Skip to content

Commit 215c0d2

Browse files
davemgreendavemgreen
committed
[AArch64][GlobalISel] Reorder getActionDefinitionsBuilders. NFC
These had grown fairly organically and the order wasn't very clear. I've tried to move some similar nodes closer to one another.
1 parent d928a67 commit 215c0d2

File tree

1 file changed

+141
-142
lines changed

1 file changed

+141
-142
lines changed

‎llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp‎

Lines changed: 141 additions & 142 deletions
Original file line numberDiff line numberDiff line change
@@ -120,14 +120,26 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
120120
.clampMaxNumElements(0, s64, 2)
121121
.clampMaxNumElements(0, p0, 2);
122122

123-
getActionDefinitionsBuilder(G_BSWAP)
124-
.legalFor({s32, s64, v4s16, v8s16, v2s32, v4s32, v2s64})
125-
.widenScalarOrEltToNextPow2(0, 16)
123+
getActionDefinitionsBuilder(G_INSERT)
124+
.legalIf(all(typeInSet(0, {s32, s64, p0}), typeInSet(1, {s8, s16, s32}),
125+
smallerThan(1, 0)))
126+
.widenScalarToNextPow2(0)
126127
.clampScalar(0, s32, s64)
127-
.clampNumElements(0, v4s16, v8s16)
128-
.clampNumElements(0, v2s32, v4s32)
129-
.clampNumElements(0, v2s64, v2s64)
130-
.moreElementsToNextPow2(0);
128+
.widenScalarToNextPow2(1)
129+
.minScalar(1, s8)
130+
.maxScalarIf(typeInSet(0, {s32}), 1, s16)
131+
.maxScalarIf(typeInSet(0, {s64, p0}), 1, s32);
132+
133+
getActionDefinitionsBuilder(G_EXTRACT)
134+
.legalIf(all(typeInSet(0, {s16, s32, s64, p0}),
135+
typeInSet(1, {s32, s64, s128, p0}), smallerThan(0, 1)))
136+
.widenScalarToNextPow2(1)
137+
.clampScalar(1, s32, s128)
138+
.widenScalarToNextPow2(0)
139+
.minScalar(0, s16)
140+
.maxScalarIf(typeInSet(1, {s32}), 0, s16)
141+
.maxScalarIf(typeInSet(1, {s64, p0}), 0, s32)
142+
.maxScalarIf(typeInSet(1, {s128}), 0, s64);
131143

132144
getActionDefinitionsBuilder({G_ADD, G_SUB, G_AND, G_OR, G_XOR})
133145
.legalFor({s32, s64, v8s8, v16s8, v4s16, v8s16, v2s32, v4s32, v2s64})
@@ -253,13 +265,112 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
253265
.clampNumElements(0, v2s32, v4s32)
254266
.lower();
255267

268+
// FIXME: Legal vector types are only legal with NEON.
269+
getActionDefinitionsBuilder(G_ABS)
270+
.legalFor(HasCSSC, {s32, s64})
271+
.legalFor(PackedVectorAllTypeList)
272+
.customIf([=](const LegalityQuery &Q) {
273+
// TODO: Fix suboptimal codegen for 128+ bit types.
274+
LLT SrcTy = Q.Types[0];
275+
return SrcTy.isScalar() && SrcTy.getSizeInBits() < 128;
276+
})
277+
.widenScalarIf(
278+
[=](const LegalityQuery &Query) { return Query.Types[0] == v4s8; },
279+
[=](const LegalityQuery &Query) { return std::make_pair(0, v4s16); })
280+
.widenScalarIf(
281+
[=](const LegalityQuery &Query) { return Query.Types[0] == v2s16; },
282+
[=](const LegalityQuery &Query) { return std::make_pair(0, v2s32); })
283+
.clampNumElements(0, v8s8, v16s8)
284+
.clampNumElements(0, v4s16, v8s16)
285+
.clampNumElements(0, v2s32, v4s32)
286+
.clampNumElements(0, v2s64, v2s64)
287+
.moreElementsToNextPow2(0)
288+
.lower();
289+
256290
getActionDefinitionsBuilder(
257291
{G_SADDE, G_SSUBE, G_UADDE, G_USUBE, G_SADDO, G_SSUBO, G_UADDO, G_USUBO})
258292
.legalFor({{s32, s32}, {s64, s32}})
259293
.clampScalar(0, s32, s64)
260294
.clampScalar(1, s32, s64)
261295
.widenScalarToNextPow2(0);
262296

297+
getActionDefinitionsBuilder({G_FSHL, G_FSHR})
298+
.customFor({{s32, s32}, {s32, s64}, {s64, s64}})
299+
.lower();
300+
301+
getActionDefinitionsBuilder(G_ROTR)
302+
.legalFor({{s32, s64}, {s64, s64}})
303+
.customIf([=](const LegalityQuery &Q) {
304+
return Q.Types[0].isScalar() && Q.Types[1].getScalarSizeInBits() < 64;
305+
})
306+
.lower();
307+
getActionDefinitionsBuilder(G_ROTL).lower();
308+
309+
getActionDefinitionsBuilder({G_SBFX, G_UBFX})
310+
.customFor({{s32, s32}, {s64, s64}});
311+
312+
auto always = [=](const LegalityQuery &Q) { return true; };
313+
getActionDefinitionsBuilder(G_CTPOP)
314+
.legalFor(HasCSSC, {{s32, s32}, {s64, s64}})
315+
.legalFor({{v8s8, v8s8}, {v16s8, v16s8}})
316+
.customFor(!HasCSSC, {{s32, s32}, {s64, s64}})
317+
.customFor({{s128, s128},
318+
{v4s16, v4s16},
319+
{v8s16, v8s16},
320+
{v2s32, v2s32},
321+
{v4s32, v4s32},
322+
{v2s64, v2s64}})
323+
.clampScalar(0, s32, s128)
324+
.widenScalarToNextPow2(0)
325+
.minScalarEltSameAsIf(always, 1, 0)
326+
.maxScalarEltSameAsIf(always, 1, 0);
327+
328+
getActionDefinitionsBuilder(G_CTLZ)
329+
.legalForCartesianProduct(
330+
{s32, s64, v8s8, v16s8, v4s16, v8s16, v2s32, v4s32})
331+
.scalarize(1)
332+
.widenScalarToNextPow2(1, /*Min=*/32)
333+
.clampScalar(1, s32, s64)
334+
.scalarSameSizeAs(0, 1);
335+
getActionDefinitionsBuilder(G_CTLZ_ZERO_UNDEF).lower();
336+
337+
getActionDefinitionsBuilder(G_CTTZ)
338+
.lowerIf(isVector(0))
339+
.widenScalarToNextPow2(1, /*Min=*/32)
340+
.clampScalar(1, s32, s64)
341+
.scalarSameSizeAs(0, 1)
342+
.legalFor(HasCSSC, {s32, s64})
343+
.customFor(!HasCSSC, {s32, s64});
344+
345+
getActionDefinitionsBuilder(G_CTTZ_ZERO_UNDEF).lower();
346+
347+
// TODO: Custom lowering for v2s32, v4s32, v2s64.
348+
getActionDefinitionsBuilder(G_BITREVERSE)
349+
.legalFor({s32, s64, v8s8, v16s8})
350+
.widenScalarToNextPow2(0, /*Min = */ 32)
351+
.clampScalar(0, s32, s64)
352+
.lower();
353+
354+
getActionDefinitionsBuilder(G_BSWAP)
355+
.legalFor({s32, s64, v4s16, v8s16, v2s32, v4s32, v2s64})
356+
.widenScalarOrEltToNextPow2(0, 16)
357+
.clampScalar(0, s32, s64)
358+
.clampNumElements(0, v4s16, v8s16)
359+
.clampNumElements(0, v2s32, v4s32)
360+
.clampNumElements(0, v2s64, v2s64)
361+
.moreElementsToNextPow2(0);
362+
363+
getActionDefinitionsBuilder({G_UADDSAT, G_SADDSAT, G_USUBSAT, G_SSUBSAT})
364+
.legalFor({v8s8, v16s8, v4s16, v8s16, v2s32, v4s32, v2s64})
365+
.legalFor(HasSVE, {nxv16s8, nxv8s16, nxv4s32, nxv2s64})
366+
.clampNumElements(0, v8s8, v16s8)
367+
.clampNumElements(0, v4s16, v8s16)
368+
.clampNumElements(0, v2s32, v4s32)
369+
.clampMaxNumElements(0, s64, 2)
370+
.scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
371+
.moreElementsToNextPow2(0)
372+
.lower();
373+
263374
getActionDefinitionsBuilder(
264375
{G_FADD, G_FSUB, G_FMUL, G_FDIV, G_FMA, G_FSQRT, G_FMAXNUM, G_FMINNUM,
265376
G_FMAXIMUM, G_FMINIMUM, G_FCEIL, G_FFLOOR, G_FRINT, G_FNEARBYINT,
@@ -309,27 +420,22 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
309420
.minScalar(0, s32)
310421
.libcallFor({{s32, s32}, {s64, s32}, {s128, s32}});
311422

312-
getActionDefinitionsBuilder(G_INSERT)
313-
.legalIf(all(typeInSet(0, {s32, s64, p0}),
314-
typeInSet(1, {s8, s16, s32}), smallerThan(1, 0)))
315-
.widenScalarToNextPow2(0)
316-
.clampScalar(0, s32, s64)
317-
.widenScalarToNextPow2(1)
318-
.minScalar(1, s8)
319-
.maxScalarIf(typeInSet(0, {s32}), 1, s16)
320-
.maxScalarIf(typeInSet(0, {s64, p0}), 1, s32);
423+
// TODO: Libcall support for s128.
424+
// TODO: s16 should be legal with full FP16 support.
425+
getActionDefinitionsBuilder({G_LROUND, G_LLROUND})
426+
.legalFor({{s64, s32}, {s64, s64}});
321427

322-
getActionDefinitionsBuilder(G_EXTRACT)
323-
.legalIf(all(typeInSet(0, {s16, s32, s64, p0}),
324-
typeInSet(1, {s32, s64, s128, p0}), smallerThan(0, 1)))
325-
.widenScalarToNextPow2(1)
326-
.clampScalar(1, s32, s128)
327-
.widenScalarToNextPow2(0)
328-
.minScalar(0, s16)
329-
.maxScalarIf(typeInSet(1, {s32}), 0, s16)
330-
.maxScalarIf(typeInSet(1, {s64, p0}), 0, s32)
331-
.maxScalarIf(typeInSet(1, {s128}), 0, s64);
428+
// TODO: Custom legalization for mismatched types.
429+
getActionDefinitionsBuilder(G_FCOPYSIGN)
430+
.moreElementsIf(
431+
[](const LegalityQuery &Query) { return Query.Types[0].isScalar(); },
432+
[=](const LegalityQuery &Query) {
433+
const LLT Ty = Query.Types[0];
434+
return std::pair(0, LLT::fixed_vector(Ty == s16 ? 4 : 2, Ty));
435+
})
436+
.lower();
332437

438+
getActionDefinitionsBuilder(G_FMAD).lower();
333439

334440
for (unsigned Op : {G_SEXTLOAD, G_ZEXTLOAD}) {
335441
auto &Actions = getActionDefinitionsBuilder(Op);
@@ -1035,32 +1141,6 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
10351141

10361142
getActionDefinitionsBuilder(G_BUILD_VECTOR_TRUNC).lower();
10371143

1038-
getActionDefinitionsBuilder(G_CTLZ)
1039-
.legalForCartesianProduct(
1040-
{s32, s64, v8s8, v16s8, v4s16, v8s16, v2s32, v4s32})
1041-
.scalarize(1)
1042-
.widenScalarToNextPow2(1, /*Min=*/32)
1043-
.clampScalar(1, s32, s64)
1044-
.scalarSameSizeAs(0, 1);
1045-
getActionDefinitionsBuilder(G_CTLZ_ZERO_UNDEF).lower();
1046-
1047-
// TODO: Custom lowering for v2s32, v4s32, v2s64.
1048-
getActionDefinitionsBuilder(G_BITREVERSE)
1049-
.legalFor({s32, s64, v8s8, v16s8})
1050-
.widenScalarToNextPow2(0, /*Min = */ 32)
1051-
.clampScalar(0, s32, s64)
1052-
.lower();
1053-
1054-
getActionDefinitionsBuilder(G_CTTZ_ZERO_UNDEF).lower();
1055-
1056-
getActionDefinitionsBuilder(G_CTTZ)
1057-
.lowerIf(isVector(0))
1058-
.widenScalarToNextPow2(1, /*Min=*/32)
1059-
.clampScalar(1, s32, s64)
1060-
.scalarSameSizeAs(0, 1)
1061-
.legalFor(HasCSSC, {s32, s64})
1062-
.customFor(!HasCSSC, {s32, s64});
1063-
10641144
getActionDefinitionsBuilder(G_SHUFFLE_VECTOR)
10651145
.legalIf([=](const LegalityQuery &Query) {
10661146
const LLT &DstTy = Query.Types[0];
@@ -1122,6 +1202,15 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
11221202
SrcTy.getNumElements())));
11231203
});
11241204

1205+
getActionDefinitionsBuilder(G_EXTRACT_SUBVECTOR)
1206+
.legalFor({{v8s8, v16s8}, {v4s16, v8s16}, {v2s32, v4s32}})
1207+
.widenScalarOrEltToNextPow2(0)
1208+
.immIdx(0); // Inform verifier imm idx 0 is handled.
1209+
1210+
// TODO: {nxv16s8, s8}, {nxv8s16, s16}
1211+
getActionDefinitionsBuilder(G_SPLAT_VECTOR)
1212+
.legalFor(HasSVE, {{nxv4s32, s32}, {nxv2s64, s64}});
1213+
11251214
getActionDefinitionsBuilder(G_JUMP_TABLE).legalFor({p0});
11261215

11271216
getActionDefinitionsBuilder(G_BRJT).legalFor({{p0, s64}});
@@ -1153,28 +1242,6 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
11531242
.libcall();
11541243
}
11551244

1156-
// FIXME: Legal vector types are only legal with NEON.
1157-
getActionDefinitionsBuilder(G_ABS)
1158-
.legalFor(HasCSSC, {s32, s64})
1159-
.legalFor(PackedVectorAllTypeList)
1160-
.customIf([=](const LegalityQuery &Q) {
1161-
// TODO: Fix suboptimal codegen for 128+ bit types.
1162-
LLT SrcTy = Q.Types[0];
1163-
return SrcTy.isScalar() && SrcTy.getSizeInBits() < 128;
1164-
})
1165-
.widenScalarIf(
1166-
[=](const LegalityQuery &Query) { return Query.Types[0] == v4s8; },
1167-
[=](const LegalityQuery &Query) { return std::make_pair(0, v4s16); })
1168-
.widenScalarIf(
1169-
[=](const LegalityQuery &Query) { return Query.Types[0] == v2s16; },
1170-
[=](const LegalityQuery &Query) { return std::make_pair(0, v2s32); })
1171-
.clampNumElements(0, v8s8, v16s8)
1172-
.clampNumElements(0, v4s16, v8s16)
1173-
.clampNumElements(0, v2s32, v4s32)
1174-
.clampNumElements(0, v2s64, v2s64)
1175-
.moreElementsToNextPow2(0)
1176-
.lower();
1177-
11781245
// For fadd reductions we have pairwise operations available. We treat the
11791246
// usual legal types as legal and handle the lowering to pairwise instructions
11801247
// later.
@@ -1284,65 +1351,6 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
12841351
// TODO: Update this to correct handling when adding AArch64/SVE support.
12851352
getActionDefinitionsBuilder(G_VECTOR_COMPRESS).lower();
12861353

1287-
getActionDefinitionsBuilder({G_FSHL, G_FSHR})
1288-
.customFor({{s32, s32}, {s32, s64}, {s64, s64}})
1289-
.lower();
1290-
1291-
getActionDefinitionsBuilder(G_ROTR)
1292-
.legalFor({{s32, s64}, {s64, s64}})
1293-
.customIf([=](const LegalityQuery &Q) {
1294-
return Q.Types[0].isScalar() && Q.Types[1].getScalarSizeInBits() < 64;
1295-
})
1296-
.lower();
1297-
getActionDefinitionsBuilder(G_ROTL).lower();
1298-
1299-
getActionDefinitionsBuilder({G_SBFX, G_UBFX})
1300-
.customFor({{s32, s32}, {s64, s64}});
1301-
1302-
auto always = [=](const LegalityQuery &Q) { return true; };
1303-
getActionDefinitionsBuilder(G_CTPOP)
1304-
.legalFor(HasCSSC, {{s32, s32}, {s64, s64}})
1305-
.legalFor({{v8s8, v8s8}, {v16s8, v16s8}})
1306-
.customFor(!HasCSSC, {{s32, s32}, {s64, s64}})
1307-
.customFor({{s128, s128},
1308-
{v2s32, v2s32},
1309-
{v4s32, v4s32},
1310-
{v4s16, v4s16},
1311-
{v8s16, v8s16},
1312-
{v2s64, v2s64}})
1313-
.clampScalar(0, s32, s128)
1314-
.widenScalarToNextPow2(0)
1315-
.minScalarEltSameAsIf(always, 1, 0)
1316-
.maxScalarEltSameAsIf(always, 1, 0);
1317-
1318-
getActionDefinitionsBuilder({G_UADDSAT, G_SADDSAT, G_USUBSAT, G_SSUBSAT})
1319-
.legalFor({v8s8, v16s8, v4s16, v8s16, v2s32, v4s32, v2s64})
1320-
.legalFor(HasSVE, {nxv16s8, nxv8s16, nxv4s32, nxv2s64})
1321-
.clampNumElements(0, v8s8, v16s8)
1322-
.clampNumElements(0, v4s16, v8s16)
1323-
.clampNumElements(0, v2s32, v4s32)
1324-
.clampMaxNumElements(0, s64, 2)
1325-
.scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
1326-
.moreElementsToNextPow2(0)
1327-
.lower();
1328-
1329-
// TODO: Libcall support for s128.
1330-
// TODO: s16 should be legal with full FP16 support.
1331-
getActionDefinitionsBuilder({G_LROUND, G_LLROUND})
1332-
.legalFor({{s64, s32}, {s64, s64}});
1333-
1334-
// TODO: Custom legalization for mismatched types.
1335-
getActionDefinitionsBuilder(G_FCOPYSIGN)
1336-
.moreElementsIf(
1337-
[](const LegalityQuery &Query) { return Query.Types[0].isScalar(); },
1338-
[=](const LegalityQuery &Query) {
1339-
const LLT Ty = Query.Types[0];
1340-
return std::pair(0, LLT::fixed_vector(Ty == s16 ? 4 : 2, Ty));
1341-
})
1342-
.lower();
1343-
1344-
getActionDefinitionsBuilder(G_FMAD).lower();
1345-
13461354
// Access to floating-point environment.
13471355
getActionDefinitionsBuilder({G_GET_FPENV, G_SET_FPENV, G_RESET_FPENV,
13481356
G_GET_FPMODE, G_SET_FPMODE, G_RESET_FPMODE})
@@ -1354,15 +1362,6 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
13541362

13551363
getActionDefinitionsBuilder({G_SCMP, G_UCMP}).lower();
13561364

1357-
getActionDefinitionsBuilder(G_EXTRACT_SUBVECTOR)
1358-
.legalFor({{v8s8, v16s8}, {v4s16, v8s16}, {v2s32, v4s32}})
1359-
.widenScalarOrEltToNextPow2(0)
1360-
.immIdx(0); // Inform verifier imm idx 0 is handled.
1361-
1362-
// TODO: {nxv16s8, s8}, {nxv8s16, s16}
1363-
getActionDefinitionsBuilder(G_SPLAT_VECTOR)
1364-
.legalFor(HasSVE, {{nxv4s32, s32}, {nxv2s64, s64}});
1365-
13661365
getLegacyLegalizerInfo().computeTables();
13671366
verify(*ST.getInstrInfo());
13681367
}

0 commit comments

Comments
 (0)