Skip to content

Commit bfef161

Browse files
authored
[AArch64][GlobalISel] Legalize Shifts for Smaller/Larger Vectors (#78750)
Legalize shl/lshr/ashr for smaller/larger vector widths with legal element sizes Smaller than legal vector types does not work at the moment as it relies on G_ANYEXT to work with smaller than legal vector types
1 parent 4f46905 commit bfef161

File tree

6 files changed

+412
-237
lines changed

6 files changed

+412
-237
lines changed

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -5198,7 +5198,10 @@ LegalizerHelper::moreElementsVector(MachineInstr &MI, unsigned TypeIdx,
51985198
case TargetOpcode::G_FMAXIMUM:
51995199
case TargetOpcode::G_STRICT_FADD:
52005200
case TargetOpcode::G_STRICT_FSUB:
5201-
case TargetOpcode::G_STRICT_FMUL: {
5201+
case TargetOpcode::G_STRICT_FMUL:
5202+
case TargetOpcode::G_SHL:
5203+
case TargetOpcode::G_ASHR:
5204+
case TargetOpcode::G_LSHR: {
52025205
Observer.changingInstr(MI);
52035206
moreElementsVectorSrc(MI, MoreTy, 1);
52045207
moreElementsVectorSrc(MI, MoreTy, 2);

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

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -169,6 +169,8 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
169169
.widenScalarToNextPow2(0)
170170
.clampScalar(1, s32, s64)
171171
.clampScalar(0, s32, s64)
172+
.clampNumElements(0, v8s8, v16s8)
173+
.clampNumElements(0, v4s16, v8s16)
172174
.clampNumElements(0, v2s32, v4s32)
173175
.clampNumElements(0, v2s64, v2s64)
174176
.moreElementsToNextPow2(0)

llvm/test/CodeGen/AArch64/fcmp.ll

Lines changed: 174 additions & 129 deletions
Original file line numberDiff line numberDiff line change
@@ -8,9 +8,9 @@
88
; CHECK-GI-NEXT: warning: Instruction selection used fallback path for v3f64_i32
99
; CHECK-GI-NEXT: warning: Instruction selection used fallback path for v3f32_float
1010
; CHECK-GI-NEXT: warning: Instruction selection used fallback path for v3f32_i32
11-
; CHECK-GI-NEXT: warning: Instruction selection used fallback path for v7f16_half
11+
; CHECK-GI-FP16-NEXT: warning: Instruction selection used fallback path for v7f16_half
1212
; CHECK-GI-FP16-NEXT: warning: Instruction selection used fallback path for v16f16_half
13-
; CHECK-GI-NEXT: warning: Instruction selection used fallback path for v7f16_i32
13+
; CHECK-GI-FP16-NEXT: warning: Instruction selection used fallback path for v7f16_i32
1414
; CHECK-GI-FP16-NEXT: warning: Instruction selection used fallback path for v16f16_i32
1515

1616
define double @f64_double(double %a, double %b, double %d, double %e) {
@@ -437,62 +437,87 @@ define <7 x half> @v7f16_half(<7 x half> %a, <7 x half> %b, <7 x half> %d, <7 x
437437
;
438438
; CHECK-GI-NOFP16-LABEL: v7f16_half:
439439
; CHECK-GI-NOFP16: // %bb.0: // %entry
440-
; CHECK-GI-NOFP16-NEXT: mov h4, v1.h[1]
441-
; CHECK-GI-NOFP16-NEXT: mov h5, v0.h[1]
442-
; CHECK-GI-NOFP16-NEXT: fcvt s6, h1
443-
; CHECK-GI-NOFP16-NEXT: fcvt s7, h0
444-
; CHECK-GI-NOFP16-NEXT: mov h16, v1.h[2]
445-
; CHECK-GI-NOFP16-NEXT: fcvt s4, h4
440+
; CHECK-GI-NOFP16-NEXT: mov h4, v0.h[1]
441+
; CHECK-GI-NOFP16-NEXT: mov h5, v1.h[1]
442+
; CHECK-GI-NOFP16-NEXT: mov w9, #15 // =0xf
443+
; CHECK-GI-NOFP16-NEXT: fcvt s6, h0
444+
; CHECK-GI-NOFP16-NEXT: fcvt s7, h1
445+
; CHECK-GI-NOFP16-NEXT: mov h16, v0.h[2]
446+
; CHECK-GI-NOFP16-NEXT: mov h17, v1.h[2]
447+
; CHECK-GI-NOFP16-NEXT: mov h19, v1.h[4]
448+
; CHECK-GI-NOFP16-NEXT: mov h20, v1.h[5]
449+
; CHECK-GI-NOFP16-NEXT: fcvt s18, h4
446450
; CHECK-GI-NOFP16-NEXT: fcvt s5, h5
447-
; CHECK-GI-NOFP16-NEXT: fcmp s5, s4
448-
; CHECK-GI-NOFP16-NEXT: mov h4, v0.h[2]
449-
; CHECK-GI-NOFP16-NEXT: mov h5, v1.h[3]
450-
; CHECK-GI-NOFP16-NEXT: csetm w8, mi
451-
; CHECK-GI-NOFP16-NEXT: fcmp s7, s6
451+
; CHECK-GI-NOFP16-NEXT: mov h4, v0.h[3]
452+
; CHECK-GI-NOFP16-NEXT: fcmp s6, s7
452453
; CHECK-GI-NOFP16-NEXT: fcvt s7, h16
453-
; CHECK-GI-NOFP16-NEXT: fcvt s4, h4
454-
; CHECK-GI-NOFP16-NEXT: mov h6, v0.h[3]
455-
; CHECK-GI-NOFP16-NEXT: fcvt s5, h5
456-
; CHECK-GI-NOFP16-NEXT: mov h16, v0.h[4]
457-
; CHECK-GI-NOFP16-NEXT: csetm w9, mi
458-
; CHECK-GI-NOFP16-NEXT: fcmp s4, s7
459-
; CHECK-GI-NOFP16-NEXT: fmov s4, w9
460-
; CHECK-GI-NOFP16-NEXT: fcvt s6, h6
461-
; CHECK-GI-NOFP16-NEXT: mov h7, v1.h[4]
462-
; CHECK-GI-NOFP16-NEXT: fcvt s16, h16
463-
; CHECK-GI-NOFP16-NEXT: mov v4.h[1], w8
464-
; CHECK-GI-NOFP16-NEXT: csetm w8, mi
465-
; CHECK-GI-NOFP16-NEXT: fcmp s6, s5
466-
; CHECK-GI-NOFP16-NEXT: mov h5, v1.h[5]
467-
; CHECK-GI-NOFP16-NEXT: mov h6, v0.h[5]
468-
; CHECK-GI-NOFP16-NEXT: fcvt s7, h7
469-
; CHECK-GI-NOFP16-NEXT: mov v4.h[2], w8
470-
; CHECK-GI-NOFP16-NEXT: csetm w8, mi
471-
; CHECK-GI-NOFP16-NEXT: fcvt s5, h5
472-
; CHECK-GI-NOFP16-NEXT: fcvt s6, h6
473-
; CHECK-GI-NOFP16-NEXT: fcmp s16, s7
474-
; CHECK-GI-NOFP16-NEXT: mov h7, v1.h[6]
475-
; CHECK-GI-NOFP16-NEXT: mov h16, v0.h[6]
476-
; CHECK-GI-NOFP16-NEXT: mov h1, v1.h[7]
477-
; CHECK-GI-NOFP16-NEXT: mov h0, v0.h[7]
478-
; CHECK-GI-NOFP16-NEXT: mov v4.h[3], w8
479-
; CHECK-GI-NOFP16-NEXT: csetm w8, mi
480-
; CHECK-GI-NOFP16-NEXT: fcmp s6, s5
481-
; CHECK-GI-NOFP16-NEXT: fcvt s5, h7
482-
; CHECK-GI-NOFP16-NEXT: fcvt s6, h16
454+
; CHECK-GI-NOFP16-NEXT: fmov s6, w9
455+
; CHECK-GI-NOFP16-NEXT: fcvt s16, h17
456+
; CHECK-GI-NOFP16-NEXT: mov h17, v1.h[3]
457+
; CHECK-GI-NOFP16-NEXT: fcvt s19, h19
458+
; CHECK-GI-NOFP16-NEXT: mov h1, v1.h[6]
459+
; CHECK-GI-NOFP16-NEXT: cset w8, mi
460+
; CHECK-GI-NOFP16-NEXT: fcmp s18, s5
461+
; CHECK-GI-NOFP16-NEXT: mov h18, v0.h[4]
462+
; CHECK-GI-NOFP16-NEXT: fmov s5, w8
463+
; CHECK-GI-NOFP16-NEXT: cset w9, mi
464+
; CHECK-GI-NOFP16-NEXT: fcmp s7, s16
465+
; CHECK-GI-NOFP16-NEXT: fcvt s7, h4
466+
; CHECK-GI-NOFP16-NEXT: fcvt s16, h17
467+
; CHECK-GI-NOFP16-NEXT: fmov s17, w9
468+
; CHECK-GI-NOFP16-NEXT: mov v4.16b, v6.16b
469+
; CHECK-GI-NOFP16-NEXT: fcvt s18, h18
483470
; CHECK-GI-NOFP16-NEXT: fcvt s1, h1
471+
; CHECK-GI-NOFP16-NEXT: cset w8, mi
472+
; CHECK-GI-NOFP16-NEXT: mov v5.h[1], v17.h[0]
473+
; CHECK-GI-NOFP16-NEXT: mov v4.h[1], v6.h[0]
474+
; CHECK-GI-NOFP16-NEXT: mov h17, v0.h[5]
475+
; CHECK-GI-NOFP16-NEXT: fcmp s7, s16
476+
; CHECK-GI-NOFP16-NEXT: fmov s7, w8
477+
; CHECK-GI-NOFP16-NEXT: mov w8, #65535 // =0xffff
478+
; CHECK-GI-NOFP16-NEXT: mov h0, v0.h[6]
479+
; CHECK-GI-NOFP16-NEXT: mov v5.h[2], v7.h[0]
480+
; CHECK-GI-NOFP16-NEXT: fmov s7, w8
481+
; CHECK-GI-NOFP16-NEXT: mov v4.h[2], v6.h[0]
482+
; CHECK-GI-NOFP16-NEXT: cset w8, mi
483+
; CHECK-GI-NOFP16-NEXT: fcmp s18, s19
484+
; CHECK-GI-NOFP16-NEXT: fcvt s17, h17
485+
; CHECK-GI-NOFP16-NEXT: fcvt s18, h20
486+
; CHECK-GI-NOFP16-NEXT: fmov s19, w8
484487
; CHECK-GI-NOFP16-NEXT: fcvt s0, h0
485-
; CHECK-GI-NOFP16-NEXT: mov v4.h[4], w8
486-
; CHECK-GI-NOFP16-NEXT: csetm w8, mi
487-
; CHECK-GI-NOFP16-NEXT: fcmp s6, s5
488-
; CHECK-GI-NOFP16-NEXT: mov v4.h[5], w8
489-
; CHECK-GI-NOFP16-NEXT: csetm w8, mi
488+
; CHECK-GI-NOFP16-NEXT: mov v16.16b, v7.16b
489+
; CHECK-GI-NOFP16-NEXT: mov v4.h[3], v6.h[0]
490+
; CHECK-GI-NOFP16-NEXT: cset w8, mi
491+
; CHECK-GI-NOFP16-NEXT: mov v5.h[3], v19.h[0]
492+
; CHECK-GI-NOFP16-NEXT: mov v16.h[1], v7.h[0]
493+
; CHECK-GI-NOFP16-NEXT: fcmp s17, s18
494+
; CHECK-GI-NOFP16-NEXT: fmov s17, w8
495+
; CHECK-GI-NOFP16-NEXT: mov v4.h[4], v6.h[0]
496+
; CHECK-GI-NOFP16-NEXT: mov v5.h[4], v17.h[0]
497+
; CHECK-GI-NOFP16-NEXT: cset w8, mi
490498
; CHECK-GI-NOFP16-NEXT: fcmp s0, s1
491-
; CHECK-GI-NOFP16-NEXT: mov v4.h[6], w8
492-
; CHECK-GI-NOFP16-NEXT: csetm w8, mi
493-
; CHECK-GI-NOFP16-NEXT: mov v4.h[7], w8
494-
; CHECK-GI-NOFP16-NEXT: mov v0.16b, v4.16b
495-
; CHECK-GI-NOFP16-NEXT: bsl v0.16b, v2.16b, v3.16b
499+
; CHECK-GI-NOFP16-NEXT: mov v16.h[2], v7.h[0]
500+
; CHECK-GI-NOFP16-NEXT: fmov s0, w8
501+
; CHECK-GI-NOFP16-NEXT: mov v4.h[5], v6.h[0]
502+
; CHECK-GI-NOFP16-NEXT: cset w8, mi
503+
; CHECK-GI-NOFP16-NEXT: mov v5.h[5], v0.h[0]
504+
; CHECK-GI-NOFP16-NEXT: fmov s0, w8
505+
; CHECK-GI-NOFP16-NEXT: mov v16.h[3], v7.h[0]
506+
; CHECK-GI-NOFP16-NEXT: mov v4.h[6], v6.h[0]
507+
; CHECK-GI-NOFP16-NEXT: mov v5.h[6], v0.h[0]
508+
; CHECK-GI-NOFP16-NEXT: mov v16.h[4], v7.h[0]
509+
; CHECK-GI-NOFP16-NEXT: mov v4.h[7], v0.h[0]
510+
; CHECK-GI-NOFP16-NEXT: mov v5.h[7], v0.h[0]
511+
; CHECK-GI-NOFP16-NEXT: mov v16.h[5], v7.h[0]
512+
; CHECK-GI-NOFP16-NEXT: neg v0.8h, v4.8h
513+
; CHECK-GI-NOFP16-NEXT: ushl v1.8h, v5.8h, v4.8h
514+
; CHECK-GI-NOFP16-NEXT: mov v16.h[6], v7.h[0]
515+
; CHECK-GI-NOFP16-NEXT: mov v16.h[7], v0.h[0]
516+
; CHECK-GI-NOFP16-NEXT: sshl v0.8h, v1.8h, v0.8h
517+
; CHECK-GI-NOFP16-NEXT: eor v1.16b, v0.16b, v16.16b
518+
; CHECK-GI-NOFP16-NEXT: and v0.16b, v2.16b, v0.16b
519+
; CHECK-GI-NOFP16-NEXT: and v1.16b, v3.16b, v1.16b
520+
; CHECK-GI-NOFP16-NEXT: orr v0.16b, v0.16b, v1.16b
496521
; CHECK-GI-NOFP16-NEXT: ret
497522
;
498523
; CHECK-GI-FP16-LABEL: v7f16_half:
@@ -1112,90 +1137,110 @@ define <7 x i32> @v7f16_i32(<7 x half> %a, <7 x half> %b, <7 x i32> %d, <7 x i32
11121137
;
11131138
; CHECK-GI-NOFP16-LABEL: v7f16_i32:
11141139
; CHECK-GI-NOFP16: // %bb.0: // %entry
1115-
; CHECK-GI-NOFP16-NEXT: mov h2, v1.h[1]
1116-
; CHECK-GI-NOFP16-NEXT: mov h3, v0.h[1]
1117-
; CHECK-GI-NOFP16-NEXT: mov h4, v1.h[2]
1118-
; CHECK-GI-NOFP16-NEXT: mov h5, v0.h[2]
1119-
; CHECK-GI-NOFP16-NEXT: fcvt s6, h1
1120-
; CHECK-GI-NOFP16-NEXT: fcvt s7, h0
1140+
; CHECK-GI-NOFP16-NEXT: mov h2, v0.h[1]
1141+
; CHECK-GI-NOFP16-NEXT: mov h3, v1.h[1]
1142+
; CHECK-GI-NOFP16-NEXT: mov w13, #31 // =0x1f
1143+
; CHECK-GI-NOFP16-NEXT: fcvt s4, h0
1144+
; CHECK-GI-NOFP16-NEXT: fcvt s5, h1
1145+
; CHECK-GI-NOFP16-NEXT: ldr s17, [sp, #40]
1146+
; CHECK-GI-NOFP16-NEXT: mov h6, v0.h[2]
1147+
; CHECK-GI-NOFP16-NEXT: mov h7, v1.h[2]
1148+
; CHECK-GI-NOFP16-NEXT: fmov s16, w0
11211149
; CHECK-GI-NOFP16-NEXT: fcvt s2, h2
11221150
; CHECK-GI-NOFP16-NEXT: fcvt s3, h3
1151+
; CHECK-GI-NOFP16-NEXT: fcmp s4, s5
1152+
; CHECK-GI-NOFP16-NEXT: mov h4, v0.h[3]
1153+
; CHECK-GI-NOFP16-NEXT: mov h5, v1.h[3]
1154+
; CHECK-GI-NOFP16-NEXT: fcvt s6, h6
1155+
; CHECK-GI-NOFP16-NEXT: fcvt s7, h7
1156+
; CHECK-GI-NOFP16-NEXT: mov v16.s[1], w1
1157+
; CHECK-GI-NOFP16-NEXT: cset w9, mi
1158+
; CHECK-GI-NOFP16-NEXT: fcmp s2, s3
1159+
; CHECK-GI-NOFP16-NEXT: mov h2, v0.h[4]
1160+
; CHECK-GI-NOFP16-NEXT: mov h3, v1.h[4]
11231161
; CHECK-GI-NOFP16-NEXT: fcvt s4, h4
11241162
; CHECK-GI-NOFP16-NEXT: fcvt s5, h5
1125-
; CHECK-GI-NOFP16-NEXT: fcmp s3, s2
1126-
; CHECK-GI-NOFP16-NEXT: mov h2, v1.h[3]
1127-
; CHECK-GI-NOFP16-NEXT: mov h3, v0.h[3]
1128-
; CHECK-GI-NOFP16-NEXT: csetm w8, mi
1129-
; CHECK-GI-NOFP16-NEXT: fcmp s7, s6
1130-
; CHECK-GI-NOFP16-NEXT: mov h6, v1.h[5]
1131-
; CHECK-GI-NOFP16-NEXT: mov h7, v0.h[5]
1163+
; CHECK-GI-NOFP16-NEXT: mov v16.s[2], w2
1164+
; CHECK-GI-NOFP16-NEXT: cset w10, mi
1165+
; CHECK-GI-NOFP16-NEXT: fcmp s6, s7
1166+
; CHECK-GI-NOFP16-NEXT: mov h6, v0.h[5]
1167+
; CHECK-GI-NOFP16-NEXT: mov h7, v1.h[5]
11321168
; CHECK-GI-NOFP16-NEXT: fcvt s2, h2
1169+
; CHECK-GI-NOFP16-NEXT: mov h0, v0.h[6]
11331170
; CHECK-GI-NOFP16-NEXT: fcvt s3, h3
1134-
; CHECK-GI-NOFP16-NEXT: csetm w9, mi
1135-
; CHECK-GI-NOFP16-NEXT: fcmp s5, s4
1136-
; CHECK-GI-NOFP16-NEXT: mov h4, v1.h[4]
1137-
; CHECK-GI-NOFP16-NEXT: mov h5, v0.h[4]
1138-
; CHECK-GI-NOFP16-NEXT: fcvt s6, h6
1139-
; CHECK-GI-NOFP16-NEXT: fcvt s7, h7
1140-
; CHECK-GI-NOFP16-NEXT: csetm w10, mi
1141-
; CHECK-GI-NOFP16-NEXT: fcmp s3, s2
1142-
; CHECK-GI-NOFP16-NEXT: fcvt s2, h4
1143-
; CHECK-GI-NOFP16-NEXT: mov h4, v1.h[6]
1144-
; CHECK-GI-NOFP16-NEXT: mov h1, v1.h[7]
1145-
; CHECK-GI-NOFP16-NEXT: fcvt s3, h5
1146-
; CHECK-GI-NOFP16-NEXT: mov h5, v0.h[6]
1147-
; CHECK-GI-NOFP16-NEXT: mov h0, v0.h[7]
1148-
; CHECK-GI-NOFP16-NEXT: csetm w11, mi
1149-
; CHECK-GI-NOFP16-NEXT: fcmp s7, s6
1150-
; CHECK-GI-NOFP16-NEXT: fcvt s1, h1
1151-
; CHECK-GI-NOFP16-NEXT: csetm w12, mi
1152-
; CHECK-GI-NOFP16-NEXT: fcmp s3, s2
1153-
; CHECK-GI-NOFP16-NEXT: fcvt s2, h4
1154-
; CHECK-GI-NOFP16-NEXT: fcvt s3, h5
1155-
; CHECK-GI-NOFP16-NEXT: fmov s4, w9
1171+
; CHECK-GI-NOFP16-NEXT: mov h1, v1.h[6]
1172+
; CHECK-GI-NOFP16-NEXT: cset w11, mi
1173+
; CHECK-GI-NOFP16-NEXT: fcmp s4, s5
1174+
; CHECK-GI-NOFP16-NEXT: mov v16.s[3], w3
1175+
; CHECK-GI-NOFP16-NEXT: fcvt s4, h6
1176+
; CHECK-GI-NOFP16-NEXT: ldr s6, [sp, #32]
1177+
; CHECK-GI-NOFP16-NEXT: fcvt s5, h7
11561178
; CHECK-GI-NOFP16-NEXT: fcvt s0, h0
1157-
; CHECK-GI-NOFP16-NEXT: add x9, sp, #8
1158-
; CHECK-GI-NOFP16-NEXT: csetm w13, mi
1159-
; CHECK-GI-NOFP16-NEXT: fmov s5, w13
1160-
; CHECK-GI-NOFP16-NEXT: mov v4.h[1], w8
1161-
; CHECK-GI-NOFP16-NEXT: mov x8, sp
1162-
; CHECK-GI-NOFP16-NEXT: fcmp s3, s2
1163-
; CHECK-GI-NOFP16-NEXT: fmov s2, w7
1164-
; CHECK-GI-NOFP16-NEXT: fmov s3, w0
1165-
; CHECK-GI-NOFP16-NEXT: mov v5.h[1], w12
1166-
; CHECK-GI-NOFP16-NEXT: ld1 { v2.s }[1], [x8]
1167-
; CHECK-GI-NOFP16-NEXT: mov v3.s[1], w1
1168-
; CHECK-GI-NOFP16-NEXT: csetm w8, mi
1169-
; CHECK-GI-NOFP16-NEXT: mov v4.h[2], w10
1179+
; CHECK-GI-NOFP16-NEXT: fmov s7, w4
1180+
; CHECK-GI-NOFP16-NEXT: cset w8, mi
1181+
; CHECK-GI-NOFP16-NEXT: fcmp s2, s3
1182+
; CHECK-GI-NOFP16-NEXT: fcvt s1, h1
1183+
; CHECK-GI-NOFP16-NEXT: fmov s2, w13
1184+
; CHECK-GI-NOFP16-NEXT: mov v7.s[1], w5
1185+
; CHECK-GI-NOFP16-NEXT: cset w12, mi
1186+
; CHECK-GI-NOFP16-NEXT: fcmp s4, s5
1187+
; CHECK-GI-NOFP16-NEXT: ldr s5, [sp]
1188+
; CHECK-GI-NOFP16-NEXT: fmov s3, w12
1189+
; CHECK-GI-NOFP16-NEXT: mov v2.s[1], w13
1190+
; CHECK-GI-NOFP16-NEXT: cset w14, mi
11701191
; CHECK-GI-NOFP16-NEXT: fcmp s0, s1
1171-
; CHECK-GI-NOFP16-NEXT: fmov s1, w4
1172-
; CHECK-GI-NOFP16-NEXT: ldr s0, [sp, #24]
1173-
; CHECK-GI-NOFP16-NEXT: mov v5.h[2], w8
1174-
; CHECK-GI-NOFP16-NEXT: ld1 { v2.s }[2], [x9]
1175-
; CHECK-GI-NOFP16-NEXT: add x9, sp, #32
1176-
; CHECK-GI-NOFP16-NEXT: mov v3.s[2], w2
1177-
; CHECK-GI-NOFP16-NEXT: mov v1.s[1], w5
1178-
; CHECK-GI-NOFP16-NEXT: csetm w8, mi
1179-
; CHECK-GI-NOFP16-NEXT: mov v4.h[3], w11
1180-
; CHECK-GI-NOFP16-NEXT: ld1 { v0.s }[1], [x9]
1181-
; CHECK-GI-NOFP16-NEXT: mov v5.h[3], w8
1182-
; CHECK-GI-NOFP16-NEXT: add x8, sp, #16
1183-
; CHECK-GI-NOFP16-NEXT: ld1 { v2.s }[3], [x8]
1184-
; CHECK-GI-NOFP16-NEXT: mov v3.s[3], w3
1185-
; CHECK-GI-NOFP16-NEXT: add x8, sp, #40
1186-
; CHECK-GI-NOFP16-NEXT: mov v1.s[2], w6
1187-
; CHECK-GI-NOFP16-NEXT: sshll v4.4s, v4.4h, #0
1188-
; CHECK-GI-NOFP16-NEXT: ld1 { v0.s }[2], [x8]
1189-
; CHECK-GI-NOFP16-NEXT: sshll v5.4s, v5.4h, #0
1190-
; CHECK-GI-NOFP16-NEXT: bit v2.16b, v3.16b, v4.16b
1191-
; CHECK-GI-NOFP16-NEXT: bit v0.16b, v1.16b, v5.16b
1192-
; CHECK-GI-NOFP16-NEXT: mov w1, v2.s[1]
1193-
; CHECK-GI-NOFP16-NEXT: mov w2, v2.s[2]
1194-
; CHECK-GI-NOFP16-NEXT: mov w3, v2.s[3]
1195-
; CHECK-GI-NOFP16-NEXT: fmov w0, s2
1196-
; CHECK-GI-NOFP16-NEXT: mov w5, v0.s[1]
1197-
; CHECK-GI-NOFP16-NEXT: mov w6, v0.s[2]
1198-
; CHECK-GI-NOFP16-NEXT: fmov w4, s0
1192+
; CHECK-GI-NOFP16-NEXT: fmov s0, w9
1193+
; CHECK-GI-NOFP16-NEXT: mov v3.s[1], w14
1194+
; CHECK-GI-NOFP16-NEXT: ldr s1, [sp, #24]
1195+
; CHECK-GI-NOFP16-NEXT: mov v7.s[2], w6
1196+
; CHECK-GI-NOFP16-NEXT: mov v2.s[2], w13
1197+
; CHECK-GI-NOFP16-NEXT: cset w9, mi
1198+
; CHECK-GI-NOFP16-NEXT: mov v0.s[1], w10
1199+
; CHECK-GI-NOFP16-NEXT: fmov w10, s5
1200+
; CHECK-GI-NOFP16-NEXT: fmov s5, w7
1201+
; CHECK-GI-NOFP16-NEXT: mov v1.s[1], v6.s[0]
1202+
; CHECK-GI-NOFP16-NEXT: ldr s6, [sp, #8]
1203+
; CHECK-GI-NOFP16-NEXT: mov v3.s[2], w9
1204+
; CHECK-GI-NOFP16-NEXT: mov w9, #-1 // =0xffffffff
1205+
; CHECK-GI-NOFP16-NEXT: fmov s4, w9
1206+
; CHECK-GI-NOFP16-NEXT: mov v2.s[3], w8
1207+
; CHECK-GI-NOFP16-NEXT: mov v0.s[2], w11
1208+
; CHECK-GI-NOFP16-NEXT: mov v5.s[1], w10
1209+
; CHECK-GI-NOFP16-NEXT: mov v1.s[2], v17.s[0]
1210+
; CHECK-GI-NOFP16-NEXT: mov v4.s[1], w9
1211+
; CHECK-GI-NOFP16-NEXT: mov v3.s[3], w8
1212+
; CHECK-GI-NOFP16-NEXT: neg v18.4s, v2.4s
1213+
; CHECK-GI-NOFP16-NEXT: mov v0.s[3], w8
1214+
; CHECK-GI-NOFP16-NEXT: fmov w8, s6
1215+
; CHECK-GI-NOFP16-NEXT: mov v4.s[2], w9
1216+
; CHECK-GI-NOFP16-NEXT: ushl v2.4s, v3.4s, v2.4s
1217+
; CHECK-GI-NOFP16-NEXT: ldr s3, [sp, #16]
1218+
; CHECK-GI-NOFP16-NEXT: mov v5.s[2], w8
1219+
; CHECK-GI-NOFP16-NEXT: mov v7.s[3], w8
1220+
; CHECK-GI-NOFP16-NEXT: shl v0.4s, v0.4s, #31
1221+
; CHECK-GI-NOFP16-NEXT: sshl v2.4s, v2.4s, v18.4s
1222+
; CHECK-GI-NOFP16-NEXT: mov v4.s[3], w8
1223+
; CHECK-GI-NOFP16-NEXT: fmov w8, s3
1224+
; CHECK-GI-NOFP16-NEXT: mov v1.s[3], v0.s[0]
1225+
; CHECK-GI-NOFP16-NEXT: sshr v0.4s, v0.4s, #31
1226+
; CHECK-GI-NOFP16-NEXT: mov v5.s[3], w8
1227+
; CHECK-GI-NOFP16-NEXT: eor v3.16b, v2.16b, v4.16b
1228+
; CHECK-GI-NOFP16-NEXT: and v2.16b, v7.16b, v2.16b
1229+
; CHECK-GI-NOFP16-NEXT: and v1.16b, v1.16b, v3.16b
1230+
; CHECK-GI-NOFP16-NEXT: bsl v0.16b, v16.16b, v5.16b
1231+
; CHECK-GI-NOFP16-NEXT: orr v1.16b, v2.16b, v1.16b
1232+
; CHECK-GI-NOFP16-NEXT: mov s2, v0.s[1]
1233+
; CHECK-GI-NOFP16-NEXT: mov s3, v0.s[2]
1234+
; CHECK-GI-NOFP16-NEXT: mov s4, v0.s[3]
1235+
; CHECK-GI-NOFP16-NEXT: fmov w0, s0
1236+
; CHECK-GI-NOFP16-NEXT: mov s5, v1.s[1]
1237+
; CHECK-GI-NOFP16-NEXT: mov s6, v1.s[2]
1238+
; CHECK-GI-NOFP16-NEXT: fmov w4, s1
1239+
; CHECK-GI-NOFP16-NEXT: fmov w1, s2
1240+
; CHECK-GI-NOFP16-NEXT: fmov w2, s3
1241+
; CHECK-GI-NOFP16-NEXT: fmov w3, s4
1242+
; CHECK-GI-NOFP16-NEXT: fmov w5, s5
1243+
; CHECK-GI-NOFP16-NEXT: fmov w6, s6
11991244
; CHECK-GI-NOFP16-NEXT: ret
12001245
;
12011246
; CHECK-GI-FP16-LABEL: v7f16_i32:

0 commit comments

Comments
 (0)