Set([...saveSlots.values()].filter((r) => isCalleeSaved(target, r))),
3730
+ // a call preserves every callee-saved register, and a restore routine is a restore
3731
+ (ins, r) => !ins.mnemonic.startsWith('b') && mayWriteGpr(ins, r),
3732
+ restoredBy,
3733
+ );
3734
+ returnUntouchedArg();
3735
+ // What f1 holds at each return, and, where the home is undecided and f1 holds a float this function
3736
+ // computed on every path into every return, that float — read now, before the SSA is finished.
3737
+ const homes = fpu ? returns.map(({ bi }) => floatHomeAt(ssa, fpu.returnReg, bi)) : [];
3738
+ const carriesFloat = carriesFloatHome(homeRule, homes);
3739
+ if (carriesFloat) {
3740
+ returns.forEach(({ op, bi }) => op.operands.push(readReg(fpu!.returnReg, bi)));
3741
+ }
1647
3742
  // A high half no `@l` ever completed — every `@l` consumer this frontend does not model (`ori`, a
1648
3743
  // float load's displacement) lands here, which is what keeps "not modelled" from becoming "not
1649
3744
  // emitted".
1650
3745
  highHalves.assertAllConsumed(name);
3746
+ tableBases.prove(name, ssa);
3747
+ refuseStaleIdiomReads(name, irBlocks, staleIdiom, ssa, prototypes[name] ?? contextOwnProto?.proto, ppcRefusal, calls);
1651
3748
  // A value stored to the frame that nothing reloads is either a home the body never reads or an
1652
3749
  // OUTGOING stack argument (`stw r4,8(r1)` before a nine-argument `bl` passes the ninth in the
1653
3750
  // callee's parameter area, as ac-decomp's `aQMgr_order_decide_trade_*` do). The two are the same
1654
- // store, and dropping the second loses an argument.
1655
- for (const { addr, mem, read } of valueSlots.values()) {
1656
- if (!read) {
3751
+ // store, and dropping the second loses an argument — save a word at or above the loaded-word
3752
+ // bound of the frame it was stored in, which is no call's argument and so a dead local's.
3753
+ //
3754
+ // A STRUCT'S HOME is the other unread word that is no argument. mwcc 2.4.2 and 2.4.7 give a struct
3755
+ // a callee returned in r3 (`registerStructLocal`, call-plan.ts) a home in the frame and store the
3756
+ // register there whole even where every member read is taken off r3 itself: `s = mk1(x); gW =
3757
+ // s.x;` is `bl mk1; stw r3,gW; ...; stw r3,8(r1)`, and nothing reloads the home. A word every
3758
+ // store of which is such an image holds that struct, and is no call's stack argument where every
3759
+ // call that can run after one of those stores — later in its block, or in a block reachable from
3760
+ // it — is declared to take none: then it is the local the call assigns, which
3761
+ // `liftRegisterStructCalls` (frontend/register-struct-call.ts) declares. A call before every store
3762
+ // reads the word before it is written, so it cannot be handed the struct there.
3763
+ const imageCalls = new Map(
3764
+ irBlocks.flatMap((b) =>
3765
+ b.ops.filter((op) => op.attrs[REGISTER_STRUCT] !== undefined).map((op) => [op.results[0], op] as const),
3766
+ ),
3767
+ );
3768
+ const images = new Set(imageCalls.keys());
3769
+ // a struct with a home is one the source names (ir/core.ts `STRUCT_HOMED`)
3770
+ for (const { stored } of valueSlots.values()) {
3771
+ for (const { v } of stored) {
3772
+ const call = imageCalls.get(v);
3773
+ if (call !== undefined) {
3774
+ call.attrs = { ...call.attrs, [REGISTER_STRUCT]: STRUCT_HOMED };
3775
+ }
3776
+ }
3777
+ }
3778
+ const takesNoStackWord = (ins: Instr): boolean => {
3779
+ if (ins.mnemonic === 'bctrl' || ins.mnemonic === 'blrl') {
3780
+ return false; // nothing declares what an indirect call takes
3781
+ }
3782
+ if (ins.mnemonic !== 'bl' || saveRoutineOf(target, ins)) {
3783
+ return true;
3784
+ }
3785
+ const declared = calls.declaredCall(ins.reloc?.sym ?? '');
3786
+ return declared !== null && wordsOf(declared.widths, target) <= ARG_REGS.length;
3787
+ };
3788
+ const blockOfAddr = new Map(blocks.flatMap((b, bi) => b.body.map((ins, k) => [ins.addr, { bi, k }] as const)));
3789
+ const noCallAfterTakesStack = (addr: number): boolean => {
3790
+ const at = blockOfAddr.get(addr);
3791
+ if (at === undefined) {
3792
+ return false;
3793
+ }
3794
+ if (!blocks[at.bi].body.slice(at.k + 1).every(takesNoStackWord)) {
3795
+ return false;
3796
+ }
3797
+ const seen = new Set<number>();
3798
+ const work = [...succIdx[at.bi]];
3799
+ while (work.length > 0) {
3800
+ const bi = work.pop()!;
3801
+ if (seen.has(bi)) {
3802
+ continue;
3803
+ }
3804
+ seen.add(bi);
3805
+ if (!blocks[bi].body.every(takesNoStackWord)) {
3806
+ return false;
3807
+ }
3808
+ work.push(...succIdx[bi]);
3809
+ }
3810
+ return true;
3811
+ };
3812
+ const depthOfAddr = new Map(blocks.flatMap((b) => b.body.map((ins) => [ins.addr, r1At.get(ins)] as const)));
3813
+ for (const [off, { addr, mem, read, stored }] of valueSlots) {
3814
+ const depth = depthOfAddr.get(addr);
3815
+ const deadLocal = typeof depth === 'number' && off < 0 && off >= loadedWordBound(depth);
3816
+ const structHome = stored.length > 0 && stored.every(({ v, at }) => images.has(v) && noCallAfterTakesStack(at));
3817
+ if (!read && !deadLocal && !structHome) {
1657
3818
  throw new PpcUnsupportedError(
1658
3819
  `cannot lift '${name}': local stack frames not supported — '${mem}' is stored at 0x${addr.toString(16)} ` +
1659
3820
  `and never read back, and an unread home and a call's argument passed on the stack are the same store`,
1660
3821
  );
1661
3822
  }
1662
3823
  }
1663
- const argSlots = fpuArgSlots(name, fpu, ARG_REGS, isFpKey, fpType);
3824
+ const argSlots = ppcArgSlots(name, fpuArgSlots(name, fpu, ARG_REGS, isFpKey, fpType), ARG_REGS, prototypes[name]);
1664
3825
  mintArgSlotHoles(ssa, preds[0].length > 0, argSlots);
3826
+ const droppedPairHalf = computedPairHalf(
3827
+ irBlocks,
3828
+ pairReturns,
3829
+ target.compilerBehaviors.keepsUnusedReads === true
3830
+ ? (v) => paramReg.get(v) === RET && irBlocks[0].params.includes(v)
3831
+ : undefined,
3832
+ truthDifferences,
3833
+ (v) => compared.has(v) || readWithoutOp.has(v) || ssa.frameWrites.has(v),
3834
+ );
1665
3835
  ssa.finish();
1666
- highHalves.assertNoneEscaped(name, irBlocks);
1667
- // Checked on the finished function, after `finish` pruned the joins nothing reads, because a
1668
- // return address reaches a use through a join's incoming arguments as well as directly.
1669
- for (const b of irBlocks) {
1670
- for (const op of b.ops) {
1671
- const lr = [...op.operands, ...op.successors.flatMap((s) => s.args)].find((v) => returnAddresses.has(v));
1672
- if (lr !== undefined) {
3836
+ // A copy of the condition register is no integer: only a one-bit extract in its own block reads
3837
+ // it, and a read across an edge resolves to it only now.
3838
+ for (const blk of irBlocks) {
3839
+ for (const op of blk.ops) {
3840
+ if ([...op.operands, ...op.successors.flatMap((x) => x.args)].some((v) => crCopies.has(v))) {
1673
3841
  throw new PpcUnsupportedError(
1674
- `cannot lift '${name}': the return address the 'mflr' at 0x${returnAddresses.get(lr)!.toString(16)} ` +
1675
- `copies out is used as a value — a function reading its own return address has no C spelling`,
3842
+ `cannot lift '${name}': a copy of the condition register ('mfcr') is read as an integer, which is not modelled`,
1676
3843
  );
1677
3844
  }
1678
3845
  }
1679
3846
  }
3847
+ refuseFpWordCrossings(name, ssa.fn);
3848
+ // Asked once the lift is done, so a refusal the stream meets first keeps its own reason.
3849
+ if (fpu) {
3850
+ settleReturnHome(name, ssa.fn, homeRule, returns, homes, carriesFloat, fpu, RET);
3851
+ }
3852
+ resolveJoinedCarryConsts(name, ssa.fn, joinedCarryConsts);
3853
+ if (droppedPairHalf) {
3854
+ throw new PpcUnsupportedError(
3855
+ `cannot lift '${name}': at a word return no declaration sizes, r4 holds a value this function computed ` +
3856
+ 'and nothing reads, which is the other word of a 64-bit return (r3:r4); returning r3 alone would drop it',
3857
+ );
3858
+ }
3859
+ highHalves.assertNoneEscaped(name, irBlocks);
3860
+ tableBases.assertNoneEscaped(name, irBlocks);
3861
+ // A non-value reaches a use through a join's incoming arguments as well as directly (`nonValues`).
3862
+ assertNoNonValueEscaped(irBlocks);
3863
+ assertNonValuesConsumed();
3864
+
3865
+ if (frame !== null) {
3866
+ // Prove every `laddr` names storage of this frame at a shape the machine states, or decline
3867
+ // (`auditFrameObjects`); an accepted function then gets the device pin the audit answers. The
3868
+ // slot model's words are named from the pushed r1, as the audit names every offset.
3869
+ const verdict = FRAME_OBJECT_AUDIT.run({
3870
+ name,
3871
+ irBlocks,
3872
+ ownedLocals: frame.area,
3873
+ declaredLocals: frame.area,
3874
+ usedSlotOffsets: new Set(
3875
+ [...valueSlots.keys()].map((o) => o + frame.size).filter((o) => o >= 0 && o < frame.size),
3876
+ ),
3877
+ // the bound of the pushed frame, named from the pushed r1 as every offset here is
3878
+ loadedWords:
3879
+ loadedFrom === undefined ? undefined : { from: loadedFrom, bound: loadedWordBound(-frame.size) + frame.size },
3880
+ capturedObjectIsTheWholeFrame: false,
3881
+ movedCaptures,
3882
+ returnsWithoutHiddenPointer: calls.returnsWithoutHiddenPointer,
3883
+ symbols: undefined,
3884
+ target,
3885
+ oneObject,
3886
+ });
3887
+ if (verdict !== undefined && 'oneObject' in verdict) {
3888
+ return verdict;
3889
+ }
3890
+ if (verdict !== undefined) {
3891
+ pinDeviceAccesses(irBlocks, verdict, target, undefined);
3892
+ }
3893
+ }
3894
+ pinDeadReads(irBlocks, target, {
3895
+ name,
3896
+ own: prototypes[name],
3897
+ atWordReturn: unreadHigh,
3898
+ atUndeclaredCall: calls.heldAtUndeclaredCall,
3899
+ compared,
3900
+ spilled: ssa.frameWrites,
3901
+ discarded: leftInDiscardRegister(irBlocks, discarded, writtenElsewhere),
3902
+ declarations,
3903
+ });
1680
3904
 
1681
3905
  abiSortEntryParams(irBlocks[0], preds[0].length > 0, paramReg, argSlots);
1682
3906
  ssa.fn.localObjects = statics.finish();
1683
3907
  return ssa.fn;
1684
3908
  }
1685
3909
 
3910
+ /** WHETHER A WORD RETURN'S r4 WAS COMPUTED FOR NOTHING, which the lift refuses. `returns` holds,
3911
+ * for each word return no declaration pins the width of, the word r3 returns and what r4, the
3912
+ * second register of an r3:r4 return, may hold there (frontend/dead-reads.ts `unreadPairHalf`).
3913
+ * Where one of those values is the result of an op that nothing reads — directly, or through joins
3914
+ * nothing reads — the compiler computed it for the caller: an optimizing compiler deletes a pure
3915
+ * computation nobody reads, so a value that survives in r4 with no reader is the other word of a
3916
+ * 64-bit return, and returning r3 alone drops it. mwcc's `(unsigned long long)a - b` is `subfc
3917
+ * r4,r4,r3; subfe r3,r4,r4`, whose high word is the very `subfe` a 32-bit `-(a < b)` is, and `long
3918
+ * long x = a; if (c) x = -x; return x;` joins `0 - a` into r4 below the branch; both would lift as
3919
+ * their high word.
3920
+ *
3921
+ * AT A LEVEL THAT KEEPS DEAD CODE (`entryWord` given, where `keepsUnusedReads` holds) that premise
3922
+ * fails for a parameter: mwcc -O0 keeps `b = a + 1;` in b's register r4. There a return whose r3
3923
+ * is still the entry value (`entryWord`, directly or through joins of only it) is not asked, since
3924
+ * -O0 spells a 64-bit op as both words' ops and writes r3 even where its word is unchanged (`x |
3925
+ * 0xff` is `or r4,r4,r5; or r3,r3,r0` with r0 = 0), in every 64-bit shape measured on 2.4.2, 2.4.7
3926
+ * and 2.3.3. At -O4 an untouched r3 is evidence of nothing: `((u64)a << 32) | (b + 1)` is `addi
3927
+ * r4,r4,1`. A dead parameter store beside a computed r3 (`b = a * 3; return a + 1;`) still
3928
+ * refuses.
3929
+ *
3930
+ * Asked BEFORE `ssa.finish`, while every join is still a parameter with its incoming values wired,
3931
+ * so a join the trivial-phi pass would fold is walked, not lost, and refused AFTER it, so an
3932
+ * unmodelled instruction or an unresolved `subfze` keeps its own message. `exemptValue` names a
3933
+ * value the machine did use where the lift left no op for the use: one a compare consumed (`compared`),
3934
+ * one a carry reader folded away (`readWithoutOp`: ac-decomp's `mSP_GetRealShopLevel` returns `(x
3935
+ * >= 25000)` as `li r4,-1; subfc r0,r0,r3; subfze r3,r4`), a `subfc`/`subfic` difference whose
3936
+ * carry a reader turned into a 0/1 truth value (`truthDifferences`, kept only where that reader is
3937
+ * the carry's sole consumer in its block and neither is r3's word: 2.3.3's `a + (a < b)` is
3938
+ * `subfc r4,r4,r3; subfe r0,r0,r0; neg r0,r0; add r3,r3,r0` over the dead `b`, and the signed one
3939
+ * keeps bit 0 of an `addze`; a 64-bit high word reads the borrow as `CA - 1`, which r3 returns),
3940
+ * and one stored to the frame (`frameWrites`). A function that holds an
3941
+ * unmodelled instruction is not asked: the `opaque` may read r4 without naming it, and it declines
3942
+ * under its own name. A value r4 holds that the function did not compute — its entry value, the
3943
+ * caller's `a` that `long long z(unsigned a) { return a; }` moves there — is not seen, and that
3944
+ * word return still drops the low word. */
3945
+ type PairReturn = { word: Value; high: ReadonlySet<Value> };
3946
+ /** A `subfc`/`subfic` difference (`diff`) whose carry a reader made a 0/1 truth value: `carry` is
3947
+ * the `subfe rX,rY,rY` (CA - 1) or `addze` result, `truth` the `neg` or `clrlwi 31` over it. */
3948
+ type TruthDifference = { diff: Value; carry: Value; truth: Value };
3949
+ function computedPairHalf(
3950
+ blocks: readonly Block[],
3951
+ returns: readonly PairReturn[],
3952
+ entryWord: ((v: Value) => boolean) | undefined,
3953
+ truthDifferences: readonly TruthDifference[],
3954
+ exemptValue: (v: Value) => boolean,
3955
+ ): boolean {
3956
+ if (returns.every((r) => r.high.size === 0) || blocks.some((b) => b.ops.some((o) => o.opcode === 'opaque'))) {
3957
+ return false;
3958
+ }
3959
+ const defOf = new Map<Value, Op>();
3960
+ const blockOf = new Map<Op, Block>();
3961
+ const incoming = new Map<Value, Value[]>();
3962
+ // Every reader of each value: the ops that take it as an operand, a join argument counting as one.
3963
+ const readers = new Map<Value, number>();
3964
+ const readBy = (v: Value) => readers.set(v, (readers.get(v) ?? 0) + 1);
3965
+ for (const b of blocks) {
3966
+ for (const o of b.ops) {
3967
+ blockOf.set(o, b);
3968
+ o.results.forEach((r) => defOf.set(r, o));
3969
+ o.operands.forEach(readBy);
3970
+ for (const succ of o.successors) {
3971
+ succ.args.forEach((a, i) => {
3972
+ readBy(a);
3973
+ const p = succ.block.params[i];
3974
+ incoming.set(p, [...(incoming.get(p) ?? []), a]);
3975
+ });
3976
+ }
3977
+ }
3978
+ }
3979
+ // What r3 may hold at a word return: the word itself, and every value a join of it carries in.
3980
+ const returned = new Set<Value>();
3981
+ const toReturn = returns.map((r) => r.word);
3982
+ while (toReturn.length > 0) {
3983
+ const v = toReturn.pop()!;
3984
+ if (!returned.has(v)) {
3985
+ returned.add(v);
3986
+ toReturn.push(...(incoming.get(v) ?? []));
3987
+ }
3988
+ }
3989
+ // A truth value's difference is the compare's by-product only where nothing else saw the carry:
3990
+ // the reader is its sole consumer and runs in its block, so on every path the difference takes,
3991
+ // and neither the carry nor the truth value is what r3 returns. mwcc's `u64 d = (u64)a - b; *p =
3992
+ // -(int)(d >> 32); return d;` is `subfc r4,r4,r3; subfe r3,r4,r4; neg r0,r3`: that `subfe` is
3993
+ // the high word, the `neg` beside it is a second reader, and r4 is the low word.
3994
+ const comparedDifference = new Set<Value>();
3995
+ for (const { diff, carry, truth } of truthDifferences) {
3996
+ const reader = defOf.get(truth);
3997
+ const writer = defOf.get(carry);
3998
+ if (
3999
+ reader !== undefined &&
4000
+ writer !== undefined &&
4001
+ reader.operands.filter((x) => x === carry).length === 1 &&
4002
+ readers.get(carry) === 1 &&
4003
+ !exemptValue(carry) &&
4004
+ blockOf.get(reader) === blockOf.get(writer) &&
4005
+ !returned.has(carry) &&
4006
+ !returned.has(truth)
4007
+ ) {
4008
+ comparedDifference.add(diff);
4009
+ }
4010
+ }
4011
+ const exempt = (v: Value) => exemptValue(v) || comparedDifference.has(v);
4012
+ // What some kept op reads, through the joins that carry it there.
4013
+ const read = new Set<Value>();
4014
+ const work: Value[] = [];
4015
+ const mark = (v: Value) => {
4016
+ if (!read.has(v)) {
4017
+ read.add(v);
4018
+ work.push(v);
4019
+ }
4020
+ };
4021
+ for (const b of blocks) {
4022
+ for (const o of b.ops) {
4023
+ if (!deletableWhenDead(o)) {
4024
+ o.operands.forEach(mark);
4025
+ }
4026
+ }
4027
+ }
4028
+ while (work.length > 0) {
4029
+ const v = work.pop()!;
4030
+ defOf.get(v)?.operands.forEach(mark);
4031
+ incoming.get(v)?.forEach(mark);
4032
+ }
4033
+ const seen = new Set<Value>();
4034
+ const computedForNothing = (v: Value): boolean => {
4035
+ if (seen.has(v) || read.has(v) || exempt(v)) {
4036
+ return false;
4037
+ }
4038
+ seen.add(v);
4039
+ const def = defOf.get(v);
4040
+ if (def !== undefined) {
4041
+ return def.opcode !== 'undef';
4042
+ }
4043
+ return (incoming.get(v) ?? []).some(computedForNothing);
4044
+ };
4045
+ // Whether `v` is r3's entry value on every path: directly, or through joins of nothing else.
4046
+ const entry = (v: Value, on: Set<Value>): boolean => {
4047
+ if (entryWord!(v)) {
4048
+ return true;
4049
+ }
4050
+ const ins = incoming.get(v);
4051
+ if (ins === undefined || on.has(v)) {
4052
+ return ins !== undefined;
4053
+ }
4054
+ on.add(v);
4055
+ return ins.every((x) => entry(x, on));
4056
+ };
4057
+ return returns.some(
4058
+ (r) => !(entryWord !== undefined && entry(r.word, new Set())) && [...r.high].some(computedForNothing),
4059
+ );
4060
+ }
4061
+
4062
+ /** Settle each `subfze rD,rA` whose rA came in through a join (the `subfze` decode): where every
4063
+ * path brings the same constant K — the join's incoming values followed back, each a `const` or a
4064
+ * join of its own, a cycle adding nothing — the `~rA + CA` becomes `~K + CA`, which is CA itself
4065
+ * for mwcc's `li -1`. Anything else refuses, as an unmodelled `subfze` does in its own block: over
4066
+ * a register that is no constant it is the high half of a 64-bit negate. */
4067
+ function resolveJoinedCarryConsts(
4068
+ name: string,
4069
+ fn: Fn,
4070
+ pending: readonly { ins: Instr; not: Op; add: Op; ca: Value }[],
4071
+ ): void {
4072
+ if (pending.length === 0) {
4073
+ return;
4074
+ }
4075
+ const defs = new Map<Value, Op>();
4076
+ const incoming = new Map<Value, Value[]>();
4077
+ for (const b of fn.blocks) {
4078
+ for (const o of b.ops) {
4079
+ o.results.forEach((r) => defs.set(r, o));
4080
+ for (const succ of o.successors) {
4081
+ succ.args.forEach((a, i) => {
4082
+ const p = succ.block.params[i];
4083
+ incoming.set(p, [...(incoming.get(p) ?? []), a]);
4084
+ });
4085
+ }
4086
+ }
4087
+ }
4088
+ // The one constant `v` is on every path, `null` where it is none, `undefined` where only a cycle
4089
+ // brings it (which says nothing).
4090
+ const constantOf = (v: Value, seen: Set<Value>): number | null | undefined => {
4091
+ const def = defs.get(v);
4092
+ if (def !== undefined) {
4093
+ return def.opcode === 'const' && typeof def.attrs.value === 'number' ? def.attrs.value | 0 : null;
4094
+ }
4095
+ const ins = incoming.get(v);
4096
+ if (ins === undefined) {
4097
+ return null; // an entry parameter
4098
+ }
4099
+ if (seen.has(v)) {
4100
+ return undefined;
4101
+ }
4102
+ seen.add(v);
4103
+ let k: number | undefined;
4104
+ for (const a of ins) {
4105
+ const c = constantOf(a, seen);
4106
+ if (c === null || (c !== undefined && k !== undefined && c !== k)) {
4107
+ return null;
4108
+ }
4109
+ k ??= c;
4110
+ }
4111
+ return k ?? null;
4112
+ };
4113
+ for (const { ins, not, add, ca } of pending) {
4114
+ const k = constantOf(not.operands[0], new Set());
4115
+ if (k === null || k === undefined) {
4116
+ throw new PpcUnsupportedError(
4117
+ `cannot lift '${name}': unmodelled instruction '${ins.mnemonic}' at 0x${ins.addr.toString(16)} — it reads ` +
4118
+ `${ins.ops[1]}, which is not one constant on every path into its block, and a subfze over a register is the ` +
4119
+ 'upper word of a 64-bit negate',
4120
+ );
4121
+ }
4122
+ const home = fn.blocks.find((b) => b.ops.includes(add))!;
4123
+ if (~k === 0) {
4124
+ replaceAllUsesWith(fn, add.results[0], ca);
4125
+ home.ops.splice(home.ops.indexOf(add), 1);
4126
+ home.ops.splice(home.ops.indexOf(not), 1);
4127
+ } else {
4128
+ Object.assign(not, { opcode: 'const', operands: [], attrs: { value: ~k } });
4129
+ }
4130
+ }
4131
+ }
4132
+
1686
4133
  function mkCmp(ops: Op[], opc: Opcode, l: Value, r: Value): Value {
1687
4134
  const v = mkValue(T.unk(32));
1688
4135
  ops.push(mkOp(opc, { operands: [l, r], results: [v] }));
1689
4136
  return v;
1690
4137
  }
1691
4138
 
4139
+ /** The EABI's slots: the shared register-then-stack-word rule (`stackArgSlots`), refusing a stack
4140
+ * word read beside a floating-point argument — the FPRs are a separate file, so where a float
4141
+ * parameter falls among the integer ones is not in the object — or in a function whose own
4142
+ * declaration has a 64-bit or floating parameter, which takes words the integer count does not. */
4143
+ function ppcArgSlots(name: string, regs: ArgSlots, argRegs: readonly string[], own: FnProto | undefined): ArgSlots {
4144
+ return stackArgSlots(regs, argRegs, (readKeys, top) => {
4145
+ const float = readKeys.find(isFpKey);
4146
+ const declared = Array.isArray(own?.params)
4147
+ ? own.params.find((p) => isFloatingSpelling(p) || (declaredWidth(p) ?? 0) > 32)
4148
+ : undefined;
4149
+ if (float !== undefined || declared !== undefined) {
4150
+ throw new PpcUnsupportedError(
4151
+ `cannot lift '${name}': incoming stack argument ${top + 1} is read beside ` +
4152
+ (float !== undefined ? `the floating-point argument in ${float}` : `a parameter declared '${declared}'`) +
4153
+ ` — which words that argument takes ahead of the stack words is not modelled`,
4154
+ );
4155
+ }
4156
+ });
4157
+ }
4158
+
1692
4159
  /** The PowerPC / CodeWarrior frontend, registered for the `ppc` target. */
1693
4160
  export const ppcFrontend: Frontend = { id: 'ppc', inputFormat: 'objdump', lift };
4161
+
4162
+ /** Internal surface for this module's own tests, and for nothing else (`frontend/frame-objects.ts`
4163
+ * says why it is not a plain export). */
4164
+ export const __testing = {
4165
+ PARAM_AREA_AT,
4166
+ /** the local area `lift` draws for one function of objdump text (`liftPrelude`), or its refusal */
4167
+ frameOf: (name: string, asm: string, target: TargetDescription, prototypes: Prototypes = {}, asmData?: AsmData) => {
4168
+ const instrs = parseDisasm(sliceSymbol(asm, name));
4169
+ const jts = asmData ? recoverPpcJumpTables(instrs, asmData) : new Map<number, PpcJT>();
4170
+ return liftPrelude(name, instrs, target, prototypes, jts).frame;
4171
+ },
4172
+ };