This is an archive of the discontinued LLVM Phabricator instance.

Differential D111787

[AArch64] Improve fptosi.sat vector lowering
ClosedPublic

Authored by dmgreen on Oct 14 2021, 3:28 AM.

Details

Reviewers
SjoerdMeijer
efriedma
kmclaughlin
samtebbs
t.p.northover
Commits
rGfa1a68285e40: [AArch64] Improve fptosi.sat vector lowering
Summary

Similar to D111236, this improves the lowering of vector fptosi.sat and fptoui.sat, using legal converts and further saturating from there with min/max. f64 are excluded for the moment due to producing worse code in places compared to the unrolling.

Diff Detail

Event Timeline

dmgreen created this revision.Oct 14 2021, 3:28 AM
Herald added subscribers: hiraditya, kristof.beyls. · View Herald TranscriptOct 14 2021, 3:28 AM
dmgreen requested review of this revision.Oct 14 2021, 3:28 AM
Herald added a project: Restricted Project. · View Herald TranscriptOct 14 2021, 3:28 AM
Harbormaster completed remote builds in B128813: Diff 379644.Oct 14 2021, 3:28 AM
samtebbs accepted this revision.Oct 14 2021, 6:36 AM

Very nice. LGTM

This revision is now accepted and ready to land.Oct 14 2021, 6:36 AM
This revision was landed with ongoing or failed builds.Oct 15 2021, 3:38 AM
Closed by commit rGfa1a68285e40: [AArch64] Improve fptosi.sat vector lowering (authored by dmgreen). · Explain Why
This revision was automatically updated to reflect the committed changes.
dmgreen added a commit: rGfa1a68285e40: [AArch64] Improve fptosi.sat vector lowering.

Revision Contents

PathSize
llvm/
lib/
Target/
AArch64/
58 lines
test/
CodeGen/
AArch64/
1654 lines
1280 lines

Diff 379953

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

  • This file is larger than 256 KB, so syntax highlighting is disabled by default.
Show First 20 Lines • Show All 3,416 Lines • ▼ Show 20 Lines assert(SatWidth <= DstElementWidth &&
"Saturation width cannot exceed result width"); "Saturation width cannot exceed result width");
// TODO: Consider lowering to SVE operations, as in LowerVectorFP_TO_INT. // TODO: Consider lowering to SVE operations, as in LowerVectorFP_TO_INT.
// Currently, the `llvm.fpto[su]i.sat.*` instrinsics don't accept scalable // Currently, the `llvm.fpto[su]i.sat.*` instrinsics don't accept scalable
// types, so this is hard to reach. // types, so this is hard to reach.
if (DstVT.isScalableVector()) if (DstVT.isScalableVector())
return SDValue(); return SDValue();
// TODO: Saturate to SatWidth explicitly.
if (SatWidth != DstElementWidth)
return SDValue();
EVT SrcElementVT = SrcVT.getVectorElementType(); EVT SrcElementVT = SrcVT.getVectorElementType();
// In the absence of FP16 support, promote f16 to f32, like // In the absence of FP16 support, promote f16 to f32 and saturate the result.
// LowerVectorFP_TO_INT(). if (SrcElementVT == MVT::f16 &&
if (SrcElementVT == MVT::f16 && !Subtarget->hasFullFP16()) { (!Subtarget->hasFullFP16() || DstElementWidth > 16)) {
MVT F32VT = MVT::getVectorVT(MVT::f32, SrcVT.getVectorNumElements()); MVT F32VT = MVT::getVectorVT(MVT::f32, SrcVT.getVectorNumElements());
return DAG.getNode(Op.getOpcode(), SDLoc(Op), DstVT, SrcVal = DAG.getNode(ISD::FP_EXTEND, SDLoc(Op), F32VT, SrcVal);
DAG.getNode(ISD::FP_EXTEND, SDLoc(Op), F32VT, SrcVal), SrcVT = F32VT;
Op.getOperand(1)); SrcElementVT = MVT::f32;
} SrcElementWidth = 32;
} else if (SrcElementVT != MVT::f64 && SrcElementVT != MVT::f32 &&
SrcElementVT != MVT::f16)
return SDValue();
SDLoc DL(Op);
// Cases that we can emit directly. // Cases that we can emit directly.
if ((SrcElementWidth == DstElementWidth) && if (SrcElementWidth == DstElementWidth && SrcElementWidth == SatWidth)
(SrcElementVT == MVT::f64 || SrcElementVT == MVT::f32 || return DAG.getNode(Op.getOpcode(), DL, DstVT, SrcVal,
(SrcElementVT == MVT::f16 && Subtarget->hasFullFP16()))) { DAG.getValueType(DstVT.getScalarType()));
return Op;
}
// For all other cases, fall back on the expanded form. // Otherwise we emit a cvt that saturates to a higher BW, and saturate the
// result. This is only valid if the legal cvt is larger than the saturate
// width. For double, as we don't have MIN/MAX, it can be simpler to scalarize
// (at least until sqxtn is selected).
if (SrcElementWidth < SatWidth || SrcElementVT == MVT::f64)
return SDValue(); return SDValue();
EVT IntVT = SrcVT.changeVectorElementTypeToInteger();
SDValue NativeCvt = DAG.getNode(Op.getOpcode(), DL, IntVT, SrcVal,
DAG.getValueType(IntVT.getScalarType()));
SDValue Sat;
if (Op.getOpcode() == ISD::FP_TO_SINT_SAT) {
SDValue MinC = DAG.getConstant(
APInt::getSignedMaxValue(SatWidth).sextOrSelf(SrcElementWidth), DL,
IntVT);
SDValue Min = DAG.getNode(ISD::SMIN, DL, IntVT, NativeCvt, MinC);
SDValue MaxC = DAG.getConstant(
APInt::getSignedMinValue(SatWidth).sextOrSelf(SrcElementWidth), DL,
IntVT);
Sat = DAG.getNode(ISD::SMAX, DL, IntVT, Min, MaxC);
} else {
SDValue MinC = DAG.getConstant(
APInt::getAllOnesValue(SatWidth).zextOrSelf(SrcElementWidth), DL,
IntVT);
Sat = DAG.getNode(ISD::UMIN, DL, IntVT, NativeCvt, MinC);
}
return DAG.getNode(ISD::TRUNCATE, DL, DstVT, Sat);
} }
SDValue AArch64TargetLowering::LowerFP_TO_INT_SAT(SDValue Op, SDValue AArch64TargetLowering::LowerFP_TO_INT_SAT(SDValue Op,
SelectionDAG &DAG) const { SelectionDAG &DAG) const {
// AArch64 FP-to-int conversions saturate to the destination register size, so // AArch64 FP-to-int conversions saturate to the destination register size, so
// we can lower common saturating conversions to simple instructions. // we can lower common saturating conversions to simple instructions.
SDValue SrcVal = Op.getOperand(0); SDValue SrcVal = Op.getOperand(0);
EVT SrcVT = SrcVal.getValueType(); EVT SrcVT = SrcVal.getValueType();
▲ Show 20 Lines • Show All 15,638 Lines • Show Last 20 Lines

llvm/test/CodeGen/AArch64/fptosi-sat-vector.ll

Show First 20 Lines • Show All 576 Lines • ▼ Show 20 Lines
; CHECK-FP16-NEXT: fcvtzs w8, h0 ; CHECK-FP16-NEXT: fcvtzs w8, h0
; CHECK-FP16-NEXT: fmov s0, w8 ; CHECK-FP16-NEXT: fmov s0, w8
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <1 x i32> @llvm.fptosi.sat.v1f16.v1i32(<1 x half> %f) %x = call <1 x i32> @llvm.fptosi.sat.v1f16.v1i32(<1 x half> %f)
ret <1 x i32> %x ret <1 x i32> %x
} }
define <2 x i32> @test_signed_v2f16_v2i32(<2 x half> %f) { define <2 x i32> @test_signed_v2f16_v2i32(<2 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v2f16_v2i32: ; CHECK-LABEL: test_signed_v2f16_v2i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: fcvtzs v0.4s, v0.4s ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 killed $q0 ; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_signed_v2f16_v2i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0
; CHECK-FP16-NEXT: mov h1, v0.h[1]
; CHECK-FP16-NEXT: fcvtzs w8, h0
; CHECK-FP16-NEXT: fmov s0, w8
; CHECK-FP16-NEXT: fcvtzs w8, h1
; CHECK-FP16-NEXT: mov v0.s[1], w8
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-FP16-NEXT: ret
%x = call <2 x i32> @llvm.fptosi.sat.v2f16.v2i32(<2 x half> %f) %x = call <2 x i32> @llvm.fptosi.sat.v2f16.v2i32(<2 x half> %f)
ret <2 x i32> %x ret <2 x i32> %x
} }
define <3 x i32> @test_signed_v3f16_v3i32(<3 x half> %f) { define <3 x i32> @test_signed_v3f16_v3i32(<3 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v3f16_v3i32: ; CHECK-LABEL: test_signed_v3f16_v3i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: fcvtzs v0.4s, v0.4s ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_signed_v3f16_v3i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0
; CHECK-FP16-NEXT: mov h2, v0.h[1]
; CHECK-FP16-NEXT: fcvtzs w8, h0
; CHECK-FP16-NEXT: fmov s1, w8
; CHECK-FP16-NEXT: fcvtzs w8, h2
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: mov v1.s[1], w8
; CHECK-FP16-NEXT: fcvtzs w8, h2
; CHECK-FP16-NEXT: mov v1.s[2], w8
; CHECK-FP16-NEXT: fcvtzs w8, h0
; CHECK-FP16-NEXT: mov v1.s[3], w8
; CHECK-FP16-NEXT: mov v0.16b, v1.16b
; CHECK-FP16-NEXT: ret
%x = call <3 x i32> @llvm.fptosi.sat.v3f16.v3i32(<3 x half> %f) %x = call <3 x i32> @llvm.fptosi.sat.v3f16.v3i32(<3 x half> %f)
ret <3 x i32> %x ret <3 x i32> %x
} }
define <4 x i32> @test_signed_v4f16_v4i32(<4 x half> %f) { define <4 x i32> @test_signed_v4f16_v4i32(<4 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v4f16_v4i32: ; CHECK-LABEL: test_signed_v4f16_v4i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: fcvtzs v0.4s, v0.4s ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_signed_v4f16_v4i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0
; CHECK-FP16-NEXT: mov h2, v0.h[1]
; CHECK-FP16-NEXT: fcvtzs w8, h0
; CHECK-FP16-NEXT: fmov s1, w8
; CHECK-FP16-NEXT: fcvtzs w8, h2
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: mov v1.s[1], w8
; CHECK-FP16-NEXT: fcvtzs w8, h2
; CHECK-FP16-NEXT: mov v1.s[2], w8
; CHECK-FP16-NEXT: fcvtzs w8, h0
; CHECK-FP16-NEXT: mov v1.s[3], w8
; CHECK-FP16-NEXT: mov v0.16b, v1.16b
; CHECK-FP16-NEXT: ret
%x = call <4 x i32> @llvm.fptosi.sat.v4f16.v4i32(<4 x half> %f) %x = call <4 x i32> @llvm.fptosi.sat.v4f16.v4i32(<4 x half> %f)
ret <4 x i32> %x ret <4 x i32> %x
} }
define <5 x i32> @test_signed_v5f16_v5i32(<5 x half> %f) { define <5 x i32> @test_signed_v5f16_v5i32(<5 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v5f16_v5i32: ; CHECK-LABEL: test_signed_v5f16_v5i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v1.4s, v0.4h ; CHECK-NEXT: fcvtl v1.4s, v0.4h
; CHECK-CVT-NEXT: fcvtl2 v0.4s, v0.8h ; CHECK-NEXT: fcvtl2 v0.4s, v0.8h
; CHECK-CVT-NEXT: fcvtzs v1.4s, v1.4s ; CHECK-NEXT: fcvtzs v1.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzs v0.4s, v0.4s ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: mov w1, v1.s[1] ; CHECK-NEXT: mov w1, v1.s[1]
; CHECK-CVT-NEXT: mov w2, v1.s[2] ; CHECK-NEXT: mov w2, v1.s[2]
; CHECK-CVT-NEXT: mov w3, v1.s[3] ; CHECK-NEXT: mov w3, v1.s[3]
; CHECK-CVT-NEXT: fmov w0, s1 ; CHECK-NEXT: fmov w0, s1
; CHECK-CVT-NEXT: fmov w4, s0 ; CHECK-NEXT: fmov w4, s0
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_signed_v5f16_v5i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: mov h1, v0.h[1]
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: ext v3.16b, v0.16b, v0.16b, #8
; CHECK-FP16-NEXT: mov h4, v0.h[3]
; CHECK-FP16-NEXT: fcvtzs w0, h0
; CHECK-FP16-NEXT: fcvtzs w1, h1
; CHECK-FP16-NEXT: fcvtzs w2, h2
; CHECK-FP16-NEXT: fcvtzs w4, h3
; CHECK-FP16-NEXT: fcvtzs w3, h4
; CHECK-FP16-NEXT: ret
%x = call <5 x i32> @llvm.fptosi.sat.v5f16.v5i32(<5 x half> %f) %x = call <5 x i32> @llvm.fptosi.sat.v5f16.v5i32(<5 x half> %f)
ret <5 x i32> %x ret <5 x i32> %x
} }
define <6 x i32> @test_signed_v6f16_v6i32(<6 x half> %f) { define <6 x i32> @test_signed_v6f16_v6i32(<6 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v6f16_v6i32: ; CHECK-LABEL: test_signed_v6f16_v6i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v1.4s, v0.4h ; CHECK-NEXT: fcvtl v1.4s, v0.4h
; CHECK-CVT-NEXT: fcvtl2 v0.4s, v0.8h ; CHECK-NEXT: fcvtl2 v0.4s, v0.8h
; CHECK-CVT-NEXT: fcvtzs v1.4s, v1.4s ; CHECK-NEXT: fcvtzs v1.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzs v0.4s, v0.4s ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: mov w1, v1.s[1] ; CHECK-NEXT: mov w1, v1.s[1]
; CHECK-CVT-NEXT: mov w2, v1.s[2] ; CHECK-NEXT: mov w2, v1.s[2]
; CHECK-CVT-NEXT: mov w3, v1.s[3] ; CHECK-NEXT: mov w3, v1.s[3]
; CHECK-CVT-NEXT: mov w5, v0.s[1] ; CHECK-NEXT: mov w5, v0.s[1]
; CHECK-CVT-NEXT: fmov w0, s1 ; CHECK-NEXT: fmov w0, s1
; CHECK-CVT-NEXT: fmov w4, s0 ; CHECK-NEXT: fmov w4, s0
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_signed_v6f16_v6i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: ext v1.16b, v0.16b, v0.16b, #8
; CHECK-FP16-NEXT: mov h3, v0.h[2]
; CHECK-FP16-NEXT: mov h4, v0.h[3]
; CHECK-FP16-NEXT: fcvtzs w0, h0
; CHECK-FP16-NEXT: mov h2, v1.h[1]
; CHECK-FP16-NEXT: fcvtzs w8, h1
; CHECK-FP16-NEXT: fcvtzs w2, h3
; CHECK-FP16-NEXT: fcvtzs w3, h4
; CHECK-FP16-NEXT: fmov s1, w8
; CHECK-FP16-NEXT: fcvtzs w5, h2
; CHECK-FP16-NEXT: mov h2, v0.h[1]
; CHECK-FP16-NEXT: mov v1.s[1], w5
; CHECK-FP16-NEXT: fcvtzs w1, h2
; CHECK-FP16-NEXT: fmov w4, s1
; CHECK-FP16-NEXT: ret
%x = call <6 x i32> @llvm.fptosi.sat.v6f16.v6i32(<6 x half> %f) %x = call <6 x i32> @llvm.fptosi.sat.v6f16.v6i32(<6 x half> %f)
ret <6 x i32> %x ret <6 x i32> %x
} }
define <7 x i32> @test_signed_v7f16_v7i32(<7 x half> %f) { define <7 x i32> @test_signed_v7f16_v7i32(<7 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v7f16_v7i32: ; CHECK-LABEL: test_signed_v7f16_v7i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v1.4s, v0.4h ; CHECK-NEXT: fcvtl v1.4s, v0.4h
; CHECK-CVT-NEXT: fcvtl2 v0.4s, v0.8h ; CHECK-NEXT: fcvtl2 v0.4s, v0.8h
; CHECK-CVT-NEXT: fcvtzs v1.4s, v1.4s ; CHECK-NEXT: fcvtzs v1.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzs v0.4s, v0.4s ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: mov w1, v1.s[1] ; CHECK-NEXT: mov w1, v1.s[1]
; CHECK-CVT-NEXT: mov w2, v1.s[2] ; CHECK-NEXT: mov w2, v1.s[2]
; CHECK-CVT-NEXT: mov w3, v1.s[3] ; CHECK-NEXT: mov w3, v1.s[3]
; CHECK-CVT-NEXT: mov w5, v0.s[1] ; CHECK-NEXT: mov w5, v0.s[1]
; CHECK-CVT-NEXT: mov w6, v0.s[2] ; CHECK-NEXT: mov w6, v0.s[2]
; CHECK-CVT-NEXT: fmov w0, s1 ; CHECK-NEXT: fmov w0, s1
; CHECK-CVT-NEXT: fmov w4, s0 ; CHECK-NEXT: fmov w4, s0
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_signed_v7f16_v7i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: ext v1.16b, v0.16b, v0.16b, #8
; CHECK-FP16-NEXT: mov h4, v0.h[3]
; CHECK-FP16-NEXT: fcvtzs w0, h0
; CHECK-FP16-NEXT: mov h2, v1.h[1]
; CHECK-FP16-NEXT: fcvtzs w8, h1
; CHECK-FP16-NEXT: mov h1, v1.h[2]
; CHECK-FP16-NEXT: fcvtzs w3, h4
; CHECK-FP16-NEXT: fmov s3, w8
; CHECK-FP16-NEXT: fcvtzs w8, h2
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: fcvtzs w6, h1
; CHECK-FP16-NEXT: mov h1, v0.h[1]
; CHECK-FP16-NEXT: mov v3.s[1], w8
; CHECK-FP16-NEXT: fcvtzs w2, h2
; CHECK-FP16-NEXT: fcvtzs w1, h1
; CHECK-FP16-NEXT: mov v3.s[2], w6
; CHECK-FP16-NEXT: mov w5, v3.s[1]
; CHECK-FP16-NEXT: fmov w4, s3
; CHECK-FP16-NEXT: ret
%x = call <7 x i32> @llvm.fptosi.sat.v7f16.v7i32(<7 x half> %f) %x = call <7 x i32> @llvm.fptosi.sat.v7f16.v7i32(<7 x half> %f)
ret <7 x i32> %x ret <7 x i32> %x
} }
define <8 x i32> @test_signed_v8f16_v8i32(<8 x half> %f) { define <8 x i32> @test_signed_v8f16_v8i32(<8 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v8f16_v8i32: ; CHECK-LABEL: test_signed_v8f16_v8i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl2 v1.4s, v0.8h ; CHECK-NEXT: fcvtl2 v1.4s, v0.8h
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: fcvtzs v1.4s, v1.4s ; CHECK-NEXT: fcvtzs v1.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzs v0.4s, v0.4s ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_signed_v8f16_v8i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: ext v3.16b, v0.16b, v0.16b, #8
; CHECK-FP16-NEXT: mov h4, v0.h[1]
; CHECK-FP16-NEXT: fcvtzs w9, h0
; CHECK-FP16-NEXT: mov h2, v3.h[1]
; CHECK-FP16-NEXT: fcvtzs w8, h3
; CHECK-FP16-NEXT: mov h5, v3.h[2]
; CHECK-FP16-NEXT: mov h3, v3.h[3]
; CHECK-FP16-NEXT: fmov s1, w8
; CHECK-FP16-NEXT: fcvtzs w8, h2
; CHECK-FP16-NEXT: fmov s2, w9
; CHECK-FP16-NEXT: fcvtzs w9, h4
; CHECK-FP16-NEXT: mov h4, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: mov v1.s[1], w8
; CHECK-FP16-NEXT: fcvtzs w8, h5
; CHECK-FP16-NEXT: mov v2.s[1], w9
; CHECK-FP16-NEXT: fcvtzs w9, h4
; CHECK-FP16-NEXT: mov v1.s[2], w8
; CHECK-FP16-NEXT: fcvtzs w8, h3
; CHECK-FP16-NEXT: mov v2.s[2], w9
; CHECK-FP16-NEXT: fcvtzs w9, h0
; CHECK-FP16-NEXT: mov v1.s[3], w8
; CHECK-FP16-NEXT: mov v2.s[3], w9
; CHECK-FP16-NEXT: mov v0.16b, v2.16b
; CHECK-FP16-NEXT: ret
%x = call <8 x i32> @llvm.fptosi.sat.v8f16.v8i32(<8 x half> %f) %x = call <8 x i32> @llvm.fptosi.sat.v8f16.v8i32(<8 x half> %f)
ret <8 x i32> %x ret <8 x i32> %x
} }
; ;
; 2-Vector float to signed integer -- result size variation ; 2-Vector float to signed integer -- result size variation
; ;
declare <2 x i1> @llvm.fptosi.sat.v2f32.v2i1 (<2 x float>) declare <2 x i1> @llvm.fptosi.sat.v2f32.v2i1 (<2 x float>)
declare <2 x i8> @llvm.fptosi.sat.v2f32.v2i8 (<2 x float>) declare <2 x i8> @llvm.fptosi.sat.v2f32.v2i8 (<2 x float>)
declare <2 x i13> @llvm.fptosi.sat.v2f32.v2i13 (<2 x float>) declare <2 x i13> @llvm.fptosi.sat.v2f32.v2i13 (<2 x float>)
declare <2 x i16> @llvm.fptosi.sat.v2f32.v2i16 (<2 x float>) declare <2 x i16> @llvm.fptosi.sat.v2f32.v2i16 (<2 x float>)
declare <2 x i19> @llvm.fptosi.sat.v2f32.v2i19 (<2 x float>) declare <2 x i19> @llvm.fptosi.sat.v2f32.v2i19 (<2 x float>)
declare <2 x i50> @llvm.fptosi.sat.v2f32.v2i50 (<2 x float>) declare <2 x i50> @llvm.fptosi.sat.v2f32.v2i50 (<2 x float>)
declare <2 x i64> @llvm.fptosi.sat.v2f32.v2i64 (<2 x float>) declare <2 x i64> @llvm.fptosi.sat.v2f32.v2i64 (<2 x float>)
declare <2 x i100> @llvm.fptosi.sat.v2f32.v2i100(<2 x float>) declare <2 x i100> @llvm.fptosi.sat.v2f32.v2i100(<2 x float>)
declare <2 x i128> @llvm.fptosi.sat.v2f32.v2i128(<2 x float>) declare <2 x i128> @llvm.fptosi.sat.v2f32.v2i128(<2 x float>)
define <2 x i1> @test_signed_v2f32_v2i1(<2 x float> %f) { define <2 x i1> @test_signed_v2f32_v2i1(<2 x float> %f) {
; CHECK-LABEL: test_signed_v2f32_v2i1: ; CHECK-LABEL: test_signed_v2f32_v2i1:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-NEXT: movi v1.2d, #0000000000000000
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: fcvtzs v0.2s, v0.2s
; CHECK-NEXT: fcvtzs w9, s0 ; CHECK-NEXT: smin v0.2s, v0.2s, v1.2s
; CHECK-NEXT: fcvtzs w8, s1 ; CHECK-NEXT: movi v1.2d, #0xffffffffffffffff
; CHECK-NEXT: cmp w8, #0 ; CHECK-NEXT: smax v0.2s, v0.2s, v1.2s
; CHECK-NEXT: csel w8, w8, wzr, lt
; CHECK-NEXT: cmp w8, #0
; CHECK-NEXT: csinv w8, w8, wzr, ge
; CHECK-NEXT: cmp w9, #0
; CHECK-NEXT: csel w9, w9, wzr, lt
; CHECK-NEXT: cmp w9, #0
; CHECK-NEXT: csinv w9, w9, wzr, ge
; CHECK-NEXT: fmov s0, w9
; CHECK-NEXT: mov v0.s[1], w8
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <2 x i1> @llvm.fptosi.sat.v2f32.v2i1(<2 x float> %f) %x = call <2 x i1> @llvm.fptosi.sat.v2f32.v2i1(<2 x float> %f)
ret <2 x i1> %x ret <2 x i1> %x
} }
define <2 x i8> @test_signed_v2f32_v2i8(<2 x float> %f) { define <2 x i8> @test_signed_v2f32_v2i8(<2 x float> %f) {
; CHECK-LABEL: test_signed_v2f32_v2i8: ; CHECK-LABEL: test_signed_v2f32_v2i8:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-NEXT: movi v1.2s, #127
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: fcvtzs v0.2s, v0.2s
; CHECK-NEXT: mov w8, #127 ; CHECK-NEXT: smin v0.2s, v0.2s, v1.2s
; CHECK-NEXT: fcvtzs w10, s0 ; CHECK-NEXT: mvni v1.2s, #127
; CHECK-NEXT: mov w11, #-128 ; CHECK-NEXT: smax v0.2s, v0.2s, v1.2s
; CHECK-NEXT: fcvtzs w9, s1
; CHECK-NEXT: cmp w9, #127
; CHECK-NEXT: csel w9, w9, w8, lt
; CHECK-NEXT: cmn w9, #128
; CHECK-NEXT: csel w9, w9, w11, gt
; CHECK-NEXT: cmp w10, #127
; CHECK-NEXT: csel w8, w10, w8, lt
; CHECK-NEXT: cmn w8, #128
; CHECK-NEXT: csel w8, w8, w11, gt
; CHECK-NEXT: fmov s0, w8
; CHECK-NEXT: mov v0.s[1], w9
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <2 x i8> @llvm.fptosi.sat.v2f32.v2i8(<2 x float> %f) %x = call <2 x i8> @llvm.fptosi.sat.v2f32.v2i8(<2 x float> %f)
ret <2 x i8> %x ret <2 x i8> %x
} }
define <2 x i13> @test_signed_v2f32_v2i13(<2 x float> %f) { define <2 x i13> @test_signed_v2f32_v2i13(<2 x float> %f) {
; CHECK-LABEL: test_signed_v2f32_v2i13: ; CHECK-LABEL: test_signed_v2f32_v2i13:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-NEXT: movi v1.2s, #15, msl #8
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: fcvtzs v0.2s, v0.2s
; CHECK-NEXT: mov w8, #4095 ; CHECK-NEXT: smin v0.2s, v0.2s, v1.2s
; CHECK-NEXT: fcvtzs w10, s0 ; CHECK-NEXT: mvni v1.2s, #15, msl #8
; CHECK-NEXT: mov w11, #-4096 ; CHECK-NEXT: smax v0.2s, v0.2s, v1.2s
; CHECK-NEXT: fcvtzs w9, s1
; CHECK-NEXT: cmp w9, #4095
; CHECK-NEXT: csel w9, w9, w8, lt
; CHECK-NEXT: cmn w9, #1, lsl #12 // =4096
; CHECK-NEXT: csel w9, w9, w11, gt
; CHECK-NEXT: cmp w10, #4095
; CHECK-NEXT: csel w8, w10, w8, lt
; CHECK-NEXT: cmn w8, #1, lsl #12 // =4096
; CHECK-NEXT: csel w8, w8, w11, gt
; CHECK-NEXT: fmov s0, w8
; CHECK-NEXT: mov v0.s[1], w9
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <2 x i13> @llvm.fptosi.sat.v2f32.v2i13(<2 x float> %f) %x = call <2 x i13> @llvm.fptosi.sat.v2f32.v2i13(<2 x float> %f)
ret <2 x i13> %x ret <2 x i13> %x
} }
define <2 x i16> @test_signed_v2f32_v2i16(<2 x float> %f) { define <2 x i16> @test_signed_v2f32_v2i16(<2 x float> %f) {
; CHECK-LABEL: test_signed_v2f32_v2i16: ; CHECK-LABEL: test_signed_v2f32_v2i16:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-NEXT: movi v1.2s, #127, msl #8
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: fcvtzs v0.2s, v0.2s
; CHECK-NEXT: mov w8, #32767 ; CHECK-NEXT: smin v0.2s, v0.2s, v1.2s
; CHECK-NEXT: fcvtzs w10, s0 ; CHECK-NEXT: mvni v1.2s, #127, msl #8
; CHECK-NEXT: mov w11, #-32768 ; CHECK-NEXT: smax v0.2s, v0.2s, v1.2s
; CHECK-NEXT: fcvtzs w9, s1
; CHECK-NEXT: cmp w9, w8
; CHECK-NEXT: csel w9, w9, w8, lt
; CHECK-NEXT: cmn w9, #8, lsl #12 // =32768
; CHECK-NEXT: csel w9, w9, w11, gt
; CHECK-NEXT: cmp w10, w8
; CHECK-NEXT: csel w8, w10, w8, lt
; CHECK-NEXT: cmn w8, #8, lsl #12 // =32768
; CHECK-NEXT: csel w8, w8, w11, gt
; CHECK-NEXT: fmov s0, w8
; CHECK-NEXT: mov v0.s[1], w9
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <2 x i16> @llvm.fptosi.sat.v2f32.v2i16(<2 x float> %f) %x = call <2 x i16> @llvm.fptosi.sat.v2f32.v2i16(<2 x float> %f)
ret <2 x i16> %x ret <2 x i16> %x
} }
define <2 x i19> @test_signed_v2f32_v2i19(<2 x float> %f) { define <2 x i19> @test_signed_v2f32_v2i19(<2 x float> %f) {
; CHECK-LABEL: test_signed_v2f32_v2i19: ; CHECK-LABEL: test_signed_v2f32_v2i19:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-NEXT: movi v1.2s, #3, msl #16
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: fcvtzs v0.2s, v0.2s
; CHECK-NEXT: mov w8, #262143 ; CHECK-NEXT: smin v0.2s, v0.2s, v1.2s
; CHECK-NEXT: fcvtzs w10, s0 ; CHECK-NEXT: mvni v1.2s, #3, msl #16
; CHECK-NEXT: mov w11, #-262144 ; CHECK-NEXT: smax v0.2s, v0.2s, v1.2s
; CHECK-NEXT: fcvtzs w9, s1
; CHECK-NEXT: cmp w9, w8
; CHECK-NEXT: csel w9, w9, w8, lt
; CHECK-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-NEXT: csel w9, w9, w11, gt
; CHECK-NEXT: cmp w10, w8
; CHECK-NEXT: csel w8, w10, w8, lt
; CHECK-NEXT: cmn w8, #64, lsl #12 // =262144
; CHECK-NEXT: csel w8, w8, w11, gt
; CHECK-NEXT: fmov s0, w8
; CHECK-NEXT: mov v0.s[1], w9
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <2 x i19> @llvm.fptosi.sat.v2f32.v2i19(<2 x float> %f) %x = call <2 x i19> @llvm.fptosi.sat.v2f32.v2i19(<2 x float> %f)
ret <2 x i19> %x ret <2 x i19> %x
} }
define <2 x i32> @test_signed_v2f32_v2i32_duplicate(<2 x float> %f) { define <2 x i32> @test_signed_v2f32_v2i32_duplicate(<2 x float> %f) {
; CHECK-LABEL: test_signed_v2f32_v2i32_duplicate: ; CHECK-LABEL: test_signed_v2f32_v2i32_duplicate:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
▲ Show 20 Lines • Show All 187 Lines • ▼ Show 20 Lines
declare <4 x i50> @llvm.fptosi.sat.v4f32.v4i50 (<4 x float>) declare <4 x i50> @llvm.fptosi.sat.v4f32.v4i50 (<4 x float>)
declare <4 x i64> @llvm.fptosi.sat.v4f32.v4i64 (<4 x float>) declare <4 x i64> @llvm.fptosi.sat.v4f32.v4i64 (<4 x float>)
declare <4 x i100> @llvm.fptosi.sat.v4f32.v4i100(<4 x float>) declare <4 x i100> @llvm.fptosi.sat.v4f32.v4i100(<4 x float>)
declare <4 x i128> @llvm.fptosi.sat.v4f32.v4i128(<4 x float>) declare <4 x i128> @llvm.fptosi.sat.v4f32.v4i128(<4 x float>)
define <4 x i1> @test_signed_v4f32_v4i1(<4 x float> %f) { define <4 x i1> @test_signed_v4f32_v4i1(<4 x float> %f) {
; CHECK-LABEL: test_signed_v4f32_v4i1: ; CHECK-LABEL: test_signed_v4f32_v4i1:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: movi v1.2d, #0000000000000000
; CHECK-NEXT: fcvtzs w9, s0 ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-NEXT: fcvtzs w8, s1 ; CHECK-NEXT: smin v0.4s, v0.4s, v1.4s
; CHECK-NEXT: mov s1, v0.s[2] ; CHECK-NEXT: movi v1.2d, #0xffffffffffffffff
; CHECK-NEXT: cmp w8, #0 ; CHECK-NEXT: smax v0.4s, v0.4s, v1.4s
; CHECK-NEXT: csel w8, w8, wzr, lt ; CHECK-NEXT: xtn v0.4h, v0.4s
; CHECK-NEXT: fcvtzs w10, s1
; CHECK-NEXT: cmp w8, #0
; CHECK-NEXT: mov s1, v0.s[3]
; CHECK-NEXT: csinv w8, w8, wzr, ge
; CHECK-NEXT: cmp w9, #0
; CHECK-NEXT: csel w9, w9, wzr, lt
; CHECK-NEXT: cmp w9, #0
; CHECK-NEXT: csinv w9, w9, wzr, ge
; CHECK-NEXT: cmp w10, #0
; CHECK-NEXT: fmov s0, w9
; CHECK-NEXT: csel w9, w10, wzr, lt
; CHECK-NEXT: fcvtzs w10, s1
; CHECK-NEXT: cmp w9, #0
; CHECK-NEXT: mov v0.h[1], w8
; CHECK-NEXT: csinv w8, w9, wzr, ge
; CHECK-NEXT: cmp w10, #0
; CHECK-NEXT: csel w9, w10, wzr, lt
; CHECK-NEXT: cmp w9, #0
; CHECK-NEXT: mov v0.h[2], w8
; CHECK-NEXT: csinv w8, w9, wzr, ge
; CHECK-NEXT: mov v0.h[3], w8
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <4 x i1> @llvm.fptosi.sat.v4f32.v4i1(<4 x float> %f) %x = call <4 x i1> @llvm.fptosi.sat.v4f32.v4i1(<4 x float> %f)
ret <4 x i1> %x ret <4 x i1> %x
} }
define <4 x i8> @test_signed_v4f32_v4i8(<4 x float> %f) { define <4 x i8> @test_signed_v4f32_v4i8(<4 x float> %f) {
; CHECK-LABEL: test_signed_v4f32_v4i8: ; CHECK-LABEL: test_signed_v4f32_v4i8:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: movi v1.4s, #127
; CHECK-NEXT: mov w8, #127 ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-NEXT: fcvtzs w10, s0 ; CHECK-NEXT: smin v0.4s, v0.4s, v1.4s
; CHECK-NEXT: mov w11, #-128 ; CHECK-NEXT: mvni v1.4s, #127
; CHECK-NEXT: fcvtzs w9, s1 ; CHECK-NEXT: smax v0.4s, v0.4s, v1.4s
; CHECK-NEXT: mov s1, v0.s[2] ; CHECK-NEXT: xtn v0.4h, v0.4s
; CHECK-NEXT: cmp w9, #127
; CHECK-NEXT: csel w9, w9, w8, lt
; CHECK-NEXT: fcvtzs w12, s1
; CHECK-NEXT: cmn w9, #128
; CHECK-NEXT: mov s1, v0.s[3]
; CHECK-NEXT: csel w9, w9, w11, gt
; CHECK-NEXT: cmp w10, #127
; CHECK-NEXT: csel w10, w10, w8, lt
; CHECK-NEXT: cmn w10, #128
; CHECK-NEXT: csel w10, w10, w11, gt
; CHECK-NEXT: cmp w12, #127
; CHECK-NEXT: fmov s0, w10
; CHECK-NEXT: csel w10, w12, w8, lt
; CHECK-NEXT: fcvtzs w12, s1
; CHECK-NEXT: cmn w10, #128
; CHECK-NEXT: mov v0.h[1], w9
; CHECK-NEXT: csel w9, w10, w11, gt
; CHECK-NEXT: cmp w12, #127
; CHECK-NEXT: csel w8, w12, w8, lt
; CHECK-NEXT: cmn w8, #128
; CHECK-NEXT: mov v0.h[2], w9
; CHECK-NEXT: csel w8, w8, w11, gt
; CHECK-NEXT: mov v0.h[3], w8
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <4 x i8> @llvm.fptosi.sat.v4f32.v4i8(<4 x float> %f) %x = call <4 x i8> @llvm.fptosi.sat.v4f32.v4i8(<4 x float> %f)
ret <4 x i8> %x ret <4 x i8> %x
} }
define <4 x i13> @test_signed_v4f32_v4i13(<4 x float> %f) { define <4 x i13> @test_signed_v4f32_v4i13(<4 x float> %f) {
; CHECK-LABEL: test_signed_v4f32_v4i13: ; CHECK-LABEL: test_signed_v4f32_v4i13:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: movi v1.4s, #15, msl #8
; CHECK-NEXT: mov w8, #4095 ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-NEXT: fcvtzs w10, s0 ; CHECK-NEXT: smin v0.4s, v0.4s, v1.4s
; CHECK-NEXT: mov w11, #-4096 ; CHECK-NEXT: mvni v1.4s, #15, msl #8
; CHECK-NEXT: fcvtzs w9, s1 ; CHECK-NEXT: smax v0.4s, v0.4s, v1.4s
; CHECK-NEXT: mov s1, v0.s[2] ; CHECK-NEXT: xtn v0.4h, v0.4s
; CHECK-NEXT: cmp w9, #4095
; CHECK-NEXT: csel w9, w9, w8, lt
; CHECK-NEXT: fcvtzs w12, s1
; CHECK-NEXT: cmn w9, #1, lsl #12 // =4096
; CHECK-NEXT: mov s1, v0.s[3]
; CHECK-NEXT: csel w9, w9, w11, gt
; CHECK-NEXT: cmp w10, #4095
; CHECK-NEXT: csel w10, w10, w8, lt
; CHECK-NEXT: cmn w10, #1, lsl #12 // =4096
; CHECK-NEXT: csel w10, w10, w11, gt
; CHECK-NEXT: cmp w12, #4095
; CHECK-NEXT: fmov s0, w10
; CHECK-NEXT: csel w10, w12, w8, lt
; CHECK-NEXT: fcvtzs w12, s1
; CHECK-NEXT: cmn w10, #1, lsl #12 // =4096
; CHECK-NEXT: mov v0.h[1], w9
; CHECK-NEXT: csel w9, w10, w11, gt
; CHECK-NEXT: cmp w12, #4095
; CHECK-NEXT: csel w8, w12, w8, lt
; CHECK-NEXT: cmn w8, #1, lsl #12 // =4096
; CHECK-NEXT: mov v0.h[2], w9
; CHECK-NEXT: csel w8, w8, w11, gt
; CHECK-NEXT: mov v0.h[3], w8
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <4 x i13> @llvm.fptosi.sat.v4f32.v4i13(<4 x float> %f) %x = call <4 x i13> @llvm.fptosi.sat.v4f32.v4i13(<4 x float> %f)
ret <4 x i13> %x ret <4 x i13> %x
} }
define <4 x i16> @test_signed_v4f32_v4i16(<4 x float> %f) { define <4 x i16> @test_signed_v4f32_v4i16(<4 x float> %f) {
; CHECK-LABEL: test_signed_v4f32_v4i16: ; CHECK-LABEL: test_signed_v4f32_v4i16:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-NEXT: mov w8, #32767 ; CHECK-NEXT: sqxtn v0.4h, v0.4s
; CHECK-NEXT: fcvtzs w10, s0
; CHECK-NEXT: mov w11, #-32768
; CHECK-NEXT: fcvtzs w9, s1
; CHECK-NEXT: mov s1, v0.s[2]
; CHECK-NEXT: cmp w9, w8
; CHECK-NEXT: csel w9, w9, w8, lt
; CHECK-NEXT: fcvtzs w12, s1
; CHECK-NEXT: cmn w9, #8, lsl #12 // =32768
; CHECK-NEXT: mov s1, v0.s[3]
; CHECK-NEXT: csel w9, w9, w11, gt
; CHECK-NEXT: cmp w10, w8
; CHECK-NEXT: csel w10, w10, w8, lt
; CHECK-NEXT: cmn w10, #8, lsl #12 // =32768
; CHECK-NEXT: csel w10, w10, w11, gt
; CHECK-NEXT: cmp w12, w8
; CHECK-NEXT: fmov s0, w10
; CHECK-NEXT: csel w10, w12, w8, lt
; CHECK-NEXT: fcvtzs w12, s1
; CHECK-NEXT: cmn w10, #8, lsl #12 // =32768
; CHECK-NEXT: mov v0.h[1], w9
; CHECK-NEXT: csel w9, w10, w11, gt
; CHECK-NEXT: cmp w12, w8
; CHECK-NEXT: csel w8, w12, w8, lt
; CHECK-NEXT: cmn w8, #8, lsl #12 // =32768
; CHECK-NEXT: mov v0.h[2], w9
; CHECK-NEXT: csel w8, w8, w11, gt
; CHECK-NEXT: mov v0.h[3], w8
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <4 x i16> @llvm.fptosi.sat.v4f32.v4i16(<4 x float> %f) %x = call <4 x i16> @llvm.fptosi.sat.v4f32.v4i16(<4 x float> %f)
ret <4 x i16> %x ret <4 x i16> %x
} }
define <4 x i19> @test_signed_v4f32_v4i19(<4 x float> %f) { define <4 x i19> @test_signed_v4f32_v4i19(<4 x float> %f) {
; CHECK-LABEL: test_signed_v4f32_v4i19: ; CHECK-LABEL: test_signed_v4f32_v4i19:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: movi v1.4s, #3, msl #16
; CHECK-NEXT: mov w8, #262143 ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-NEXT: fcvtzs w10, s0 ; CHECK-NEXT: smin v0.4s, v0.4s, v1.4s
; CHECK-NEXT: mov w11, #-262144 ; CHECK-NEXT: mvni v1.4s, #3, msl #16
; CHECK-NEXT: fcvtzs w9, s1 ; CHECK-NEXT: smax v0.4s, v0.4s, v1.4s
; CHECK-NEXT: mov s1, v0.s[2]
; CHECK-NEXT: cmp w9, w8
; CHECK-NEXT: csel w9, w9, w8, lt
; CHECK-NEXT: fcvtzs w12, s1
; CHECK-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-NEXT: mov s1, v0.s[3]
; CHECK-NEXT: csel w9, w9, w11, gt
; CHECK-NEXT: cmp w10, w8
; CHECK-NEXT: csel w10, w10, w8, lt
; CHECK-NEXT: cmn w10, #64, lsl #12 // =262144
; CHECK-NEXT: csel w10, w10, w11, gt
; CHECK-NEXT: cmp w12, w8
; CHECK-NEXT: fmov s0, w10
; CHECK-NEXT: csel w10, w12, w8, lt
; CHECK-NEXT: fcvtzs w12, s1
; CHECK-NEXT: cmn w10, #64, lsl #12 // =262144
; CHECK-NEXT: mov v0.s[1], w9
; CHECK-NEXT: csel w9, w10, w11, gt
; CHECK-NEXT: cmp w12, w8
; CHECK-NEXT: csel w8, w12, w8, lt
; CHECK-NEXT: cmn w8, #64, lsl #12 // =262144
; CHECK-NEXT: mov v0.s[2], w9
; CHECK-NEXT: csel w8, w8, w11, gt
; CHECK-NEXT: mov v0.s[3], w8
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <4 x i19> @llvm.fptosi.sat.v4f32.v4i19(<4 x float> %f) %x = call <4 x i19> @llvm.fptosi.sat.v4f32.v4i19(<4 x float> %f)
ret <4 x i19> %x ret <4 x i19> %x
} }
define <4 x i32> @test_signed_v4f32_v4i32_duplicate(<4 x float> %f) { define <4 x i32> @test_signed_v4f32_v4i32_duplicate(<4 x float> %f) {
; CHECK-LABEL: test_signed_v4f32_v4i32_duplicate: ; CHECK-LABEL: test_signed_v4f32_v4i32_duplicate:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
▲ Show 20 Lines • Show All 587 Lines • ▼ Show 20 Lines
declare <4 x i50> @llvm.fptosi.sat.v4f16.v4i50 (<4 x half>) declare <4 x i50> @llvm.fptosi.sat.v4f16.v4i50 (<4 x half>)
declare <4 x i64> @llvm.fptosi.sat.v4f16.v4i64 (<4 x half>) declare <4 x i64> @llvm.fptosi.sat.v4f16.v4i64 (<4 x half>)
declare <4 x i100> @llvm.fptosi.sat.v4f16.v4i100(<4 x half>) declare <4 x i100> @llvm.fptosi.sat.v4f16.v4i100(<4 x half>)
declare <4 x i128> @llvm.fptosi.sat.v4f16.v4i128(<4 x half>) declare <4 x i128> @llvm.fptosi.sat.v4f16.v4i128(<4 x half>)
define <4 x i1> @test_signed_v4f16_v4i1(<4 x half> %f) { define <4 x i1> @test_signed_v4f16_v4i1(<4 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v4f16_v4i1: ; CHECK-CVT-LABEL: test_signed_v4f16_v4i1:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov h1, v0.h[1] ; CHECK-CVT-NEXT: movi v1.2d, #0000000000000000
; CHECK-CVT-NEXT: fcvt s2, h0 ; CHECK-CVT-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-CVT-NEXT: smin v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzs w9, s2 ; CHECK-CVT-NEXT: movi v1.2d, #0xffffffffffffffff
; CHECK-CVT-NEXT: fcvtzs w8, s1 ; CHECK-CVT-NEXT: smax v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: mov h1, v0.h[2] ; CHECK-CVT-NEXT: xtn v0.4h, v0.4s
; CHECK-CVT-NEXT: mov h0, v0.h[3]
; CHECK-CVT-NEXT: cmp w8, #0
; CHECK-CVT-NEXT: csel w8, w8, wzr, lt
; CHECK-CVT-NEXT: fcvt s1, h1
; CHECK-CVT-NEXT: cmp w8, #0
; CHECK-CVT-NEXT: csinv w8, w8, wzr, ge
; CHECK-CVT-NEXT: cmp w9, #0
; CHECK-CVT-NEXT: csel w9, w9, wzr, lt
; CHECK-CVT-NEXT: cmp w9, #0
; CHECK-CVT-NEXT: fcvtzs w10, s1
; CHECK-CVT-NEXT: csinv w9, w9, wzr, ge
; CHECK-CVT-NEXT: fcvt s1, h0
; CHECK-CVT-NEXT: cmp w10, #0
; CHECK-CVT-NEXT: fmov s0, w9
; CHECK-CVT-NEXT: csel w9, w10, wzr, lt
; CHECK-CVT-NEXT: fcvtzs w10, s1
; CHECK-CVT-NEXT: cmp w9, #0
; CHECK-CVT-NEXT: mov v0.h[1], w8
; CHECK-CVT-NEXT: csinv w8, w9, wzr, ge
; CHECK-CVT-NEXT: cmp w10, #0
; CHECK-CVT-NEXT: csel w9, w10, wzr, lt
; CHECK-CVT-NEXT: cmp w9, #0
; CHECK-CVT-NEXT: mov v0.h[2], w8
; CHECK-CVT-NEXT: csinv w8, w9, wzr, ge
; CHECK-CVT-NEXT: mov v0.h[3], w8
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_signed_v4f16_v4i1: ; CHECK-FP16-LABEL: test_signed_v4f16_v4i1:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-FP16-NEXT: movi v1.2d, #0000000000000000
; CHECK-FP16-NEXT: mov h1, v0.h[1] ; CHECK-FP16-NEXT: fcvtzs v0.4h, v0.4h
; CHECK-FP16-NEXT: fcvtzs w9, h0 ; CHECK-FP16-NEXT: smin v0.4h, v0.4h, v1.4h
; CHECK-FP16-NEXT: fcvtzs w8, h1 ; CHECK-FP16-NEXT: movi v1.2d, #0xffffffffffffffff
; CHECK-FP16-NEXT: mov h1, v0.h[2] ; CHECK-FP16-NEXT: smax v0.4h, v0.4h, v1.4h
; CHECK-FP16-NEXT: cmp w8, #0
; CHECK-FP16-NEXT: csel w8, w8, wzr, lt
; CHECK-FP16-NEXT: fcvtzs w10, h1
; CHECK-FP16-NEXT: cmp w8, #0
; CHECK-FP16-NEXT: mov h1, v0.h[3]
; CHECK-FP16-NEXT: csinv w8, w8, wzr, ge
; CHECK-FP16-NEXT: cmp w9, #0
; CHECK-FP16-NEXT: csel w9, w9, wzr, lt
; CHECK-FP16-NEXT: cmp w9, #0
; CHECK-FP16-NEXT: csinv w9, w9, wzr, ge
; CHECK-FP16-NEXT: cmp w10, #0
; CHECK-FP16-NEXT: fmov s0, w9
; CHECK-FP16-NEXT: csel w9, w10, wzr, lt
; CHECK-FP16-NEXT: fcvtzs w10, h1
; CHECK-FP16-NEXT: cmp w9, #0
; CHECK-FP16-NEXT: mov v0.h[1], w8
; CHECK-FP16-NEXT: csinv w8, w9, wzr, ge
; CHECK-FP16-NEXT: cmp w10, #0
; CHECK-FP16-NEXT: csel w9, w10, wzr, lt
; CHECK-FP16-NEXT: cmp w9, #0
; CHECK-FP16-NEXT: mov v0.h[2], w8
; CHECK-FP16-NEXT: csinv w8, w9, wzr, ge
; CHECK-FP16-NEXT: mov v0.h[3], w8
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <4 x i1> @llvm.fptosi.sat.v4f16.v4i1(<4 x half> %f) %x = call <4 x i1> @llvm.fptosi.sat.v4f16.v4i1(<4 x half> %f)
ret <4 x i1> %x ret <4 x i1> %x
} }
define <4 x i8> @test_signed_v4f16_v4i8(<4 x half> %f) { define <4 x i8> @test_signed_v4f16_v4i8(<4 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v4f16_v4i8: ; CHECK-CVT-LABEL: test_signed_v4f16_v4i8:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov h1, v0.h[1] ; CHECK-CVT-NEXT: movi v1.4s, #127
; CHECK-CVT-NEXT: fcvt s2, h0 ; CHECK-CVT-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: mov w8, #127 ; CHECK-CVT-NEXT: smin v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: mov w11, #-128 ; CHECK-CVT-NEXT: mvni v1.4s, #127
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-CVT-NEXT: smax v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzs w10, s2 ; CHECK-CVT-NEXT: xtn v0.4h, v0.4s
; CHECK-CVT-NEXT: fcvtzs w9, s1
; CHECK-CVT-NEXT: mov h1, v0.h[2]
; CHECK-CVT-NEXT: mov h0, v0.h[3]
; CHECK-CVT-NEXT: cmp w9, #127
; CHECK-CVT-NEXT: csel w9, w9, w8, lt
; CHECK-CVT-NEXT: fcvt s1, h1
; CHECK-CVT-NEXT: cmn w9, #128
; CHECK-CVT-NEXT: csel w9, w9, w11, gt
; CHECK-CVT-NEXT: cmp w10, #127
; CHECK-CVT-NEXT: csel w10, w10, w8, lt
; CHECK-CVT-NEXT: cmn w10, #128
; CHECK-CVT-NEXT: fcvtzs w12, s1
; CHECK-CVT-NEXT: csel w10, w10, w11, gt
; CHECK-CVT-NEXT: fcvt s1, h0
; CHECK-CVT-NEXT: cmp w12, #127
; CHECK-CVT-NEXT: fmov s0, w10
; CHECK-CVT-NEXT: csel w10, w12, w8, lt
; CHECK-CVT-NEXT: fcvtzs w12, s1
; CHECK-CVT-NEXT: cmn w10, #128
; CHECK-CVT-NEXT: mov v0.h[1], w9
; CHECK-CVT-NEXT: csel w9, w10, w11, gt
; CHECK-CVT-NEXT: cmp w12, #127
; CHECK-CVT-NEXT: csel w8, w12, w8, lt
; CHECK-CVT-NEXT: cmn w8, #128
; CHECK-CVT-NEXT: mov v0.h[2], w9
; CHECK-CVT-NEXT: csel w8, w8, w11, gt
; CHECK-CVT-NEXT: mov v0.h[3], w8
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_signed_v4f16_v4i8: ; CHECK-FP16-LABEL: test_signed_v4f16_v4i8:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-FP16-NEXT: movi v1.4h, #127
; CHECK-FP16-NEXT: mov h1, v0.h[1] ; CHECK-FP16-NEXT: fcvtzs v0.4h, v0.4h
; CHECK-FP16-NEXT: mov w8, #127 ; CHECK-FP16-NEXT: smin v0.4h, v0.4h, v1.4h
; CHECK-FP16-NEXT: fcvtzs w10, h0 ; CHECK-FP16-NEXT: mvni v1.4h, #127
; CHECK-FP16-NEXT: mov w11, #-128 ; CHECK-FP16-NEXT: smax v0.4h, v0.4h, v1.4h
; CHECK-FP16-NEXT: fcvtzs w9, h1
; CHECK-FP16-NEXT: mov h1, v0.h[2]
; CHECK-FP16-NEXT: cmp w9, #127
; CHECK-FP16-NEXT: csel w9, w9, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h1
; CHECK-FP16-NEXT: cmn w9, #128
; CHECK-FP16-NEXT: mov h1, v0.h[3]
; CHECK-FP16-NEXT: csel w9, w9, w11, gt
; CHECK-FP16-NEXT: cmp w10, #127
; CHECK-FP16-NEXT: csel w10, w10, w8, lt
; CHECK-FP16-NEXT: cmn w10, #128
; CHECK-FP16-NEXT: csel w10, w10, w11, gt
; CHECK-FP16-NEXT: cmp w12, #127
; CHECK-FP16-NEXT: fmov s0, w10
; CHECK-FP16-NEXT: csel w10, w12, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h1
; CHECK-FP16-NEXT: cmn w10, #128
; CHECK-FP16-NEXT: mov v0.h[1], w9
; CHECK-FP16-NEXT: csel w9, w10, w11, gt
; CHECK-FP16-NEXT: cmp w12, #127
; CHECK-FP16-NEXT: csel w8, w12, w8, lt
; CHECK-FP16-NEXT: cmn w8, #128
; CHECK-FP16-NEXT: mov v0.h[2], w9
; CHECK-FP16-NEXT: csel w8, w8, w11, gt
; CHECK-FP16-NEXT: mov v0.h[3], w8
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <4 x i8> @llvm.fptosi.sat.v4f16.v4i8(<4 x half> %f) %x = call <4 x i8> @llvm.fptosi.sat.v4f16.v4i8(<4 x half> %f)
ret <4 x i8> %x ret <4 x i8> %x
} }
define <4 x i13> @test_signed_v4f16_v4i13(<4 x half> %f) { define <4 x i13> @test_signed_v4f16_v4i13(<4 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v4f16_v4i13: ; CHECK-CVT-LABEL: test_signed_v4f16_v4i13:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov h1, v0.h[1] ; CHECK-CVT-NEXT: movi v1.4s, #15, msl #8
; CHECK-CVT-NEXT: fcvt s2, h0 ; CHECK-CVT-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: mov w8, #4095 ; CHECK-CVT-NEXT: smin v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: mov w11, #-4096 ; CHECK-CVT-NEXT: mvni v1.4s, #15, msl #8
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-CVT-NEXT: smax v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzs w10, s2 ; CHECK-CVT-NEXT: xtn v0.4h, v0.4s
; CHECK-CVT-NEXT: fcvtzs w9, s1
; CHECK-CVT-NEXT: mov h1, v0.h[2]
; CHECK-CVT-NEXT: mov h0, v0.h[3]
; CHECK-CVT-NEXT: cmp w9, #4095
; CHECK-CVT-NEXT: csel w9, w9, w8, lt
; CHECK-CVT-NEXT: fcvt s1, h1
; CHECK-CVT-NEXT: cmn w9, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: csel w9, w9, w11, gt
; CHECK-CVT-NEXT: cmp w10, #4095
; CHECK-CVT-NEXT: csel w10, w10, w8, lt
; CHECK-CVT-NEXT: cmn w10, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: fcvtzs w12, s1
; CHECK-CVT-NEXT: csel w10, w10, w11, gt
; CHECK-CVT-NEXT: fcvt s1, h0
; CHECK-CVT-NEXT: cmp w12, #4095
; CHECK-CVT-NEXT: fmov s0, w10
; CHECK-CVT-NEXT: csel w10, w12, w8, lt
; CHECK-CVT-NEXT: fcvtzs w12, s1
; CHECK-CVT-NEXT: cmn w10, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: mov v0.h[1], w9
; CHECK-CVT-NEXT: csel w9, w10, w11, gt
; CHECK-CVT-NEXT: cmp w12, #4095
; CHECK-CVT-NEXT: csel w8, w12, w8, lt
; CHECK-CVT-NEXT: cmn w8, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: mov v0.h[2], w9
; CHECK-CVT-NEXT: csel w8, w8, w11, gt
; CHECK-CVT-NEXT: mov v0.h[3], w8
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_signed_v4f16_v4i13: ; CHECK-FP16-LABEL: test_signed_v4f16_v4i13:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-FP16-NEXT: mvni v1.4h, #240, lsl #8
; CHECK-FP16-NEXT: mov h1, v0.h[1] ; CHECK-FP16-NEXT: fcvtzs v0.4h, v0.4h
; CHECK-FP16-NEXT: mov w8, #4095 ; CHECK-FP16-NEXT: smin v0.4h, v0.4h, v1.4h
; CHECK-FP16-NEXT: fcvtzs w10, h0 ; CHECK-FP16-NEXT: movi v1.4h, #240, lsl #8
; CHECK-FP16-NEXT: mov w11, #-4096 ; CHECK-FP16-NEXT: smax v0.4h, v0.4h, v1.4h
; CHECK-FP16-NEXT: fcvtzs w9, h1
; CHECK-FP16-NEXT: mov h1, v0.h[2]
; CHECK-FP16-NEXT: cmp w9, #4095
; CHECK-FP16-NEXT: csel w9, w9, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h1
; CHECK-FP16-NEXT: cmn w9, #1, lsl #12 // =4096
; CHECK-FP16-NEXT: mov h1, v0.h[3]
; CHECK-FP16-NEXT: csel w9, w9, w11, gt
; CHECK-FP16-NEXT: cmp w10, #4095
; CHECK-FP16-NEXT: csel w10, w10, w8, lt
; CHECK-FP16-NEXT: cmn w10, #1, lsl #12 // =4096
; CHECK-FP16-NEXT: csel w10, w10, w11, gt
; CHECK-FP16-NEXT: cmp w12, #4095
; CHECK-FP16-NEXT: fmov s0, w10
; CHECK-FP16-NEXT: csel w10, w12, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h1
; CHECK-FP16-NEXT: cmn w10, #1, lsl #12 // =4096
; CHECK-FP16-NEXT: mov v0.h[1], w9
; CHECK-FP16-NEXT: csel w9, w10, w11, gt
; CHECK-FP16-NEXT: cmp w12, #4095
; CHECK-FP16-NEXT: csel w8, w12, w8, lt
; CHECK-FP16-NEXT: cmn w8, #1, lsl #12 // =4096
; CHECK-FP16-NEXT: mov v0.h[2], w9
; CHECK-FP16-NEXT: csel w8, w8, w11, gt
; CHECK-FP16-NEXT: mov v0.h[3], w8
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <4 x i13> @llvm.fptosi.sat.v4f16.v4i13(<4 x half> %f) %x = call <4 x i13> @llvm.fptosi.sat.v4f16.v4i13(<4 x half> %f)
ret <4 x i13> %x ret <4 x i13> %x
} }
define <4 x i16> @test_signed_v4f16_v4i16(<4 x half> %f) { define <4 x i16> @test_signed_v4f16_v4i16(<4 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v4f16_v4i16: ; CHECK-CVT-LABEL: test_signed_v4f16_v4i16:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov w8, #32767 ; CHECK-CVT-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: mov w11, #-32768 ; CHECK-CVT-NEXT: sqxtn v0.4h, v0.4s
; CHECK-CVT-NEXT: mov s1, v0.s[1]
; CHECK-CVT-NEXT: fcvtzs w10, s0
; CHECK-CVT-NEXT: fcvtzs w9, s1
; CHECK-CVT-NEXT: mov s1, v0.s[2]
; CHECK-CVT-NEXT: cmp w9, w8
; CHECK-CVT-NEXT: csel w9, w9, w8, lt
; CHECK-CVT-NEXT: fcvtzs w12, s1
; CHECK-CVT-NEXT: cmn w9, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: mov s1, v0.s[3]
; CHECK-CVT-NEXT: csel w9, w9, w11, gt
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w10, w10, w8, lt
; CHECK-CVT-NEXT: cmn w10, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: csel w10, w10, w11, gt
; CHECK-CVT-NEXT: cmp w12, w8
; CHECK-CVT-NEXT: fmov s0, w10
; CHECK-CVT-NEXT: csel w10, w12, w8, lt
; CHECK-CVT-NEXT: fcvtzs w12, s1
; CHECK-CVT-NEXT: cmn w10, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: mov v0.h[1], w9
; CHECK-CVT-NEXT: csel w9, w10, w11, gt
; CHECK-CVT-NEXT: cmp w12, w8
; CHECK-CVT-NEXT: csel w8, w12, w8, lt
; CHECK-CVT-NEXT: cmn w8, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: mov v0.h[2], w9
; CHECK-CVT-NEXT: csel w8, w8, w11, gt
; CHECK-CVT-NEXT: mov v0.h[3], w8
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_signed_v4f16_v4i16: ; CHECK-FP16-LABEL: test_signed_v4f16_v4i16:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: fcvtzs v0.4h, v0.4h ; CHECK-FP16-NEXT: fcvtzs v0.4h, v0.4h
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <4 x i16> @llvm.fptosi.sat.v4f16.v4i16(<4 x half> %f) %x = call <4 x i16> @llvm.fptosi.sat.v4f16.v4i16(<4 x half> %f)
ret <4 x i16> %x ret <4 x i16> %x
} }
define <4 x i19> @test_signed_v4f16_v4i19(<4 x half> %f) { define <4 x i19> @test_signed_v4f16_v4i19(<4 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v4f16_v4i19: ; CHECK-LABEL: test_signed_v4f16_v4i19:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov h1, v0.h[1] ; CHECK-NEXT: movi v1.4s, #3, msl #16
; CHECK-CVT-NEXT: fcvt s2, h0 ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: mov w8, #262143 ; CHECK-NEXT: smin v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: mov w11, #-262144 ; CHECK-NEXT: mvni v1.4s, #3, msl #16
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-NEXT: smax v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzs w10, s2 ; CHECK-NEXT: ret
; CHECK-CVT-NEXT: fcvtzs w9, s1
; CHECK-CVT-NEXT: mov h1, v0.h[2]
; CHECK-CVT-NEXT: mov h0, v0.h[3]
; CHECK-CVT-NEXT: cmp w9, w8
; CHECK-CVT-NEXT: csel w9, w9, w8, lt
; CHECK-CVT-NEXT: fcvt s1, h1
; CHECK-CVT-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-CVT-NEXT: csel w9, w9, w11, gt
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w10, w10, w8, lt
; CHECK-CVT-NEXT: cmn w10, #64, lsl #12 // =262144
; CHECK-CVT-NEXT: fcvtzs w12, s1
; CHECK-CVT-NEXT: csel w10, w10, w11, gt
; CHECK-CVT-NEXT: fcvt s1, h0
; CHECK-CVT-NEXT: cmp w12, w8
; CHECK-CVT-NEXT: fmov s0, w10
; CHECK-CVT-NEXT: csel w10, w12, w8, lt
; CHECK-CVT-NEXT: fcvtzs w12, s1
; CHECK-CVT-NEXT: cmn w10, #64, lsl #12 // =262144
; CHECK-CVT-NEXT: mov v0.s[1], w9
; CHECK-CVT-NEXT: csel w9, w10, w11, gt
; CHECK-CVT-NEXT: cmp w12, w8
; CHECK-CVT-NEXT: csel w8, w12, w8, lt
; CHECK-CVT-NEXT: cmn w8, #64, lsl #12 // =262144
; CHECK-CVT-NEXT: mov v0.s[2], w9
; CHECK-CVT-NEXT: csel w8, w8, w11, gt
; CHECK-CVT-NEXT: mov v0.s[3], w8
; CHECK-CVT-NEXT: ret
;
; CHECK-FP16-LABEL: test_signed_v4f16_v4i19:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0
; CHECK-FP16-NEXT: mov h1, v0.h[1]
; CHECK-FP16-NEXT: mov w8, #262143
; CHECK-FP16-NEXT: fcvtzs w10, h0
; CHECK-FP16-NEXT: mov w11, #-262144
; CHECK-FP16-NEXT: fcvtzs w9, h1
; CHECK-FP16-NEXT: mov h1, v0.h[2]
; CHECK-FP16-NEXT: cmp w9, w8
; CHECK-FP16-NEXT: csel w9, w9, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h1
; CHECK-FP16-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-FP16-NEXT: mov h1, v0.h[3]
; CHECK-FP16-NEXT: csel w9, w9, w11, gt
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w10, w10, w8, lt
; CHECK-FP16-NEXT: cmn w10, #64, lsl #12 // =262144
; CHECK-FP16-NEXT: csel w10, w10, w11, gt
; CHECK-FP16-NEXT: cmp w12, w8
; CHECK-FP16-NEXT: fmov s0, w10
; CHECK-FP16-NEXT: csel w10, w12, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h1
; CHECK-FP16-NEXT: cmn w10, #64, lsl #12 // =262144
; CHECK-FP16-NEXT: mov v0.s[1], w9
; CHECK-FP16-NEXT: csel w9, w10, w11, gt
; CHECK-FP16-NEXT: cmp w12, w8
; CHECK-FP16-NEXT: csel w8, w12, w8, lt
; CHECK-FP16-NEXT: cmn w8, #64, lsl #12 // =262144
; CHECK-FP16-NEXT: mov v0.s[2], w9
; CHECK-FP16-NEXT: csel w8, w8, w11, gt
; CHECK-FP16-NEXT: mov v0.s[3], w8
; CHECK-FP16-NEXT: ret
%x = call <4 x i19> @llvm.fptosi.sat.v4f16.v4i19(<4 x half> %f) %x = call <4 x i19> @llvm.fptosi.sat.v4f16.v4i19(<4 x half> %f)
ret <4 x i19> %x ret <4 x i19> %x
} }
define <4 x i32> @test_signed_v4f16_v4i32_duplicate(<4 x half> %f) { define <4 x i32> @test_signed_v4f16_v4i32_duplicate(<4 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v4f16_v4i32_duplicate: ; CHECK-LABEL: test_signed_v4f16_v4i32_duplicate:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: fcvtzs v0.4s, v0.4s ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_signed_v4f16_v4i32_duplicate:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0
; CHECK-FP16-NEXT: mov h2, v0.h[1]
; CHECK-FP16-NEXT: fcvtzs w8, h0
; CHECK-FP16-NEXT: fmov s1, w8
; CHECK-FP16-NEXT: fcvtzs w8, h2
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: mov v1.s[1], w8
; CHECK-FP16-NEXT: fcvtzs w8, h2
; CHECK-FP16-NEXT: mov v1.s[2], w8
; CHECK-FP16-NEXT: fcvtzs w8, h0
; CHECK-FP16-NEXT: mov v1.s[3], w8
; CHECK-FP16-NEXT: mov v0.16b, v1.16b
; CHECK-FP16-NEXT: ret
%x = call <4 x i32> @llvm.fptosi.sat.v4f16.v4i32(<4 x half> %f) %x = call <4 x i32> @llvm.fptosi.sat.v4f16.v4i32(<4 x half> %f)
ret <4 x i32> %x ret <4 x i32> %x
} }
define <4 x i50> @test_signed_v4f16_v4i50(<4 x half> %f) { define <4 x i50> @test_signed_v4f16_v4i50(<4 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v4f16_v4i50: ; CHECK-CVT-LABEL: test_signed_v4f16_v4i50:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 def $q0
▲ Show 20 Lines • Show All 329 Lines • ▼ Show 20 Lines
declare <8 x i50> @llvm.fptosi.sat.v8f16.v8i50 (<8 x half>) declare <8 x i50> @llvm.fptosi.sat.v8f16.v8i50 (<8 x half>)
declare <8 x i64> @llvm.fptosi.sat.v8f16.v8i64 (<8 x half>) declare <8 x i64> @llvm.fptosi.sat.v8f16.v8i64 (<8 x half>)
declare <8 x i100> @llvm.fptosi.sat.v8f16.v8i100(<8 x half>) declare <8 x i100> @llvm.fptosi.sat.v8f16.v8i100(<8 x half>)
declare <8 x i128> @llvm.fptosi.sat.v8f16.v8i128(<8 x half>) declare <8 x i128> @llvm.fptosi.sat.v8f16.v8i128(<8 x half>)
define <8 x i1> @test_signed_v8f16_v8i1(<8 x half> %f) { define <8 x i1> @test_signed_v8f16_v8i1(<8 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v8f16_v8i1: ; CHECK-CVT-LABEL: test_signed_v8f16_v8i1:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: mov h1, v0.h[1] ; CHECK-CVT-NEXT: fcvtl2 v1.4s, v0.8h
; CHECK-CVT-NEXT: fcvt s2, h0 ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov h3, v0.h[4] ; CHECK-CVT-NEXT: mov s2, v1.s[1]
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-CVT-NEXT: fcvtzs w9, s1
; CHECK-CVT-NEXT: fcvtzs w9, s2 ; CHECK-CVT-NEXT: fcvtzs w13, s0
; CHECK-CVT-NEXT: mov h2, v0.h[3] ; CHECK-CVT-NEXT: fcvtzs w8, s2
; CHECK-CVT-NEXT: fcvtzs w8, s1 ; CHECK-CVT-NEXT: mov s2, v1.s[2]
; CHECK-CVT-NEXT: mov h1, v0.h[2] ; CHECK-CVT-NEXT: mov s1, v1.s[3]
; CHECK-CVT-NEXT: fcvt s2, h2
; CHECK-CVT-NEXT: cmp w8, #0 ; CHECK-CVT-NEXT: cmp w8, #0
; CHECK-CVT-NEXT: csel w8, w8, wzr, lt ; CHECK-CVT-NEXT: csel w8, w8, wzr, lt
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-CVT-NEXT: fcvtzs w10, s2
; CHECK-CVT-NEXT: cmp w8, #0 ; CHECK-CVT-NEXT: cmp w8, #0
; CHECK-CVT-NEXT: mov s2, v0.s[1]
; CHECK-CVT-NEXT: csinv w8, w8, wzr, ge ; CHECK-CVT-NEXT: csinv w8, w8, wzr, ge
; CHECK-CVT-NEXT: cmp w9, #0 ; CHECK-CVT-NEXT: cmp w9, #0
; CHECK-CVT-NEXT: csel w9, w9, wzr, lt ; CHECK-CVT-NEXT: csel w9, w9, wzr, lt
; CHECK-CVT-NEXT: fcvtzs w11, s1
; CHECK-CVT-NEXT: cmp w9, #0 ; CHECK-CVT-NEXT: cmp w9, #0
; CHECK-CVT-NEXT: fcvtzs w10, s1 ; CHECK-CVT-NEXT: mov s1, v0.s[2]
; CHECK-CVT-NEXT: csinv w9, w9, wzr, ge ; CHECK-CVT-NEXT: csinv w9, w9, wzr, ge
; CHECK-CVT-NEXT: cmp w10, #0 ; CHECK-CVT-NEXT: cmp w10, #0
; CHECK-CVT-NEXT: fmov s1, w9
; CHECK-CVT-NEXT: fcvtzs w9, s2
; CHECK-CVT-NEXT: fcvt s2, h3
; CHECK-CVT-NEXT: mov h3, v0.h[5]
; CHECK-CVT-NEXT: csel w10, w10, wzr, lt ; CHECK-CVT-NEXT: csel w10, w10, wzr, lt
; CHECK-CVT-NEXT: mov v1.b[1], w8 ; CHECK-CVT-NEXT: fcvtzs w12, s2
; CHECK-CVT-NEXT: cmp w10, #0
; CHECK-CVT-NEXT: csinv w8, w10, wzr, ge
; CHECK-CVT-NEXT: cmp w9, #0
; CHECK-CVT-NEXT: fcvtzs w10, s2
; CHECK-CVT-NEXT: fcvt s2, h3
; CHECK-CVT-NEXT: csel w9, w9, wzr, lt
; CHECK-CVT-NEXT: mov h3, v0.h[6]
; CHECK-CVT-NEXT: mov v1.b[2], w8
; CHECK-CVT-NEXT: cmp w9, #0
; CHECK-CVT-NEXT: csinv w8, w9, wzr, ge
; CHECK-CVT-NEXT: cmp w10, #0
; CHECK-CVT-NEXT: fcvtzs w9, s2
; CHECK-CVT-NEXT: csel w10, w10, wzr, lt
; CHECK-CVT-NEXT: fcvt s2, h3
; CHECK-CVT-NEXT: mov h0, v0.h[7]
; CHECK-CVT-NEXT: mov v1.b[3], w8
; CHECK-CVT-NEXT: cmp w10, #0
; CHECK-CVT-NEXT: csinv w8, w10, wzr, ge
; CHECK-CVT-NEXT: cmp w9, #0
; CHECK-CVT-NEXT: csel w9, w9, wzr, lt
; CHECK-CVT-NEXT: fcvtzs w10, s2
; CHECK-CVT-NEXT: fcvt s0, h0
; CHECK-CVT-NEXT: mov v1.b[4], w8
; CHECK-CVT-NEXT: cmp w9, #0
; CHECK-CVT-NEXT: csinv w8, w9, wzr, ge
; CHECK-CVT-NEXT: cmp w10, #0
; CHECK-CVT-NEXT: csel w9, w10, wzr, lt
; CHECK-CVT-NEXT: fcvtzs w10, s0
; CHECK-CVT-NEXT: mov v1.b[5], w8
; CHECK-CVT-NEXT: cmp w9, #0
; CHECK-CVT-NEXT: csinv w8, w9, wzr, ge
; CHECK-CVT-NEXT: cmp w10, #0 ; CHECK-CVT-NEXT: cmp w10, #0
; CHECK-CVT-NEXT: csel w9, w10, wzr, lt ; CHECK-CVT-NEXT: mov s0, v0.s[3]
; CHECK-CVT-NEXT: mov v1.b[6], w8 ; CHECK-CVT-NEXT: csinv w10, w10, wzr, ge
; CHECK-CVT-NEXT: cmp w11, #0
; CHECK-CVT-NEXT: csel w11, w11, wzr, lt
; CHECK-CVT-NEXT: fcvtzs w14, s1
; CHECK-CVT-NEXT: cmp w11, #0
; CHECK-CVT-NEXT: fmov s2, w9
; CHECK-CVT-NEXT: csinv w11, w11, wzr, ge
; CHECK-CVT-NEXT: cmp w12, #0
; CHECK-CVT-NEXT: csel w12, w12, wzr, lt
; CHECK-CVT-NEXT: cmp w12, #0
; CHECK-CVT-NEXT: csinv w12, w12, wzr, ge
; CHECK-CVT-NEXT: cmp w13, #0
; CHECK-CVT-NEXT: csel w13, w13, wzr, lt
; CHECK-CVT-NEXT: cmp w13, #0
; CHECK-CVT-NEXT: csinv w13, w13, wzr, ge
; CHECK-CVT-NEXT: cmp w14, #0
; CHECK-CVT-NEXT: csel w9, w14, wzr, lt
; CHECK-CVT-NEXT: cmp w9, #0 ; CHECK-CVT-NEXT: cmp w9, #0
; CHECK-CVT-NEXT: csinv w8, w9, wzr, ge ; CHECK-CVT-NEXT: fmov s1, w13
; CHECK-CVT-NEXT: mov v1.b[7], w8 ; CHECK-CVT-NEXT: fcvtzs w13, s0
; CHECK-CVT-NEXT: fmov d0, d1 ; CHECK-CVT-NEXT: csinv w9, w9, wzr, ge
; CHECK-CVT-NEXT: mov v2.s[1], w8
; CHECK-CVT-NEXT: mov v1.s[1], w12
; CHECK-CVT-NEXT: cmp w13, #0
; CHECK-CVT-NEXT: csel w8, w13, wzr, lt
; CHECK-CVT-NEXT: cmp w8, #0
; CHECK-CVT-NEXT: csinv w8, w8, wzr, ge
; CHECK-CVT-NEXT: mov v1.s[2], w9
; CHECK-CVT-NEXT: mov v2.s[2], w10
; CHECK-CVT-NEXT: mov v1.s[3], w8
; CHECK-CVT-NEXT: mov v2.s[3], w11
; CHECK-CVT-NEXT: xtn v0.4h, v1.4s
; CHECK-CVT-NEXT: xtn2 v0.8h, v2.4s
; CHECK-CVT-NEXT: xtn v0.8b, v0.8h
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_signed_v8f16_v8i1: ; CHECK-FP16-LABEL: test_signed_v8f16_v8i1:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: mov h1, v0.h[1] ; CHECK-FP16-NEXT: movi v1.2d, #0000000000000000
; CHECK-FP16-NEXT: fcvtzs w9, h0 ; CHECK-FP16-NEXT: fcvtzs v0.8h, v0.8h
; CHECK-FP16-NEXT: mov h2, v0.h[3] ; CHECK-FP16-NEXT: smin v0.8h, v0.8h, v1.8h
; CHECK-FP16-NEXT: mov h3, v0.h[5] ; CHECK-FP16-NEXT: movi v1.2d, #0xffffffffffffffff
; CHECK-FP16-NEXT: fcvtzs w8, h1 ; CHECK-FP16-NEXT: smax v0.8h, v0.8h, v1.8h
; CHECK-FP16-NEXT: mov h1, v0.h[2] ; CHECK-FP16-NEXT: xtn v0.8b, v0.8h
; CHECK-FP16-NEXT: cmp w8, #0
; CHECK-FP16-NEXT: csel w8, w8, wzr, lt
; CHECK-FP16-NEXT: fcvtzs w10, h1
; CHECK-FP16-NEXT: cmp w8, #0
; CHECK-FP16-NEXT: csinv w8, w8, wzr, ge
; CHECK-FP16-NEXT: cmp w9, #0
; CHECK-FP16-NEXT: csel w9, w9, wzr, lt
; CHECK-FP16-NEXT: cmp w9, #0
; CHECK-FP16-NEXT: csinv w9, w9, wzr, ge
; CHECK-FP16-NEXT: cmp w10, #0
; CHECK-FP16-NEXT: fmov s1, w9
; CHECK-FP16-NEXT: csel w9, w10, wzr, lt
; CHECK-FP16-NEXT: fcvtzs w10, h2
; CHECK-FP16-NEXT: mov h2, v0.h[4]
; CHECK-FP16-NEXT: cmp w9, #0
; CHECK-FP16-NEXT: mov v1.b[1], w8
; CHECK-FP16-NEXT: csinv w8, w9, wzr, ge
; CHECK-FP16-NEXT: cmp w10, #0
; CHECK-FP16-NEXT: fcvtzs w9, h2
; CHECK-FP16-NEXT: csel w10, w10, wzr, lt
; CHECK-FP16-NEXT: cmp w10, #0
; CHECK-FP16-NEXT: mov h2, v0.h[6]
; CHECK-FP16-NEXT: mov v1.b[2], w8
; CHECK-FP16-NEXT: csinv w8, w10, wzr, ge
; CHECK-FP16-NEXT: cmp w9, #0
; CHECK-FP16-NEXT: fcvtzs w10, h3
; CHECK-FP16-NEXT: csel w9, w9, wzr, lt
; CHECK-FP16-NEXT: mov h0, v0.h[7]
; CHECK-FP16-NEXT: cmp w9, #0
; CHECK-FP16-NEXT: mov v1.b[3], w8
; CHECK-FP16-NEXT: csinv w8, w9, wzr, ge
; CHECK-FP16-NEXT: cmp w10, #0
; CHECK-FP16-NEXT: csel w9, w10, wzr, lt
; CHECK-FP16-NEXT: fcvtzs w10, h2
; CHECK-FP16-NEXT: cmp w9, #0
; CHECK-FP16-NEXT: mov v1.b[4], w8
; CHECK-FP16-NEXT: csinv w8, w9, wzr, ge
; CHECK-FP16-NEXT: cmp w10, #0
; CHECK-FP16-NEXT: csel w9, w10, wzr, lt
; CHECK-FP16-NEXT: fcvtzs w10, h0
; CHECK-FP16-NEXT: cmp w9, #0
; CHECK-FP16-NEXT: mov v1.b[5], w8
; CHECK-FP16-NEXT: csinv w8, w9, wzr, ge
; CHECK-FP16-NEXT: cmp w10, #0
; CHECK-FP16-NEXT: csel w9, w10, wzr, lt
; CHECK-FP16-NEXT: cmp w9, #0
; CHECK-FP16-NEXT: mov v1.b[6], w8
; CHECK-FP16-NEXT: csinv w8, w9, wzr, ge
; CHECK-FP16-NEXT: mov v1.b[7], w8
; CHECK-FP16-NEXT: fmov d0, d1
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <8 x i1> @llvm.fptosi.sat.v8f16.v8i1(<8 x half> %f) %x = call <8 x i1> @llvm.fptosi.sat.v8f16.v8i1(<8 x half> %f)
ret <8 x i1> %x ret <8 x i1> %x
} }
define <8 x i8> @test_signed_v8f16_v8i8(<8 x half> %f) { define <8 x i8> @test_signed_v8f16_v8i8(<8 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v8f16_v8i8: ; CHECK-CVT-LABEL: test_signed_v8f16_v8i8:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v1.4s, v0.4h ; CHECK-CVT-NEXT: fcvtl2 v1.4s, v0.8h
; CHECK-CVT-NEXT: mov w8, #127 ; CHECK-CVT-NEXT: mov w8, #127
; CHECK-CVT-NEXT: fcvtl2 v0.4s, v0.8h ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov w10, #-128
; CHECK-CVT-NEXT: mov s2, v1.s[1] ; CHECK-CVT-NEXT: mov s2, v1.s[1]
; CHECK-CVT-NEXT: fcvtzs w10, s1 ; CHECK-CVT-NEXT: fcvtzs w11, s1
; CHECK-CVT-NEXT: fcvtzs w15, s0
; CHECK-CVT-NEXT: fcvtzs w9, s2 ; CHECK-CVT-NEXT: fcvtzs w9, s2
; CHECK-CVT-NEXT: mov s2, v1.s[2] ; CHECK-CVT-NEXT: mov s2, v1.s[2]
; CHECK-CVT-NEXT: mov s1, v1.s[3]
; CHECK-CVT-NEXT: cmp w9, #127 ; CHECK-CVT-NEXT: cmp w9, #127
; CHECK-CVT-NEXT: csel w11, w9, w8, lt ; CHECK-CVT-NEXT: csel w9, w9, w8, lt
; CHECK-CVT-NEXT: mov w9, #-128
; CHECK-CVT-NEXT: cmn w11, #128
; CHECK-CVT-NEXT: fcvtzs w12, s2
; CHECK-CVT-NEXT: csel w11, w11, w9, gt
; CHECK-CVT-NEXT: cmp w10, #127
; CHECK-CVT-NEXT: csel w10, w10, w8, lt
; CHECK-CVT-NEXT: mov s2, v1.s[3]
; CHECK-CVT-NEXT: cmn w10, #128
; CHECK-CVT-NEXT: csel w10, w10, w9, gt
; CHECK-CVT-NEXT: cmp w12, #127
; CHECK-CVT-NEXT: fmov s1, w10
; CHECK-CVT-NEXT: csel w10, w12, w8, lt
; CHECK-CVT-NEXT: fcvtzs w12, s2 ; CHECK-CVT-NEXT: fcvtzs w12, s2
; CHECK-CVT-NEXT: cmn w10, #128 ; CHECK-CVT-NEXT: cmn w9, #128
; CHECK-CVT-NEXT: csel w10, w10, w9, gt
; CHECK-CVT-NEXT: mov s2, v0.s[1] ; CHECK-CVT-NEXT: mov s2, v0.s[1]
; CHECK-CVT-NEXT: mov v1.b[1], w11 ; CHECK-CVT-NEXT: csel w9, w9, w10, gt
; CHECK-CVT-NEXT: cmp w12, #127 ; CHECK-CVT-NEXT: cmp w11, #127
; CHECK-CVT-NEXT: csel w11, w12, w8, lt ; CHECK-CVT-NEXT: csel w11, w11, w8, lt
; CHECK-CVT-NEXT: fcvtzs w12, s0 ; CHECK-CVT-NEXT: fcvtzs w13, s1
; CHECK-CVT-NEXT: cmn w11, #128 ; CHECK-CVT-NEXT: cmn w11, #128
; CHECK-CVT-NEXT: mov v1.b[2], w10 ; CHECK-CVT-NEXT: mov s1, v0.s[2]
; CHECK-CVT-NEXT: csel w10, w11, w9, gt ; CHECK-CVT-NEXT: csel w11, w11, w10, gt
; CHECK-CVT-NEXT: cmp w12, #127 ; CHECK-CVT-NEXT: cmp w12, #127
; CHECK-CVT-NEXT: fcvtzs w11, s2
; CHECK-CVT-NEXT: csel w12, w12, w8, lt ; CHECK-CVT-NEXT: csel w12, w12, w8, lt
; CHECK-CVT-NEXT: mov s2, v0.s[2] ; CHECK-CVT-NEXT: fcvtzs w14, s2
; CHECK-CVT-NEXT: cmn w12, #128 ; CHECK-CVT-NEXT: cmn w12, #128
; CHECK-CVT-NEXT: mov s0, v0.s[3] ; CHECK-CVT-NEXT: mov s0, v0.s[3]
; CHECK-CVT-NEXT: mov v1.b[3], w10 ; CHECK-CVT-NEXT: csel w12, w12, w10, gt
; CHECK-CVT-NEXT: csel w10, w12, w9, gt ; CHECK-CVT-NEXT: cmp w13, #127
; CHECK-CVT-NEXT: cmp w11, #127 ; CHECK-CVT-NEXT: csel w13, w13, w8, lt
; CHECK-CVT-NEXT: csel w11, w11, w8, lt ; CHECK-CVT-NEXT: fcvtzs w16, s1
; CHECK-CVT-NEXT: fcvtzs w12, s2 ; CHECK-CVT-NEXT: cmn w13, #128
; CHECK-CVT-NEXT: cmn w11, #128 ; CHECK-CVT-NEXT: fmov s2, w11
; CHECK-CVT-NEXT: mov v1.b[4], w10 ; CHECK-CVT-NEXT: csel w13, w13, w10, gt
; CHECK-CVT-NEXT: csel w10, w11, w9, gt ; CHECK-CVT-NEXT: cmp w14, #127
; CHECK-CVT-NEXT: cmp w12, #127 ; CHECK-CVT-NEXT: csel w14, w14, w8, lt
; CHECK-CVT-NEXT: csel w11, w12, w8, lt ; CHECK-CVT-NEXT: cmn w14, #128
; CHECK-CVT-NEXT: fcvtzs w12, s0 ; CHECK-CVT-NEXT: csel w14, w14, w10, gt
; CHECK-CVT-NEXT: cmp w15, #127
; CHECK-CVT-NEXT: csel w15, w15, w8, lt
; CHECK-CVT-NEXT: cmn w15, #128
; CHECK-CVT-NEXT: csel w15, w15, w10, gt
; CHECK-CVT-NEXT: cmp w16, #127
; CHECK-CVT-NEXT: csel w11, w16, w8, lt
; CHECK-CVT-NEXT: cmn w11, #128 ; CHECK-CVT-NEXT: cmn w11, #128
; CHECK-CVT-NEXT: mov v1.b[5], w10 ; CHECK-CVT-NEXT: fmov s1, w15
; CHECK-CVT-NEXT: csel w10, w11, w9, gt ; CHECK-CVT-NEXT: fcvtzs w15, s0
; CHECK-CVT-NEXT: cmp w12, #127 ; CHECK-CVT-NEXT: csel w11, w11, w10, gt
; CHECK-CVT-NEXT: csel w8, w12, w8, lt ; CHECK-CVT-NEXT: mov v2.s[1], w9
; CHECK-CVT-NEXT: mov v1.s[1], w14
; CHECK-CVT-NEXT: cmp w15, #127
; CHECK-CVT-NEXT: csel w8, w15, w8, lt
; CHECK-CVT-NEXT: cmn w8, #128 ; CHECK-CVT-NEXT: cmn w8, #128
; CHECK-CVT-NEXT: mov v1.b[6], w10 ; CHECK-CVT-NEXT: csel w8, w8, w10, gt
; CHECK-CVT-NEXT: csel w8, w8, w9, gt ; CHECK-CVT-NEXT: mov v1.s[2], w11
; CHECK-CVT-NEXT: mov v1.b[7], w8 ; CHECK-CVT-NEXT: mov v2.s[2], w12
; CHECK-CVT-NEXT: fmov d0, d1 ; CHECK-CVT-NEXT: mov v1.s[3], w8
; CHECK-CVT-NEXT: mov v2.s[3], w13
; CHECK-CVT-NEXT: xtn v0.4h, v1.4s
; CHECK-CVT-NEXT: xtn2 v0.8h, v2.4s
; CHECK-CVT-NEXT: xtn v0.8b, v0.8h
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_signed_v8f16_v8i8: ; CHECK-FP16-LABEL: test_signed_v8f16_v8i8:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: mov h1, v0.h[1] ; CHECK-FP16-NEXT: fcvtzs v0.8h, v0.8h
; CHECK-FP16-NEXT: mov w8, #127 ; CHECK-FP16-NEXT: sqxtn v0.8b, v0.8h
; CHECK-FP16-NEXT: fcvtzs w10, h0
; CHECK-FP16-NEXT: mov h2, v0.h[3]
; CHECK-FP16-NEXT: mov h3, v0.h[5]
; CHECK-FP16-NEXT: fcvtzs w9, h1
; CHECK-FP16-NEXT: mov h1, v0.h[2]
; CHECK-FP16-NEXT: cmp w9, #127
; CHECK-FP16-NEXT: csel w11, w9, w8, lt
; CHECK-FP16-NEXT: mov w9, #-128
; CHECK-FP16-NEXT: cmn w11, #128
; CHECK-FP16-NEXT: fcvtzs w12, h1
; CHECK-FP16-NEXT: csel w11, w11, w9, gt
; CHECK-FP16-NEXT: cmp w10, #127
; CHECK-FP16-NEXT: csel w10, w10, w8, lt
; CHECK-FP16-NEXT: cmn w10, #128
; CHECK-FP16-NEXT: csel w10, w10, w9, gt
; CHECK-FP16-NEXT: cmp w12, #127
; CHECK-FP16-NEXT: fmov s1, w10
; CHECK-FP16-NEXT: csel w10, w12, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h2
; CHECK-FP16-NEXT: mov h2, v0.h[4]
; CHECK-FP16-NEXT: cmn w10, #128
; CHECK-FP16-NEXT: mov v1.b[1], w11
; CHECK-FP16-NEXT: csel w10, w10, w9, gt
; CHECK-FP16-NEXT: cmp w12, #127
; CHECK-FP16-NEXT: fcvtzs w11, h2
; CHECK-FP16-NEXT: csel w12, w12, w8, lt
; CHECK-FP16-NEXT: cmn w12, #128
; CHECK-FP16-NEXT: mov h2, v0.h[6]
; CHECK-FP16-NEXT: mov v1.b[2], w10
; CHECK-FP16-NEXT: csel w10, w12, w9, gt
; CHECK-FP16-NEXT: cmp w11, #127
; CHECK-FP16-NEXT: fcvtzs w12, h3
; CHECK-FP16-NEXT: csel w11, w11, w8, lt
; CHECK-FP16-NEXT: mov h0, v0.h[7]
; CHECK-FP16-NEXT: cmn w11, #128
; CHECK-FP16-NEXT: mov v1.b[3], w10
; CHECK-FP16-NEXT: csel w10, w11, w9, gt
; CHECK-FP16-NEXT: cmp w12, #127
; CHECK-FP16-NEXT: csel w11, w12, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h2
; CHECK-FP16-NEXT: cmn w11, #128
; CHECK-FP16-NEXT: mov v1.b[4], w10
; CHECK-FP16-NEXT: csel w10, w11, w9, gt
; CHECK-FP16-NEXT: cmp w12, #127
; CHECK-FP16-NEXT: csel w11, w12, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h0
; CHECK-FP16-NEXT: cmn w11, #128
; CHECK-FP16-NEXT: mov v1.b[5], w10
; CHECK-FP16-NEXT: csel w10, w11, w9, gt
; CHECK-FP16-NEXT: cmp w12, #127
; CHECK-FP16-NEXT: csel w8, w12, w8, lt
; CHECK-FP16-NEXT: cmn w8, #128
; CHECK-FP16-NEXT: mov v1.b[6], w10
; CHECK-FP16-NEXT: csel w8, w8, w9, gt
; CHECK-FP16-NEXT: mov v1.b[7], w8
; CHECK-FP16-NEXT: fmov d0, d1
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <8 x i8> @llvm.fptosi.sat.v8f16.v8i8(<8 x half> %f) %x = call <8 x i8> @llvm.fptosi.sat.v8f16.v8i8(<8 x half> %f)
ret <8 x i8> %x ret <8 x i8> %x
} }
define <8 x i13> @test_signed_v8f16_v8i13(<8 x half> %f) { define <8 x i13> @test_signed_v8f16_v8i13(<8 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v8f16_v8i13: ; CHECK-CVT-LABEL: test_signed_v8f16_v8i13:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: mov h1, v0.h[1] ; CHECK-CVT-NEXT: fcvtl2 v1.4s, v0.8h
; CHECK-CVT-NEXT: fcvt s2, h0 ; CHECK-CVT-NEXT: mov w8, #4095
; CHECK-CVT-NEXT: mov w9, #4095 ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov w10, #-4096 ; CHECK-CVT-NEXT: mov w10, #-4096
; CHECK-CVT-NEXT: mov h3, v0.h[4] ; CHECK-CVT-NEXT: mov s2, v1.s[1]
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-CVT-NEXT: fcvtzs w11, s1
; CHECK-CVT-NEXT: fcvtzs w11, s2 ; CHECK-CVT-NEXT: fcvtzs w15, s0
; CHECK-CVT-NEXT: mov h2, v0.h[3] ; CHECK-CVT-NEXT: fcvtzs w9, s2
; CHECK-CVT-NEXT: fcvtzs w8, s1 ; CHECK-CVT-NEXT: mov s2, v1.s[2]
; CHECK-CVT-NEXT: mov h1, v0.h[2] ; CHECK-CVT-NEXT: mov s1, v1.s[3]
; CHECK-CVT-NEXT: fcvt s2, h2 ; CHECK-CVT-NEXT: cmp w9, #4095
; CHECK-CVT-NEXT: cmp w8, #4095 ; CHECK-CVT-NEXT: csel w9, w9, w8, lt
; CHECK-CVT-NEXT: csel w8, w8, w9, lt ; CHECK-CVT-NEXT: fcvtzs w12, s2
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-CVT-NEXT: cmn w9, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: cmn w8, #1, lsl #12 // =4096 ; CHECK-CVT-NEXT: mov s2, v0.s[1]
; CHECK-CVT-NEXT: csel w8, w8, w10, gt ; CHECK-CVT-NEXT: csel w9, w9, w10, gt
; CHECK-CVT-NEXT: cmp w11, #4095 ; CHECK-CVT-NEXT: cmp w11, #4095
; CHECK-CVT-NEXT: csel w11, w11, w9, lt ; CHECK-CVT-NEXT: csel w11, w11, w8, lt
; CHECK-CVT-NEXT: fcvtzs w13, s1
; CHECK-CVT-NEXT: cmn w11, #1, lsl #12 // =4096 ; CHECK-CVT-NEXT: cmn w11, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: fcvtzs w12, s1 ; CHECK-CVT-NEXT: mov s1, v0.s[2]
; CHECK-CVT-NEXT: csel w11, w11, w10, gt ; CHECK-CVT-NEXT: csel w11, w11, w10, gt
; CHECK-CVT-NEXT: cmp w12, #4095 ; CHECK-CVT-NEXT: cmp w12, #4095
; CHECK-CVT-NEXT: fmov s1, w11 ; CHECK-CVT-NEXT: csel w12, w12, w8, lt
; CHECK-CVT-NEXT: fcvtzs w11, s2 ; CHECK-CVT-NEXT: fcvtzs w14, s2
; CHECK-CVT-NEXT: fcvt s2, h3
; CHECK-CVT-NEXT: mov h3, v0.h[5]
; CHECK-CVT-NEXT: csel w12, w12, w9, lt
; CHECK-CVT-NEXT: mov v1.h[1], w8
; CHECK-CVT-NEXT: cmn w12, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: csel w8, w12, w10, gt
; CHECK-CVT-NEXT: cmp w11, #4095
; CHECK-CVT-NEXT: fcvtzs w12, s2
; CHECK-CVT-NEXT: fcvt s2, h3
; CHECK-CVT-NEXT: csel w11, w11, w9, lt
; CHECK-CVT-NEXT: mov h3, v0.h[6]
; CHECK-CVT-NEXT: mov v1.h[2], w8
; CHECK-CVT-NEXT: cmn w11, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: csel w8, w11, w10, gt
; CHECK-CVT-NEXT: cmp w12, #4095
; CHECK-CVT-NEXT: fcvtzs w11, s2
; CHECK-CVT-NEXT: csel w12, w12, w9, lt
; CHECK-CVT-NEXT: fcvt s2, h3
; CHECK-CVT-NEXT: mov h0, v0.h[7]
; CHECK-CVT-NEXT: mov v1.h[3], w8
; CHECK-CVT-NEXT: cmn w12, #1, lsl #12 // =4096 ; CHECK-CVT-NEXT: cmn w12, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: csel w8, w12, w10, gt ; CHECK-CVT-NEXT: mov s0, v0.s[3]
; CHECK-CVT-NEXT: cmp w11, #4095 ; CHECK-CVT-NEXT: csel w12, w12, w10, gt
; CHECK-CVT-NEXT: csel w11, w11, w9, lt ; CHECK-CVT-NEXT: cmp w13, #4095
; CHECK-CVT-NEXT: fcvtzs w12, s2 ; CHECK-CVT-NEXT: csel w13, w13, w8, lt
; CHECK-CVT-NEXT: fcvt s0, h0 ; CHECK-CVT-NEXT: fcvtzs w16, s1
; CHECK-CVT-NEXT: mov v1.h[4], w8 ; CHECK-CVT-NEXT: cmn w13, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: cmn w11, #1, lsl #12 // =4096 ; CHECK-CVT-NEXT: fmov s2, w11
; CHECK-CVT-NEXT: csel w8, w11, w10, gt ; CHECK-CVT-NEXT: csel w13, w13, w10, gt
; CHECK-CVT-NEXT: cmp w12, #4095 ; CHECK-CVT-NEXT: cmp w14, #4095
; CHECK-CVT-NEXT: csel w11, w12, w9, lt ; CHECK-CVT-NEXT: csel w14, w14, w8, lt
; CHECK-CVT-NEXT: fcvtzs w12, s0 ; CHECK-CVT-NEXT: cmn w14, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: mov v1.h[5], w8 ; CHECK-CVT-NEXT: csel w14, w14, w10, gt
; CHECK-CVT-NEXT: cmp w15, #4095
; CHECK-CVT-NEXT: csel w15, w15, w8, lt
; CHECK-CVT-NEXT: cmn w15, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: csel w15, w15, w10, gt
; CHECK-CVT-NEXT: cmp w16, #4095
; CHECK-CVT-NEXT: csel w11, w16, w8, lt
; CHECK-CVT-NEXT: cmn w11, #1, lsl #12 // =4096 ; CHECK-CVT-NEXT: cmn w11, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: csel w8, w11, w10, gt ; CHECK-CVT-NEXT: fmov s1, w15
; CHECK-CVT-NEXT: cmp w12, #4095 ; CHECK-CVT-NEXT: fcvtzs w15, s0
; CHECK-CVT-NEXT: csel w9, w12, w9, lt ; CHECK-CVT-NEXT: csel w11, w11, w10, gt
; CHECK-CVT-NEXT: mov v1.h[6], w8 ; CHECK-CVT-NEXT: mov v2.s[1], w9
; CHECK-CVT-NEXT: cmn w9, #1, lsl #12 // =4096 ; CHECK-CVT-NEXT: mov v1.s[1], w14
; CHECK-CVT-NEXT: csel w8, w9, w10, gt ; CHECK-CVT-NEXT: cmp w15, #4095
; CHECK-CVT-NEXT: mov v1.h[7], w8 ; CHECK-CVT-NEXT: csel w8, w15, w8, lt
; CHECK-CVT-NEXT: mov v0.16b, v1.16b ; CHECK-CVT-NEXT: cmn w8, #1, lsl #12 // =4096
; CHECK-CVT-NEXT: csel w8, w8, w10, gt
; CHECK-CVT-NEXT: mov v1.s[2], w11
; CHECK-CVT-NEXT: mov v2.s[2], w12
; CHECK-CVT-NEXT: mov v1.s[3], w8
; CHECK-CVT-NEXT: mov v2.s[3], w13
; CHECK-CVT-NEXT: xtn v0.4h, v1.4s
; CHECK-CVT-NEXT: xtn2 v0.8h, v2.4s
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_signed_v8f16_v8i13: ; CHECK-FP16-LABEL: test_signed_v8f16_v8i13:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: mov h1, v0.h[1] ; CHECK-FP16-NEXT: mvni v1.8h, #240, lsl #8
; CHECK-FP16-NEXT: mov w8, #4095 ; CHECK-FP16-NEXT: fcvtzs v0.8h, v0.8h
; CHECK-FP16-NEXT: fcvtzs w10, h0 ; CHECK-FP16-NEXT: smin v0.8h, v0.8h, v1.8h
; CHECK-FP16-NEXT: mov w11, #-4096 ; CHECK-FP16-NEXT: movi v1.8h, #240, lsl #8
; CHECK-FP16-NEXT: mov h2, v0.h[3] ; CHECK-FP16-NEXT: smax v0.8h, v0.8h, v1.8h
; CHECK-FP16-NEXT: mov h3, v0.h[5]
; CHECK-FP16-NEXT: fcvtzs w9, h1
; CHECK-FP16-NEXT: mov h1, v0.h[2]
; CHECK-FP16-NEXT: cmp w9, #4095
; CHECK-FP16-NEXT: csel w9, w9, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h1
; CHECK-FP16-NEXT: cmn w9, #1, lsl #12 // =4096
; CHECK-FP16-NEXT: csel w9, w9, w11, gt
; CHECK-FP16-NEXT: cmp w10, #4095
; CHECK-FP16-NEXT: csel w10, w10, w8, lt
; CHECK-FP16-NEXT: cmn w10, #1, lsl #12 // =4096
; CHECK-FP16-NEXT: csel w10, w10, w11, gt
; CHECK-FP16-NEXT: cmp w12, #4095
; CHECK-FP16-NEXT: fmov s1, w10
; CHECK-FP16-NEXT: csel w10, w12, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h2
; CHECK-FP16-NEXT: mov h2, v0.h[4]
; CHECK-FP16-NEXT: cmn w10, #1, lsl #12 // =4096
; CHECK-FP16-NEXT: mov v1.h[1], w9
; CHECK-FP16-NEXT: csel w9, w10, w11, gt
; CHECK-FP16-NEXT: cmp w12, #4095
; CHECK-FP16-NEXT: fcvtzs w10, h2
; CHECK-FP16-NEXT: csel w12, w12, w8, lt
; CHECK-FP16-NEXT: cmn w12, #1, lsl #12 // =4096
; CHECK-FP16-NEXT: mov h2, v0.h[6]
; CHECK-FP16-NEXT: mov v1.h[2], w9
; CHECK-FP16-NEXT: csel w9, w12, w11, gt
; CHECK-FP16-NEXT: cmp w10, #4095
; CHECK-FP16-NEXT: fcvtzs w12, h3
; CHECK-FP16-NEXT: csel w10, w10, w8, lt
; CHECK-FP16-NEXT: mov h0, v0.h[7]
; CHECK-FP16-NEXT: cmn w10, #1, lsl #12 // =4096
; CHECK-FP16-NEXT: mov v1.h[3], w9
; CHECK-FP16-NEXT: csel w9, w10, w11, gt
; CHECK-FP16-NEXT: cmp w12, #4095
; CHECK-FP16-NEXT: csel w10, w12, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h2
; CHECK-FP16-NEXT: cmn w10, #1, lsl #12 // =4096
; CHECK-FP16-NEXT: mov v1.h[4], w9
; CHECK-FP16-NEXT: csel w9, w10, w11, gt
; CHECK-FP16-NEXT: cmp w12, #4095
; CHECK-FP16-NEXT: csel w10, w12, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h0
; CHECK-FP16-NEXT: cmn w10, #1, lsl #12 // =4096
; CHECK-FP16-NEXT: mov v1.h[5], w9
; CHECK-FP16-NEXT: csel w9, w10, w11, gt
; CHECK-FP16-NEXT: cmp w12, #4095
; CHECK-FP16-NEXT: csel w8, w12, w8, lt
; CHECK-FP16-NEXT: cmn w8, #1, lsl #12 // =4096
; CHECK-FP16-NEXT: mov v1.h[6], w9
; CHECK-FP16-NEXT: csel w8, w8, w11, gt
; CHECK-FP16-NEXT: mov v1.h[7], w8
; CHECK-FP16-NEXT: mov v0.16b, v1.16b
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <8 x i13> @llvm.fptosi.sat.v8f16.v8i13(<8 x half> %f) %x = call <8 x i13> @llvm.fptosi.sat.v8f16.v8i13(<8 x half> %f)
ret <8 x i13> %x ret <8 x i13> %x
} }
define <8 x i16> @test_signed_v8f16_v8i16(<8 x half> %f) { define <8 x i16> @test_signed_v8f16_v8i16(<8 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v8f16_v8i16: ; CHECK-CVT-LABEL: test_signed_v8f16_v8i16:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v1.4s, v0.4h ; CHECK-CVT-NEXT: fcvtl2 v1.4s, v0.8h
; CHECK-CVT-NEXT: mov w8, #32767 ; CHECK-CVT-NEXT: mov w8, #32767
; CHECK-CVT-NEXT: mov w11, #-32768 ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: fcvtl2 v0.4s, v0.8h ; CHECK-CVT-NEXT: mov w10, #-32768
; CHECK-CVT-NEXT: mov s2, v1.s[1] ; CHECK-CVT-NEXT: mov s2, v1.s[1]
; CHECK-CVT-NEXT: fcvtzs w10, s1 ; CHECK-CVT-NEXT: fcvtzs w11, s1
; CHECK-CVT-NEXT: fcvtzs w15, s0
; CHECK-CVT-NEXT: fcvtzs w9, s2 ; CHECK-CVT-NEXT: fcvtzs w9, s2
; CHECK-CVT-NEXT: mov s2, v1.s[2] ; CHECK-CVT-NEXT: mov s2, v1.s[2]
; CHECK-CVT-NEXT: mov s1, v1.s[3]
; CHECK-CVT-NEXT: cmp w9, w8 ; CHECK-CVT-NEXT: cmp w9, w8
; CHECK-CVT-NEXT: csel w9, w9, w8, lt ; CHECK-CVT-NEXT: csel w9, w9, w8, lt
; CHECK-CVT-NEXT: fcvtzs w12, s2 ; CHECK-CVT-NEXT: fcvtzs w12, s2
; CHECK-CVT-NEXT: cmn w9, #8, lsl #12 // =32768 ; CHECK-CVT-NEXT: cmn w9, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: mov s2, v1.s[3]
; CHECK-CVT-NEXT: csel w9, w9, w11, gt
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w10, w10, w8, lt
; CHECK-CVT-NEXT: cmn w10, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: csel w10, w10, w11, gt
; CHECK-CVT-NEXT: cmp w12, w8
; CHECK-CVT-NEXT: fmov s1, w10
; CHECK-CVT-NEXT: csel w10, w12, w8, lt
; CHECK-CVT-NEXT: fcvtzs w12, s2
; CHECK-CVT-NEXT: cmn w10, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: mov s2, v0.s[1] ; CHECK-CVT-NEXT: mov s2, v0.s[1]
; CHECK-CVT-NEXT: mov v1.h[1], w9 ; CHECK-CVT-NEXT: csel w9, w9, w10, gt
; CHECK-CVT-NEXT: csel w9, w10, w11, gt ; CHECK-CVT-NEXT: cmp w11, w8
; CHECK-CVT-NEXT: cmp w12, w8 ; CHECK-CVT-NEXT: csel w11, w11, w8, lt
; CHECK-CVT-NEXT: csel w10, w12, w8, lt ; CHECK-CVT-NEXT: fcvtzs w13, s1
; CHECK-CVT-NEXT: fcvtzs w12, s0 ; CHECK-CVT-NEXT: cmn w11, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: cmn w10, #8, lsl #12 // =32768 ; CHECK-CVT-NEXT: mov s1, v0.s[2]
; CHECK-CVT-NEXT: mov v1.h[2], w9 ; CHECK-CVT-NEXT: csel w11, w11, w10, gt
; CHECK-CVT-NEXT: csel w9, w10, w11, gt
; CHECK-CVT-NEXT: cmp w12, w8 ; CHECK-CVT-NEXT: cmp w12, w8
; CHECK-CVT-NEXT: fcvtzs w10, s2
; CHECK-CVT-NEXT: csel w12, w12, w8, lt ; CHECK-CVT-NEXT: csel w12, w12, w8, lt
; CHECK-CVT-NEXT: mov s2, v0.s[2] ; CHECK-CVT-NEXT: fcvtzs w14, s2
; CHECK-CVT-NEXT: cmn w12, #8, lsl #12 // =32768 ; CHECK-CVT-NEXT: cmn w12, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: mov s0, v0.s[3] ; CHECK-CVT-NEXT: mov s0, v0.s[3]
; CHECK-CVT-NEXT: mov v1.h[3], w9 ; CHECK-CVT-NEXT: csel w12, w12, w10, gt
; CHECK-CVT-NEXT: csel w9, w12, w11, gt ; CHECK-CVT-NEXT: cmp w13, w8
; CHECK-CVT-NEXT: cmp w10, w8 ; CHECK-CVT-NEXT: csel w13, w13, w8, lt
; CHECK-CVT-NEXT: csel w10, w10, w8, lt ; CHECK-CVT-NEXT: fcvtzs w16, s1
; CHECK-CVT-NEXT: fcvtzs w12, s2 ; CHECK-CVT-NEXT: cmn w13, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: cmn w10, #8, lsl #12 // =32768 ; CHECK-CVT-NEXT: fmov s2, w11
; CHECK-CVT-NEXT: mov v1.h[4], w9 ; CHECK-CVT-NEXT: csel w13, w13, w10, gt
; CHECK-CVT-NEXT: csel w9, w10, w11, gt ; CHECK-CVT-NEXT: cmp w14, w8
; CHECK-CVT-NEXT: cmp w12, w8 ; CHECK-CVT-NEXT: csel w14, w14, w8, lt
; CHECK-CVT-NEXT: csel w10, w12, w8, lt ; CHECK-CVT-NEXT: cmn w14, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: fcvtzs w12, s0 ; CHECK-CVT-NEXT: csel w14, w14, w10, gt
; CHECK-CVT-NEXT: cmn w10, #8, lsl #12 // =32768 ; CHECK-CVT-NEXT: cmp w15, w8
; CHECK-CVT-NEXT: mov v1.h[5], w9 ; CHECK-CVT-NEXT: csel w15, w15, w8, lt
; CHECK-CVT-NEXT: csel w9, w10, w11, gt ; CHECK-CVT-NEXT: cmn w15, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: cmp w12, w8 ; CHECK-CVT-NEXT: csel w15, w15, w10, gt
; CHECK-CVT-NEXT: csel w8, w12, w8, lt ; CHECK-CVT-NEXT: cmp w16, w8
; CHECK-CVT-NEXT: csel w11, w16, w8, lt
; CHECK-CVT-NEXT: cmn w11, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: fmov s1, w15
; CHECK-CVT-NEXT: fcvtzs w15, s0
; CHECK-CVT-NEXT: csel w11, w11, w10, gt
; CHECK-CVT-NEXT: mov v2.s[1], w9
; CHECK-CVT-NEXT: mov v1.s[1], w14
; CHECK-CVT-NEXT: cmp w15, w8
; CHECK-CVT-NEXT: csel w8, w15, w8, lt
; CHECK-CVT-NEXT: cmn w8, #8, lsl #12 // =32768 ; CHECK-CVT-NEXT: cmn w8, #8, lsl #12 // =32768
; CHECK-CVT-NEXT: mov v1.h[6], w9 ; CHECK-CVT-NEXT: csel w8, w8, w10, gt
; CHECK-CVT-NEXT: csel w8, w8, w11, gt ; CHECK-CVT-NEXT: mov v1.s[2], w11
; CHECK-CVT-NEXT: mov v1.h[7], w8 ; CHECK-CVT-NEXT: mov v2.s[2], w12
; CHECK-CVT-NEXT: mov v0.16b, v1.16b ; CHECK-CVT-NEXT: mov v1.s[3], w8
; CHECK-CVT-NEXT: mov v2.s[3], w13
; CHECK-CVT-NEXT: xtn v0.4h, v1.4s
; CHECK-CVT-NEXT: xtn2 v0.8h, v2.4s
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_signed_v8f16_v8i16: ; CHECK-FP16-LABEL: test_signed_v8f16_v8i16:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: fcvtzs v0.8h, v0.8h ; CHECK-FP16-NEXT: fcvtzs v0.8h, v0.8h
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <8 x i16> @llvm.fptosi.sat.v8f16.v8i16(<8 x half> %f) %x = call <8 x i16> @llvm.fptosi.sat.v8f16.v8i16(<8 x half> %f)
ret <8 x i16> %x ret <8 x i16> %x
} }
define <8 x i19> @test_signed_v8f16_v8i19(<8 x half> %f) { define <8 x i19> @test_signed_v8f16_v8i19(<8 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v8f16_v8i19: ; CHECK-LABEL: test_signed_v8f16_v8i19:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: ext v1.16b, v0.16b, v0.16b, #8 ; CHECK-NEXT: fcvtl v2.4s, v0.4h
; CHECK-CVT-NEXT: mov w8, #262143 ; CHECK-NEXT: fcvtl2 v0.4s, v0.8h
; CHECK-CVT-NEXT: mov w12, #-262144 ; CHECK-NEXT: movi v1.4s, #3, msl #16
; CHECK-CVT-NEXT: fcvt s5, h0 ; CHECK-NEXT: mvni v3.4s, #3, msl #16
; CHECK-CVT-NEXT: mov h2, v1.h[1] ; CHECK-NEXT: fcvtzs v2.4s, v2.4s
; CHECK-CVT-NEXT: fcvt s3, h1 ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: mov h4, v1.h[2] ; CHECK-NEXT: smin v2.4s, v2.4s, v1.4s
; CHECK-CVT-NEXT: mov h1, v1.h[3] ; CHECK-NEXT: smin v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzs w10, s5 ; CHECK-NEXT: smax v1.4s, v2.4s, v3.4s
; CHECK-CVT-NEXT: fcvt s2, h2 ; CHECK-NEXT: smax v0.4s, v0.4s, v3.4s
; CHECK-CVT-NEXT: fcvtzs w9, s3 ; CHECK-NEXT: mov w1, v1.s[1]
; CHECK-CVT-NEXT: fcvt s3, h4 ; CHECK-NEXT: mov w2, v1.s[2]
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-NEXT: mov w5, v0.s[1]
; CHECK-CVT-NEXT: cmp w9, w8 ; CHECK-NEXT: mov w3, v1.s[3]
; CHECK-CVT-NEXT: fcvtzs w11, s2 ; CHECK-NEXT: mov w6, v0.s[2]
; CHECK-CVT-NEXT: csel w9, w9, w8, lt ; CHECK-NEXT: mov w7, v0.s[3]
; CHECK-CVT-NEXT: cmn w9, #64, lsl #12 // =262144 ; CHECK-NEXT: fmov w4, s0
; CHECK-CVT-NEXT: fcvtzs w13, s3 ; CHECK-NEXT: fmov w0, s1
; CHECK-CVT-NEXT: csel w4, w9, w12, gt ; CHECK-NEXT: ret
; CHECK-CVT-NEXT: mov h2, v0.h[1]
; CHECK-CVT-NEXT: cmp w11, w8
; CHECK-CVT-NEXT: fcvtzs w9, s1
; CHECK-CVT-NEXT: csel w11, w11, w8, lt
; CHECK-CVT-NEXT: mov h1, v0.h[2]
; CHECK-CVT-NEXT: cmn w11, #64, lsl #12 // =262144
; CHECK-CVT-NEXT: mov h0, v0.h[3]
; CHECK-CVT-NEXT: csel w5, w11, w12, gt
; CHECK-CVT-NEXT: cmp w13, w8
; CHECK-CVT-NEXT: csel w11, w13, w8, lt
; CHECK-CVT-NEXT: fcvt s2, h2
; CHECK-CVT-NEXT: cmn w11, #64, lsl #12 // =262144
; CHECK-CVT-NEXT: fcvt s1, h1
; CHECK-CVT-NEXT: csel w6, w11, w12, gt
; CHECK-CVT-NEXT: cmp w9, w8
; CHECK-CVT-NEXT: csel w9, w9, w8, lt
; CHECK-CVT-NEXT: fcvt s0, h0
; CHECK-CVT-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-CVT-NEXT: fcvtzs w11, s2
; CHECK-CVT-NEXT: csel w7, w9, w12, gt
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w9, w10, w8, lt
; CHECK-CVT-NEXT: fcvtzs w10, s1
; CHECK-CVT-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-CVT-NEXT: csel w0, w9, w12, gt
; CHECK-CVT-NEXT: cmp w11, w8
; CHECK-CVT-NEXT: csel w9, w11, w8, lt
; CHECK-CVT-NEXT: fcvtzs w11, s0
; CHECK-CVT-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-CVT-NEXT: csel w1, w9, w12, gt
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w9, w10, w8, lt
; CHECK-CVT-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-CVT-NEXT: csel w2, w9, w12, gt
; CHECK-CVT-NEXT: cmp w11, w8
; CHECK-CVT-NEXT: csel w8, w11, w8, lt
; CHECK-CVT-NEXT: cmn w8, #64, lsl #12 // =262144
; CHECK-CVT-NEXT: csel w3, w8, w12, gt
; CHECK-CVT-NEXT: ret
;
; CHECK-FP16-LABEL: test_signed_v8f16_v8i19:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: ext v1.16b, v0.16b, v0.16b, #8
; CHECK-FP16-NEXT: mov w8, #262143
; CHECK-FP16-NEXT: mov w11, #-262144
; CHECK-FP16-NEXT: mov h2, v1.h[1]
; CHECK-FP16-NEXT: fcvtzs w9, h1
; CHECK-FP16-NEXT: mov h3, v1.h[2]
; CHECK-FP16-NEXT: mov h1, v1.h[3]
; CHECK-FP16-NEXT: cmp w9, w8
; CHECK-FP16-NEXT: fcvtzs w10, h2
; CHECK-FP16-NEXT: csel w9, w9, w8, lt
; CHECK-FP16-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-FP16-NEXT: fcvtzs w12, h3
; CHECK-FP16-NEXT: csel w4, w9, w11, gt
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w9, w10, w8, lt
; CHECK-FP16-NEXT: fcvtzs w10, h1
; CHECK-FP16-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-FP16-NEXT: mov h1, v0.h[1]
; CHECK-FP16-NEXT: csel w5, w9, w11, gt
; CHECK-FP16-NEXT: cmp w12, w8
; CHECK-FP16-NEXT: csel w9, w12, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h0
; CHECK-FP16-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: csel w6, w9, w11, gt
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w9, w10, w8, lt
; CHECK-FP16-NEXT: fcvtzs w10, h1
; CHECK-FP16-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-FP16-NEXT: csel w7, w9, w11, gt
; CHECK-FP16-NEXT: cmp w12, w8
; CHECK-FP16-NEXT: csel w9, w12, w8, lt
; CHECK-FP16-NEXT: fcvtzs w12, h2
; CHECK-FP16-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-FP16-NEXT: csel w0, w9, w11, gt
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w9, w10, w8, lt
; CHECK-FP16-NEXT: fcvtzs w10, h0
; CHECK-FP16-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-FP16-NEXT: csel w1, w9, w11, gt
; CHECK-FP16-NEXT: cmp w12, w8
; CHECK-FP16-NEXT: csel w9, w12, w8, lt
; CHECK-FP16-NEXT: cmn w9, #64, lsl #12 // =262144
; CHECK-FP16-NEXT: csel w2, w9, w11, gt
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w8, w10, w8, lt
; CHECK-FP16-NEXT: cmn w8, #64, lsl #12 // =262144
; CHECK-FP16-NEXT: csel w3, w8, w11, gt
; CHECK-FP16-NEXT: ret
%x = call <8 x i19> @llvm.fptosi.sat.v8f16.v8i19(<8 x half> %f) %x = call <8 x i19> @llvm.fptosi.sat.v8f16.v8i19(<8 x half> %f)
ret <8 x i19> %x ret <8 x i19> %x
} }
define <8 x i32> @test_signed_v8f16_v8i32_duplicate(<8 x half> %f) { define <8 x i32> @test_signed_v8f16_v8i32_duplicate(<8 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v8f16_v8i32_duplicate: ; CHECK-LABEL: test_signed_v8f16_v8i32_duplicate:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl2 v1.4s, v0.8h ; CHECK-NEXT: fcvtl2 v1.4s, v0.8h
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: fcvtzs v1.4s, v1.4s ; CHECK-NEXT: fcvtzs v1.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzs v0.4s, v0.4s ; CHECK-NEXT: fcvtzs v0.4s, v0.4s
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_signed_v8f16_v8i32_duplicate:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: ext v3.16b, v0.16b, v0.16b, #8
; CHECK-FP16-NEXT: mov h4, v0.h[1]
; CHECK-FP16-NEXT: fcvtzs w9, h0
; CHECK-FP16-NEXT: mov h2, v3.h[1]
; CHECK-FP16-NEXT: fcvtzs w8, h3
; CHECK-FP16-NEXT: mov h5, v3.h[2]
; CHECK-FP16-NEXT: mov h3, v3.h[3]
; CHECK-FP16-NEXT: fmov s1, w8
; CHECK-FP16-NEXT: fcvtzs w8, h2
; CHECK-FP16-NEXT: fmov s2, w9
; CHECK-FP16-NEXT: fcvtzs w9, h4
; CHECK-FP16-NEXT: mov h4, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: mov v1.s[1], w8
; CHECK-FP16-NEXT: fcvtzs w8, h5
; CHECK-FP16-NEXT: mov v2.s[1], w9
; CHECK-FP16-NEXT: fcvtzs w9, h4
; CHECK-FP16-NEXT: mov v1.s[2], w8
; CHECK-FP16-NEXT: fcvtzs w8, h3
; CHECK-FP16-NEXT: mov v2.s[2], w9
; CHECK-FP16-NEXT: fcvtzs w9, h0
; CHECK-FP16-NEXT: mov v1.s[3], w8
; CHECK-FP16-NEXT: mov v2.s[3], w9
; CHECK-FP16-NEXT: mov v0.16b, v2.16b
; CHECK-FP16-NEXT: ret
%x = call <8 x i32> @llvm.fptosi.sat.v8f16.v8i32(<8 x half> %f) %x = call <8 x i32> @llvm.fptosi.sat.v8f16.v8i32(<8 x half> %f)
ret <8 x i32> %x ret <8 x i32> %x
} }
define <8 x i50> @test_signed_v8f16_v8i50(<8 x half> %f) { define <8 x i50> @test_signed_v8f16_v8i50(<8 x half> %f) {
; CHECK-CVT-LABEL: test_signed_v8f16_v8i50: ; CHECK-CVT-LABEL: test_signed_v8f16_v8i50:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: ext v1.16b, v0.16b, v0.16b, #8 ; CHECK-CVT-NEXT: ext v1.16b, v0.16b, v0.16b, #8
▲ Show 20 Lines • Show All 574 Lines • Show Last 20 Lines

llvm/test/CodeGen/AArch64/fptoui-sat-vector.ll

Show First 20 Lines • Show All 505 Lines • ▼ Show 20 Lines
; CHECK-FP16-NEXT: fcvtzu w8, h0 ; CHECK-FP16-NEXT: fcvtzu w8, h0
; CHECK-FP16-NEXT: fmov s0, w8 ; CHECK-FP16-NEXT: fmov s0, w8
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <1 x i32> @llvm.fptoui.sat.v1f16.v1i32(<1 x half> %f) %x = call <1 x i32> @llvm.fptoui.sat.v1f16.v1i32(<1 x half> %f)
ret <1 x i32> %x ret <1 x i32> %x
} }
define <2 x i32> @test_unsigned_v2f16_v2i32(<2 x half> %f) { define <2 x i32> @test_unsigned_v2f16_v2i32(<2 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v2f16_v2i32: ; CHECK-LABEL: test_unsigned_v2f16_v2i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: fcvtzu v0.4s, v0.4s ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 killed $q0 ; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_unsigned_v2f16_v2i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0
; CHECK-FP16-NEXT: mov h1, v0.h[1]
; CHECK-FP16-NEXT: fcvtzu w8, h0
; CHECK-FP16-NEXT: fmov s0, w8
; CHECK-FP16-NEXT: fcvtzu w8, h1
; CHECK-FP16-NEXT: mov v0.s[1], w8
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-FP16-NEXT: ret
%x = call <2 x i32> @llvm.fptoui.sat.v2f16.v2i32(<2 x half> %f) %x = call <2 x i32> @llvm.fptoui.sat.v2f16.v2i32(<2 x half> %f)
ret <2 x i32> %x ret <2 x i32> %x
} }
define <3 x i32> @test_unsigned_v3f16_v3i32(<3 x half> %f) { define <3 x i32> @test_unsigned_v3f16_v3i32(<3 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v3f16_v3i32: ; CHECK-LABEL: test_unsigned_v3f16_v3i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: fcvtzu v0.4s, v0.4s ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_unsigned_v3f16_v3i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0
; CHECK-FP16-NEXT: mov h2, v0.h[1]
; CHECK-FP16-NEXT: fcvtzu w8, h0
; CHECK-FP16-NEXT: fmov s1, w8
; CHECK-FP16-NEXT: fcvtzu w8, h2
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: mov v1.s[1], w8
; CHECK-FP16-NEXT: fcvtzu w8, h2
; CHECK-FP16-NEXT: mov v1.s[2], w8
; CHECK-FP16-NEXT: fcvtzu w8, h0
; CHECK-FP16-NEXT: mov v1.s[3], w8
; CHECK-FP16-NEXT: mov v0.16b, v1.16b
; CHECK-FP16-NEXT: ret
%x = call <3 x i32> @llvm.fptoui.sat.v3f16.v3i32(<3 x half> %f) %x = call <3 x i32> @llvm.fptoui.sat.v3f16.v3i32(<3 x half> %f)
ret <3 x i32> %x ret <3 x i32> %x
} }
define <4 x i32> @test_unsigned_v4f16_v4i32(<4 x half> %f) { define <4 x i32> @test_unsigned_v4f16_v4i32(<4 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v4f16_v4i32: ; CHECK-LABEL: test_unsigned_v4f16_v4i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: fcvtzu v0.4s, v0.4s ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_unsigned_v4f16_v4i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0
; CHECK-FP16-NEXT: mov h2, v0.h[1]
; CHECK-FP16-NEXT: fcvtzu w8, h0
; CHECK-FP16-NEXT: fmov s1, w8
; CHECK-FP16-NEXT: fcvtzu w8, h2
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: mov v1.s[1], w8
; CHECK-FP16-NEXT: fcvtzu w8, h2
; CHECK-FP16-NEXT: mov v1.s[2], w8
; CHECK-FP16-NEXT: fcvtzu w8, h0
; CHECK-FP16-NEXT: mov v1.s[3], w8
; CHECK-FP16-NEXT: mov v0.16b, v1.16b
; CHECK-FP16-NEXT: ret
%x = call <4 x i32> @llvm.fptoui.sat.v4f16.v4i32(<4 x half> %f) %x = call <4 x i32> @llvm.fptoui.sat.v4f16.v4i32(<4 x half> %f)
ret <4 x i32> %x ret <4 x i32> %x
} }
define <5 x i32> @test_unsigned_v5f16_v5i32(<5 x half> %f) { define <5 x i32> @test_unsigned_v5f16_v5i32(<5 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v5f16_v5i32: ; CHECK-LABEL: test_unsigned_v5f16_v5i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v1.4s, v0.4h ; CHECK-NEXT: fcvtl v1.4s, v0.4h
; CHECK-CVT-NEXT: fcvtl2 v0.4s, v0.8h ; CHECK-NEXT: fcvtl2 v0.4s, v0.8h
; CHECK-CVT-NEXT: fcvtzu v1.4s, v1.4s ; CHECK-NEXT: fcvtzu v1.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzu v0.4s, v0.4s ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: mov w1, v1.s[1] ; CHECK-NEXT: mov w1, v1.s[1]
; CHECK-CVT-NEXT: mov w2, v1.s[2] ; CHECK-NEXT: mov w2, v1.s[2]
; CHECK-CVT-NEXT: mov w3, v1.s[3] ; CHECK-NEXT: mov w3, v1.s[3]
; CHECK-CVT-NEXT: fmov w0, s1 ; CHECK-NEXT: fmov w0, s1
; CHECK-CVT-NEXT: fmov w4, s0 ; CHECK-NEXT: fmov w4, s0
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_unsigned_v5f16_v5i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: mov h1, v0.h[1]
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: ext v3.16b, v0.16b, v0.16b, #8
; CHECK-FP16-NEXT: mov h4, v0.h[3]
; CHECK-FP16-NEXT: fcvtzu w0, h0
; CHECK-FP16-NEXT: fcvtzu w1, h1
; CHECK-FP16-NEXT: fcvtzu w2, h2
; CHECK-FP16-NEXT: fcvtzu w4, h3
; CHECK-FP16-NEXT: fcvtzu w3, h4
; CHECK-FP16-NEXT: ret
%x = call <5 x i32> @llvm.fptoui.sat.v5f16.v5i32(<5 x half> %f) %x = call <5 x i32> @llvm.fptoui.sat.v5f16.v5i32(<5 x half> %f)
ret <5 x i32> %x ret <5 x i32> %x
} }
define <6 x i32> @test_unsigned_v6f16_v6i32(<6 x half> %f) { define <6 x i32> @test_unsigned_v6f16_v6i32(<6 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v6f16_v6i32: ; CHECK-LABEL: test_unsigned_v6f16_v6i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v1.4s, v0.4h ; CHECK-NEXT: fcvtl v1.4s, v0.4h
; CHECK-CVT-NEXT: fcvtl2 v0.4s, v0.8h ; CHECK-NEXT: fcvtl2 v0.4s, v0.8h
; CHECK-CVT-NEXT: fcvtzu v1.4s, v1.4s ; CHECK-NEXT: fcvtzu v1.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzu v0.4s, v0.4s ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: mov w1, v1.s[1] ; CHECK-NEXT: mov w1, v1.s[1]
; CHECK-CVT-NEXT: mov w2, v1.s[2] ; CHECK-NEXT: mov w2, v1.s[2]
; CHECK-CVT-NEXT: mov w3, v1.s[3] ; CHECK-NEXT: mov w3, v1.s[3]
; CHECK-CVT-NEXT: mov w5, v0.s[1] ; CHECK-NEXT: mov w5, v0.s[1]
; CHECK-CVT-NEXT: fmov w0, s1 ; CHECK-NEXT: fmov w0, s1
; CHECK-CVT-NEXT: fmov w4, s0 ; CHECK-NEXT: fmov w4, s0
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_unsigned_v6f16_v6i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: ext v1.16b, v0.16b, v0.16b, #8
; CHECK-FP16-NEXT: mov h3, v0.h[2]
; CHECK-FP16-NEXT: mov h4, v0.h[3]
; CHECK-FP16-NEXT: fcvtzu w0, h0
; CHECK-FP16-NEXT: mov h2, v1.h[1]
; CHECK-FP16-NEXT: fcvtzu w8, h1
; CHECK-FP16-NEXT: fcvtzu w2, h3
; CHECK-FP16-NEXT: fcvtzu w3, h4
; CHECK-FP16-NEXT: fmov s1, w8
; CHECK-FP16-NEXT: fcvtzu w5, h2
; CHECK-FP16-NEXT: mov h2, v0.h[1]
; CHECK-FP16-NEXT: mov v1.s[1], w5
; CHECK-FP16-NEXT: fcvtzu w1, h2
; CHECK-FP16-NEXT: fmov w4, s1
; CHECK-FP16-NEXT: ret
%x = call <6 x i32> @llvm.fptoui.sat.v6f16.v6i32(<6 x half> %f) %x = call <6 x i32> @llvm.fptoui.sat.v6f16.v6i32(<6 x half> %f)
ret <6 x i32> %x ret <6 x i32> %x
} }
define <7 x i32> @test_unsigned_v7f16_v7i32(<7 x half> %f) { define <7 x i32> @test_unsigned_v7f16_v7i32(<7 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v7f16_v7i32: ; CHECK-LABEL: test_unsigned_v7f16_v7i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v1.4s, v0.4h ; CHECK-NEXT: fcvtl v1.4s, v0.4h
; CHECK-CVT-NEXT: fcvtl2 v0.4s, v0.8h ; CHECK-NEXT: fcvtl2 v0.4s, v0.8h
; CHECK-CVT-NEXT: fcvtzu v1.4s, v1.4s ; CHECK-NEXT: fcvtzu v1.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzu v0.4s, v0.4s ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: mov w1, v1.s[1] ; CHECK-NEXT: mov w1, v1.s[1]
; CHECK-CVT-NEXT: mov w2, v1.s[2] ; CHECK-NEXT: mov w2, v1.s[2]
; CHECK-CVT-NEXT: mov w3, v1.s[3] ; CHECK-NEXT: mov w3, v1.s[3]
; CHECK-CVT-NEXT: mov w5, v0.s[1] ; CHECK-NEXT: mov w5, v0.s[1]
; CHECK-CVT-NEXT: mov w6, v0.s[2] ; CHECK-NEXT: mov w6, v0.s[2]
; CHECK-CVT-NEXT: fmov w0, s1 ; CHECK-NEXT: fmov w0, s1
; CHECK-CVT-NEXT: fmov w4, s0 ; CHECK-NEXT: fmov w4, s0
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_unsigned_v7f16_v7i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: ext v1.16b, v0.16b, v0.16b, #8
; CHECK-FP16-NEXT: mov h4, v0.h[3]
; CHECK-FP16-NEXT: fcvtzu w0, h0
; CHECK-FP16-NEXT: mov h2, v1.h[1]
; CHECK-FP16-NEXT: fcvtzu w8, h1
; CHECK-FP16-NEXT: mov h1, v1.h[2]
; CHECK-FP16-NEXT: fcvtzu w3, h4
; CHECK-FP16-NEXT: fmov s3, w8
; CHECK-FP16-NEXT: fcvtzu w8, h2
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: fcvtzu w6, h1
; CHECK-FP16-NEXT: mov h1, v0.h[1]
; CHECK-FP16-NEXT: mov v3.s[1], w8
; CHECK-FP16-NEXT: fcvtzu w2, h2
; CHECK-FP16-NEXT: fcvtzu w1, h1
; CHECK-FP16-NEXT: mov v3.s[2], w6
; CHECK-FP16-NEXT: mov w5, v3.s[1]
; CHECK-FP16-NEXT: fmov w4, s3
; CHECK-FP16-NEXT: ret
%x = call <7 x i32> @llvm.fptoui.sat.v7f16.v7i32(<7 x half> %f) %x = call <7 x i32> @llvm.fptoui.sat.v7f16.v7i32(<7 x half> %f)
ret <7 x i32> %x ret <7 x i32> %x
} }
define <8 x i32> @test_unsigned_v8f16_v8i32(<8 x half> %f) { define <8 x i32> @test_unsigned_v8f16_v8i32(<8 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v8f16_v8i32: ; CHECK-LABEL: test_unsigned_v8f16_v8i32:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl2 v1.4s, v0.8h ; CHECK-NEXT: fcvtl2 v1.4s, v0.8h
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: fcvtzu v1.4s, v1.4s ; CHECK-NEXT: fcvtzu v1.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzu v0.4s, v0.4s ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_unsigned_v8f16_v8i32:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: ext v3.16b, v0.16b, v0.16b, #8
; CHECK-FP16-NEXT: mov h4, v0.h[1]
; CHECK-FP16-NEXT: fcvtzu w9, h0
; CHECK-FP16-NEXT: mov h2, v3.h[1]
; CHECK-FP16-NEXT: fcvtzu w8, h3
; CHECK-FP16-NEXT: mov h5, v3.h[2]
; CHECK-FP16-NEXT: mov h3, v3.h[3]
; CHECK-FP16-NEXT: fmov s1, w8
; CHECK-FP16-NEXT: fcvtzu w8, h2
; CHECK-FP16-NEXT: fmov s2, w9
; CHECK-FP16-NEXT: fcvtzu w9, h4
; CHECK-FP16-NEXT: mov h4, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: mov v1.s[1], w8
; CHECK-FP16-NEXT: fcvtzu w8, h5
; CHECK-FP16-NEXT: mov v2.s[1], w9
; CHECK-FP16-NEXT: fcvtzu w9, h4
; CHECK-FP16-NEXT: mov v1.s[2], w8
; CHECK-FP16-NEXT: fcvtzu w8, h3
; CHECK-FP16-NEXT: mov v2.s[2], w9
; CHECK-FP16-NEXT: fcvtzu w9, h0
; CHECK-FP16-NEXT: mov v1.s[3], w8
; CHECK-FP16-NEXT: mov v2.s[3], w9
; CHECK-FP16-NEXT: mov v0.16b, v2.16b
; CHECK-FP16-NEXT: ret
%x = call <8 x i32> @llvm.fptoui.sat.v8f16.v8i32(<8 x half> %f) %x = call <8 x i32> @llvm.fptoui.sat.v8f16.v8i32(<8 x half> %f)
ret <8 x i32> %x ret <8 x i32> %x
} }
; ;
; 2-Vector float to unsigned integer -- result size variation ; 2-Vector float to unsigned integer -- result size variation
; ;
declare <2 x i1> @llvm.fptoui.sat.v2f32.v2i1 (<2 x float>) declare <2 x i1> @llvm.fptoui.sat.v2f32.v2i1 (<2 x float>)
declare <2 x i8> @llvm.fptoui.sat.v2f32.v2i8 (<2 x float>) declare <2 x i8> @llvm.fptoui.sat.v2f32.v2i8 (<2 x float>)
declare <2 x i13> @llvm.fptoui.sat.v2f32.v2i13 (<2 x float>) declare <2 x i13> @llvm.fptoui.sat.v2f32.v2i13 (<2 x float>)
declare <2 x i16> @llvm.fptoui.sat.v2f32.v2i16 (<2 x float>) declare <2 x i16> @llvm.fptoui.sat.v2f32.v2i16 (<2 x float>)
declare <2 x i19> @llvm.fptoui.sat.v2f32.v2i19 (<2 x float>) declare <2 x i19> @llvm.fptoui.sat.v2f32.v2i19 (<2 x float>)
declare <2 x i50> @llvm.fptoui.sat.v2f32.v2i50 (<2 x float>) declare <2 x i50> @llvm.fptoui.sat.v2f32.v2i50 (<2 x float>)
declare <2 x i64> @llvm.fptoui.sat.v2f32.v2i64 (<2 x float>) declare <2 x i64> @llvm.fptoui.sat.v2f32.v2i64 (<2 x float>)
declare <2 x i100> @llvm.fptoui.sat.v2f32.v2i100(<2 x float>) declare <2 x i100> @llvm.fptoui.sat.v2f32.v2i100(<2 x float>)
declare <2 x i128> @llvm.fptoui.sat.v2f32.v2i128(<2 x float>) declare <2 x i128> @llvm.fptoui.sat.v2f32.v2i128(<2 x float>)
define <2 x i1> @test_unsigned_v2f32_v2i1(<2 x float> %f) { define <2 x i1> @test_unsigned_v2f32_v2i1(<2 x float> %f) {
; CHECK-LABEL: test_unsigned_v2f32_v2i1: ; CHECK-LABEL: test_unsigned_v2f32_v2i1:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-NEXT: movi v1.2s, #1
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: fcvtzu v0.2s, v0.2s
; CHECK-NEXT: fcvtzu w9, s0 ; CHECK-NEXT: umin v0.2s, v0.2s, v1.2s
; CHECK-NEXT: fcvtzu w8, s1
; CHECK-NEXT: cmp w8, #1
; CHECK-NEXT: csinc w8, w8, wzr, lo
; CHECK-NEXT: cmp w9, #1
; CHECK-NEXT: csinc w9, w9, wzr, lo
; CHECK-NEXT: fmov s0, w9
; CHECK-NEXT: mov v0.s[1], w8
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <2 x i1> @llvm.fptoui.sat.v2f32.v2i1(<2 x float> %f) %x = call <2 x i1> @llvm.fptoui.sat.v2f32.v2i1(<2 x float> %f)
ret <2 x i1> %x ret <2 x i1> %x
} }
define <2 x i8> @test_unsigned_v2f32_v2i8(<2 x float> %f) { define <2 x i8> @test_unsigned_v2f32_v2i8(<2 x float> %f) {
; CHECK-LABEL: test_unsigned_v2f32_v2i8: ; CHECK-LABEL: test_unsigned_v2f32_v2i8:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-NEXT: movi d1, #0x0000ff000000ff
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: fcvtzu v0.2s, v0.2s
; CHECK-NEXT: fcvtzu w10, s0 ; CHECK-NEXT: umin v0.2s, v0.2s, v1.2s
; CHECK-NEXT: mov w8, #255
; CHECK-NEXT: fcvtzu w9, s1
; CHECK-NEXT: cmp w9, #255
; CHECK-NEXT: csel w9, w9, w8, lo
; CHECK-NEXT: cmp w10, #255
; CHECK-NEXT: csel w8, w10, w8, lo
; CHECK-NEXT: fmov s0, w8
; CHECK-NEXT: mov v0.s[1], w9
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <2 x i8> @llvm.fptoui.sat.v2f32.v2i8(<2 x float> %f) %x = call <2 x i8> @llvm.fptoui.sat.v2f32.v2i8(<2 x float> %f)
ret <2 x i8> %x ret <2 x i8> %x
} }
define <2 x i13> @test_unsigned_v2f32_v2i13(<2 x float> %f) { define <2 x i13> @test_unsigned_v2f32_v2i13(<2 x float> %f) {
; CHECK-LABEL: test_unsigned_v2f32_v2i13: ; CHECK-LABEL: test_unsigned_v2f32_v2i13:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-NEXT: movi v1.2s, #31, msl #8
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: fcvtzu v0.2s, v0.2s
; CHECK-NEXT: fcvtzu w9, s0 ; CHECK-NEXT: umin v0.2s, v0.2s, v1.2s
; CHECK-NEXT: mov w10, #8191
; CHECK-NEXT: fcvtzu w8, s1
; CHECK-NEXT: cmp w8, w10
; CHECK-NEXT: csel w8, w8, w10, lo
; CHECK-NEXT: cmp w9, w10
; CHECK-NEXT: csel w9, w9, w10, lo
; CHECK-NEXT: fmov s0, w9
; CHECK-NEXT: mov v0.s[1], w8
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <2 x i13> @llvm.fptoui.sat.v2f32.v2i13(<2 x float> %f) %x = call <2 x i13> @llvm.fptoui.sat.v2f32.v2i13(<2 x float> %f)
ret <2 x i13> %x ret <2 x i13> %x
} }
define <2 x i16> @test_unsigned_v2f32_v2i16(<2 x float> %f) { define <2 x i16> @test_unsigned_v2f32_v2i16(<2 x float> %f) {
; CHECK-LABEL: test_unsigned_v2f32_v2i16: ; CHECK-LABEL: test_unsigned_v2f32_v2i16:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-NEXT: movi d1, #0x00ffff0000ffff
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: fcvtzu v0.2s, v0.2s
; CHECK-NEXT: fcvtzu w9, s0 ; CHECK-NEXT: umin v0.2s, v0.2s, v1.2s
; CHECK-NEXT: mov w10, #65535
; CHECK-NEXT: fcvtzu w8, s1
; CHECK-NEXT: cmp w8, w10
; CHECK-NEXT: csel w8, w8, w10, lo
; CHECK-NEXT: cmp w9, w10
; CHECK-NEXT: csel w9, w9, w10, lo
; CHECK-NEXT: fmov s0, w9
; CHECK-NEXT: mov v0.s[1], w8
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <2 x i16> @llvm.fptoui.sat.v2f32.v2i16(<2 x float> %f) %x = call <2 x i16> @llvm.fptoui.sat.v2f32.v2i16(<2 x float> %f)
ret <2 x i16> %x ret <2 x i16> %x
} }
define <2 x i19> @test_unsigned_v2f32_v2i19(<2 x float> %f) { define <2 x i19> @test_unsigned_v2f32_v2i19(<2 x float> %f) {
; CHECK-LABEL: test_unsigned_v2f32_v2i19: ; CHECK-LABEL: test_unsigned_v2f32_v2i19:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-NEXT: movi v1.2s, #7, msl #16
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: fcvtzu v0.2s, v0.2s
; CHECK-NEXT: fcvtzu w9, s0 ; CHECK-NEXT: umin v0.2s, v0.2s, v1.2s
; CHECK-NEXT: mov w10, #524287
; CHECK-NEXT: fcvtzu w8, s1
; CHECK-NEXT: cmp w8, w10
; CHECK-NEXT: csel w8, w8, w10, lo
; CHECK-NEXT: cmp w9, w10
; CHECK-NEXT: csel w9, w9, w10, lo
; CHECK-NEXT: fmov s0, w9
; CHECK-NEXT: mov v0.s[1], w8
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <2 x i19> @llvm.fptoui.sat.v2f32.v2i19(<2 x float> %f) %x = call <2 x i19> @llvm.fptoui.sat.v2f32.v2i19(<2 x float> %f)
ret <2 x i19> %x ret <2 x i19> %x
} }
define <2 x i32> @test_unsigned_v2f32_v2i32_duplicate(<2 x float> %f) { define <2 x i32> @test_unsigned_v2f32_v2i32_duplicate(<2 x float> %f) {
; CHECK-LABEL: test_unsigned_v2f32_v2i32_duplicate: ; CHECK-LABEL: test_unsigned_v2f32_v2i32_duplicate:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
▲ Show 20 Lines • Show All 150 Lines • ▼ Show 20 Lines
declare <4 x i50> @llvm.fptoui.sat.v4f32.v4i50 (<4 x float>) declare <4 x i50> @llvm.fptoui.sat.v4f32.v4i50 (<4 x float>)
declare <4 x i64> @llvm.fptoui.sat.v4f32.v4i64 (<4 x float>) declare <4 x i64> @llvm.fptoui.sat.v4f32.v4i64 (<4 x float>)
declare <4 x i100> @llvm.fptoui.sat.v4f32.v4i100(<4 x float>) declare <4 x i100> @llvm.fptoui.sat.v4f32.v4i100(<4 x float>)
declare <4 x i128> @llvm.fptoui.sat.v4f32.v4i128(<4 x float>) declare <4 x i128> @llvm.fptoui.sat.v4f32.v4i128(<4 x float>)
define <4 x i1> @test_unsigned_v4f32_v4i1(<4 x float> %f) { define <4 x i1> @test_unsigned_v4f32_v4i1(<4 x float> %f) {
; CHECK-LABEL: test_unsigned_v4f32_v4i1: ; CHECK-LABEL: test_unsigned_v4f32_v4i1:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: movi v1.4s, #1
; CHECK-NEXT: fcvtzu w9, s0 ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-NEXT: mov s2, v0.s[2] ; CHECK-NEXT: umin v0.4s, v0.4s, v1.4s
; CHECK-NEXT: mov s0, v0.s[3] ; CHECK-NEXT: xtn v0.4h, v0.4s
; CHECK-NEXT: fcvtzu w8, s1
; CHECK-NEXT: cmp w8, #1
; CHECK-NEXT: csinc w8, w8, wzr, lo
; CHECK-NEXT: cmp w9, #1
; CHECK-NEXT: csinc w9, w9, wzr, lo
; CHECK-NEXT: fmov s1, w9
; CHECK-NEXT: fcvtzu w9, s2
; CHECK-NEXT: mov v1.h[1], w8
; CHECK-NEXT: cmp w9, #1
; CHECK-NEXT: csinc w8, w9, wzr, lo
; CHECK-NEXT: fcvtzu w9, s0
; CHECK-NEXT: mov v1.h[2], w8
; CHECK-NEXT: cmp w9, #1
; CHECK-NEXT: csinc w8, w9, wzr, lo
; CHECK-NEXT: mov v1.h[3], w8
; CHECK-NEXT: fmov d0, d1
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <4 x i1> @llvm.fptoui.sat.v4f32.v4i1(<4 x float> %f) %x = call <4 x i1> @llvm.fptoui.sat.v4f32.v4i1(<4 x float> %f)
ret <4 x i1> %x ret <4 x i1> %x
} }
define <4 x i8> @test_unsigned_v4f32_v4i8(<4 x float> %f) { define <4 x i8> @test_unsigned_v4f32_v4i8(<4 x float> %f) {
; CHECK-LABEL: test_unsigned_v4f32_v4i8: ; CHECK-LABEL: test_unsigned_v4f32_v4i8:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: movi v1.2d, #0x0000ff000000ff
; CHECK-NEXT: fcvtzu w10, s0 ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-NEXT: mov w8, #255 ; CHECK-NEXT: umin v0.4s, v0.4s, v1.4s
; CHECK-NEXT: mov s2, v0.s[2] ; CHECK-NEXT: xtn v0.4h, v0.4s
; CHECK-NEXT: mov s0, v0.s[3]
; CHECK-NEXT: fcvtzu w9, s1
; CHECK-NEXT: cmp w9, #255
; CHECK-NEXT: csel w9, w9, w8, lo
; CHECK-NEXT: cmp w10, #255
; CHECK-NEXT: csel w10, w10, w8, lo
; CHECK-NEXT: fmov s1, w10
; CHECK-NEXT: fcvtzu w10, s2
; CHECK-NEXT: mov v1.h[1], w9
; CHECK-NEXT: cmp w10, #255
; CHECK-NEXT: csel w9, w10, w8, lo
; CHECK-NEXT: fcvtzu w10, s0
; CHECK-NEXT: mov v1.h[2], w9
; CHECK-NEXT: cmp w10, #255
; CHECK-NEXT: csel w8, w10, w8, lo
; CHECK-NEXT: mov v1.h[3], w8
; CHECK-NEXT: fmov d0, d1
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <4 x i8> @llvm.fptoui.sat.v4f32.v4i8(<4 x float> %f) %x = call <4 x i8> @llvm.fptoui.sat.v4f32.v4i8(<4 x float> %f)
ret <4 x i8> %x ret <4 x i8> %x
} }
define <4 x i13> @test_unsigned_v4f32_v4i13(<4 x float> %f) { define <4 x i13> @test_unsigned_v4f32_v4i13(<4 x float> %f) {
; CHECK-LABEL: test_unsigned_v4f32_v4i13: ; CHECK-LABEL: test_unsigned_v4f32_v4i13:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: movi v1.4s, #31, msl #8
; CHECK-NEXT: fcvtzu w10, s0 ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-NEXT: mov w8, #8191 ; CHECK-NEXT: umin v0.4s, v0.4s, v1.4s
; CHECK-NEXT: mov s2, v0.s[2] ; CHECK-NEXT: xtn v0.4h, v0.4s
; CHECK-NEXT: mov s0, v0.s[3]
; CHECK-NEXT: fcvtzu w9, s1
; CHECK-NEXT: cmp w9, w8
; CHECK-NEXT: csel w9, w9, w8, lo
; CHECK-NEXT: cmp w10, w8
; CHECK-NEXT: csel w10, w10, w8, lo
; CHECK-NEXT: fmov s1, w10
; CHECK-NEXT: fcvtzu w10, s2
; CHECK-NEXT: mov v1.h[1], w9
; CHECK-NEXT: cmp w10, w8
; CHECK-NEXT: csel w9, w10, w8, lo
; CHECK-NEXT: fcvtzu w10, s0
; CHECK-NEXT: mov v1.h[2], w9
; CHECK-NEXT: cmp w10, w8
; CHECK-NEXT: csel w8, w10, w8, lo
; CHECK-NEXT: mov v1.h[3], w8
; CHECK-NEXT: fmov d0, d1
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <4 x i13> @llvm.fptoui.sat.v4f32.v4i13(<4 x float> %f) %x = call <4 x i13> @llvm.fptoui.sat.v4f32.v4i13(<4 x float> %f)
ret <4 x i13> %x ret <4 x i13> %x
} }
define <4 x i16> @test_unsigned_v4f32_v4i16(<4 x float> %f) { define <4 x i16> @test_unsigned_v4f32_v4i16(<4 x float> %f) {
; CHECK-LABEL: test_unsigned_v4f32_v4i16: ; CHECK-LABEL: test_unsigned_v4f32_v4i16:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-NEXT: fcvtzu w10, s0 ; CHECK-NEXT: uqxtn v0.4h, v0.4s
; CHECK-NEXT: mov w8, #65535
; CHECK-NEXT: mov s2, v0.s[2]
; CHECK-NEXT: mov s0, v0.s[3]
; CHECK-NEXT: fcvtzu w9, s1
; CHECK-NEXT: cmp w9, w8
; CHECK-NEXT: csel w9, w9, w8, lo
; CHECK-NEXT: cmp w10, w8
; CHECK-NEXT: csel w10, w10, w8, lo
; CHECK-NEXT: fmov s1, w10
; CHECK-NEXT: fcvtzu w10, s2
; CHECK-NEXT: mov v1.h[1], w9
; CHECK-NEXT: cmp w10, w8
; CHECK-NEXT: csel w9, w10, w8, lo
; CHECK-NEXT: fcvtzu w10, s0
; CHECK-NEXT: mov v1.h[2], w9
; CHECK-NEXT: cmp w10, w8
; CHECK-NEXT: csel w8, w10, w8, lo
; CHECK-NEXT: mov v1.h[3], w8
; CHECK-NEXT: fmov d0, d1
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <4 x i16> @llvm.fptoui.sat.v4f32.v4i16(<4 x float> %f) %x = call <4 x i16> @llvm.fptoui.sat.v4f32.v4i16(<4 x float> %f)
ret <4 x i16> %x ret <4 x i16> %x
} }
define <4 x i19> @test_unsigned_v4f32_v4i19(<4 x float> %f) { define <4 x i19> @test_unsigned_v4f32_v4i19(<4 x float> %f) {
; CHECK-LABEL: test_unsigned_v4f32_v4i19: ; CHECK-LABEL: test_unsigned_v4f32_v4i19:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
; CHECK-NEXT: mov s1, v0.s[1] ; CHECK-NEXT: movi v1.4s, #7, msl #16
; CHECK-NEXT: fcvtzu w10, s0 ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-NEXT: mov w8, #524287 ; CHECK-NEXT: umin v0.4s, v0.4s, v1.4s
; CHECK-NEXT: mov s2, v0.s[2]
; CHECK-NEXT: mov s0, v0.s[3]
; CHECK-NEXT: fcvtzu w9, s1
; CHECK-NEXT: cmp w9, w8
; CHECK-NEXT: csel w9, w9, w8, lo
; CHECK-NEXT: cmp w10, w8
; CHECK-NEXT: csel w10, w10, w8, lo
; CHECK-NEXT: fmov s1, w10
; CHECK-NEXT: fcvtzu w10, s2
; CHECK-NEXT: mov v1.s[1], w9
; CHECK-NEXT: cmp w10, w8
; CHECK-NEXT: csel w9, w10, w8, lo
; CHECK-NEXT: fcvtzu w10, s0
; CHECK-NEXT: mov v1.s[2], w9
; CHECK-NEXT: cmp w10, w8
; CHECK-NEXT: csel w8, w10, w8, lo
; CHECK-NEXT: mov v1.s[3], w8
; CHECK-NEXT: mov v0.16b, v1.16b
; CHECK-NEXT: ret ; CHECK-NEXT: ret
%x = call <4 x i19> @llvm.fptoui.sat.v4f32.v4i19(<4 x float> %f) %x = call <4 x i19> @llvm.fptoui.sat.v4f32.v4i19(<4 x float> %f)
ret <4 x i19> %x ret <4 x i19> %x
} }
define <4 x i32> @test_unsigned_v4f32_v4i32_duplicate(<4 x float> %f) { define <4 x i32> @test_unsigned_v4f32_v4i32_duplicate(<4 x float> %f) {
; CHECK-LABEL: test_unsigned_v4f32_v4i32_duplicate: ; CHECK-LABEL: test_unsigned_v4f32_v4i32_duplicate:
; CHECK: // %bb.0: ; CHECK: // %bb.0:
▲ Show 20 Lines • Show All 473 Lines • ▼ Show 20 Lines
declare <4 x i50> @llvm.fptoui.sat.v4f16.v4i50 (<4 x half>) declare <4 x i50> @llvm.fptoui.sat.v4f16.v4i50 (<4 x half>)
declare <4 x i64> @llvm.fptoui.sat.v4f16.v4i64 (<4 x half>) declare <4 x i64> @llvm.fptoui.sat.v4f16.v4i64 (<4 x half>)
declare <4 x i100> @llvm.fptoui.sat.v4f16.v4i100(<4 x half>) declare <4 x i100> @llvm.fptoui.sat.v4f16.v4i100(<4 x half>)
declare <4 x i128> @llvm.fptoui.sat.v4f16.v4i128(<4 x half>) declare <4 x i128> @llvm.fptoui.sat.v4f16.v4i128(<4 x half>)
define <4 x i1> @test_unsigned_v4f16_v4i1(<4 x half> %f) { define <4 x i1> @test_unsigned_v4f16_v4i1(<4 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v4f16_v4i1: ; CHECK-CVT-LABEL: test_unsigned_v4f16_v4i1:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov h1, v0.h[1] ; CHECK-CVT-NEXT: movi v1.4s, #1
; CHECK-CVT-NEXT: fcvt s2, h0 ; CHECK-CVT-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-CVT-NEXT: umin v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzu w9, s2 ; CHECK-CVT-NEXT: xtn v0.4h, v0.4s
; CHECK-CVT-NEXT: mov h2, v0.h[3]
; CHECK-CVT-NEXT: fcvtzu w8, s1
; CHECK-CVT-NEXT: mov h1, v0.h[2]
; CHECK-CVT-NEXT: cmp w8, #1
; CHECK-CVT-NEXT: csinc w8, w8, wzr, lo
; CHECK-CVT-NEXT: cmp w9, #1
; CHECK-CVT-NEXT: fcvt s1, h1
; CHECK-CVT-NEXT: csinc w9, w9, wzr, lo
; CHECK-CVT-NEXT: fmov s0, w9
; CHECK-CVT-NEXT: fcvtzu w9, s1
; CHECK-CVT-NEXT: fcvt s1, h2
; CHECK-CVT-NEXT: mov v0.h[1], w8
; CHECK-CVT-NEXT: cmp w9, #1
; CHECK-CVT-NEXT: csinc w8, w9, wzr, lo
; CHECK-CVT-NEXT: fcvtzu w9, s1
; CHECK-CVT-NEXT: mov v0.h[2], w8
; CHECK-CVT-NEXT: cmp w9, #1
; CHECK-CVT-NEXT: csinc w8, w9, wzr, lo
; CHECK-CVT-NEXT: mov v0.h[3], w8
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_unsigned_v4f16_v4i1: ; CHECK-FP16-LABEL: test_unsigned_v4f16_v4i1:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-FP16-NEXT: movi v1.4h, #1
; CHECK-FP16-NEXT: mov h1, v0.h[1] ; CHECK-FP16-NEXT: fcvtzu v0.4h, v0.4h
; CHECK-FP16-NEXT: fcvtzu w9, h0 ; CHECK-FP16-NEXT: umin v0.4h, v0.4h, v1.4h
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: fcvtzu w8, h1
; CHECK-FP16-NEXT: cmp w8, #1
; CHECK-FP16-NEXT: csinc w8, w8, wzr, lo
; CHECK-FP16-NEXT: cmp w9, #1
; CHECK-FP16-NEXT: csinc w9, w9, wzr, lo
; CHECK-FP16-NEXT: fmov s1, w9
; CHECK-FP16-NEXT: fcvtzu w9, h2
; CHECK-FP16-NEXT: mov v1.h[1], w8
; CHECK-FP16-NEXT: cmp w9, #1
; CHECK-FP16-NEXT: csinc w8, w9, wzr, lo
; CHECK-FP16-NEXT: fcvtzu w9, h0
; CHECK-FP16-NEXT: mov v1.h[2], w8
; CHECK-FP16-NEXT: cmp w9, #1
; CHECK-FP16-NEXT: csinc w8, w9, wzr, lo
; CHECK-FP16-NEXT: mov v1.h[3], w8
; CHECK-FP16-NEXT: fmov d0, d1
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <4 x i1> @llvm.fptoui.sat.v4f16.v4i1(<4 x half> %f) %x = call <4 x i1> @llvm.fptoui.sat.v4f16.v4i1(<4 x half> %f)
ret <4 x i1> %x ret <4 x i1> %x
} }
define <4 x i8> @test_unsigned_v4f16_v4i8(<4 x half> %f) { define <4 x i8> @test_unsigned_v4f16_v4i8(<4 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v4f16_v4i8: ; CHECK-CVT-LABEL: test_unsigned_v4f16_v4i8:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov h1, v0.h[1] ; CHECK-CVT-NEXT: movi v1.2d, #0x0000ff000000ff
; CHECK-CVT-NEXT: fcvt s2, h0 ; CHECK-CVT-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: mov w8, #255 ; CHECK-CVT-NEXT: umin v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-CVT-NEXT: xtn v0.4h, v0.4s
; CHECK-CVT-NEXT: fcvtzu w10, s2
; CHECK-CVT-NEXT: mov h2, v0.h[3]
; CHECK-CVT-NEXT: fcvtzu w9, s1
; CHECK-CVT-NEXT: mov h1, v0.h[2]
; CHECK-CVT-NEXT: cmp w9, #255
; CHECK-CVT-NEXT: csel w9, w9, w8, lo
; CHECK-CVT-NEXT: cmp w10, #255
; CHECK-CVT-NEXT: fcvt s1, h1
; CHECK-CVT-NEXT: csel w10, w10, w8, lo
; CHECK-CVT-NEXT: fmov s0, w10
; CHECK-CVT-NEXT: fcvtzu w10, s1
; CHECK-CVT-NEXT: fcvt s1, h2
; CHECK-CVT-NEXT: mov v0.h[1], w9
; CHECK-CVT-NEXT: cmp w10, #255
; CHECK-CVT-NEXT: csel w9, w10, w8, lo
; CHECK-CVT-NEXT: fcvtzu w10, s1
; CHECK-CVT-NEXT: mov v0.h[2], w9
; CHECK-CVT-NEXT: cmp w10, #255
; CHECK-CVT-NEXT: csel w8, w10, w8, lo
; CHECK-CVT-NEXT: mov v0.h[3], w8
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_unsigned_v4f16_v4i8: ; CHECK-FP16-LABEL: test_unsigned_v4f16_v4i8:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-FP16-NEXT: movi d1, #0xff00ff00ff00ff
; CHECK-FP16-NEXT: mov h1, v0.h[1] ; CHECK-FP16-NEXT: fcvtzu v0.4h, v0.4h
; CHECK-FP16-NEXT: fcvtzu w10, h0 ; CHECK-FP16-NEXT: umin v0.4h, v0.4h, v1.4h
; CHECK-FP16-NEXT: mov w8, #255
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: fcvtzu w9, h1
; CHECK-FP16-NEXT: cmp w9, #255
; CHECK-FP16-NEXT: csel w9, w9, w8, lo
; CHECK-FP16-NEXT: cmp w10, #255
; CHECK-FP16-NEXT: csel w10, w10, w8, lo
; CHECK-FP16-NEXT: fmov s1, w10
; CHECK-FP16-NEXT: fcvtzu w10, h2
; CHECK-FP16-NEXT: mov v1.h[1], w9
; CHECK-FP16-NEXT: cmp w10, #255
; CHECK-FP16-NEXT: csel w9, w10, w8, lo
; CHECK-FP16-NEXT: fcvtzu w10, h0
; CHECK-FP16-NEXT: mov v1.h[2], w9
; CHECK-FP16-NEXT: cmp w10, #255
; CHECK-FP16-NEXT: csel w8, w10, w8, lo
; CHECK-FP16-NEXT: mov v1.h[3], w8
; CHECK-FP16-NEXT: fmov d0, d1
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <4 x i8> @llvm.fptoui.sat.v4f16.v4i8(<4 x half> %f) %x = call <4 x i8> @llvm.fptoui.sat.v4f16.v4i8(<4 x half> %f)
ret <4 x i8> %x ret <4 x i8> %x
} }
define <4 x i13> @test_unsigned_v4f16_v4i13(<4 x half> %f) { define <4 x i13> @test_unsigned_v4f16_v4i13(<4 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v4f16_v4i13: ; CHECK-CVT-LABEL: test_unsigned_v4f16_v4i13:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov h1, v0.h[1] ; CHECK-CVT-NEXT: movi v1.4s, #31, msl #8
; CHECK-CVT-NEXT: fcvt s2, h0 ; CHECK-CVT-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: mov w8, #8191 ; CHECK-CVT-NEXT: umin v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-CVT-NEXT: xtn v0.4h, v0.4s
; CHECK-CVT-NEXT: fcvtzu w10, s2
; CHECK-CVT-NEXT: mov h2, v0.h[3]
; CHECK-CVT-NEXT: fcvtzu w9, s1
; CHECK-CVT-NEXT: mov h1, v0.h[2]
; CHECK-CVT-NEXT: cmp w9, w8
; CHECK-CVT-NEXT: csel w9, w9, w8, lo
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: fcvt s1, h1
; CHECK-CVT-NEXT: csel w10, w10, w8, lo
; CHECK-CVT-NEXT: fmov s0, w10
; CHECK-CVT-NEXT: fcvtzu w10, s1
; CHECK-CVT-NEXT: fcvt s1, h2
; CHECK-CVT-NEXT: mov v0.h[1], w9
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w9, w10, w8, lo
; CHECK-CVT-NEXT: fcvtzu w10, s1
; CHECK-CVT-NEXT: mov v0.h[2], w9
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w8, w10, w8, lo
; CHECK-CVT-NEXT: mov v0.h[3], w8
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_unsigned_v4f16_v4i13: ; CHECK-FP16-LABEL: test_unsigned_v4f16_v4i13:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-FP16-NEXT: mvni v1.4h, #224, lsl #8
; CHECK-FP16-NEXT: mov h1, v0.h[1] ; CHECK-FP16-NEXT: fcvtzu v0.4h, v0.4h
; CHECK-FP16-NEXT: fcvtzu w10, h0 ; CHECK-FP16-NEXT: umin v0.4h, v0.4h, v1.4h
; CHECK-FP16-NEXT: mov w8, #8191
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: fcvtzu w9, h1
; CHECK-FP16-NEXT: cmp w9, w8
; CHECK-FP16-NEXT: csel w9, w9, w8, lo
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w10, w10, w8, lo
; CHECK-FP16-NEXT: fmov s1, w10
; CHECK-FP16-NEXT: fcvtzu w10, h2
; CHECK-FP16-NEXT: mov v1.h[1], w9
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w9, w10, w8, lo
; CHECK-FP16-NEXT: fcvtzu w10, h0
; CHECK-FP16-NEXT: mov v1.h[2], w9
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w8, w10, w8, lo
; CHECK-FP16-NEXT: mov v1.h[3], w8
; CHECK-FP16-NEXT: fmov d0, d1
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <4 x i13> @llvm.fptoui.sat.v4f16.v4i13(<4 x half> %f) %x = call <4 x i13> @llvm.fptoui.sat.v4f16.v4i13(<4 x half> %f)
ret <4 x i13> %x ret <4 x i13> %x
} }
define <4 x i16> @test_unsigned_v4f16_v4i16(<4 x half> %f) { define <4 x i16> @test_unsigned_v4f16_v4i16(<4 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v4f16_v4i16: ; CHECK-CVT-LABEL: test_unsigned_v4f16_v4i16:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v1.4s, v0.4h ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov w8, #65535 ; CHECK-CVT-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: mov s0, v1.s[1] ; CHECK-CVT-NEXT: uqxtn v0.4h, v0.4s
; CHECK-CVT-NEXT: fcvtzu w10, s1
; CHECK-CVT-NEXT: mov s2, v1.s[2]
; CHECK-CVT-NEXT: mov s1, v1.s[3]
; CHECK-CVT-NEXT: fcvtzu w9, s0
; CHECK-CVT-NEXT: cmp w9, w8
; CHECK-CVT-NEXT: csel w9, w9, w8, lo
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w10, w10, w8, lo
; CHECK-CVT-NEXT: fmov s0, w10
; CHECK-CVT-NEXT: fcvtzu w10, s2
; CHECK-CVT-NEXT: mov v0.h[1], w9
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w9, w10, w8, lo
; CHECK-CVT-NEXT: fcvtzu w10, s1
; CHECK-CVT-NEXT: mov v0.h[2], w9
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w8, w10, w8, lo
; CHECK-CVT-NEXT: mov v0.h[3], w8
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_unsigned_v4f16_v4i16: ; CHECK-FP16-LABEL: test_unsigned_v4f16_v4i16:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: fcvtzu v0.4h, v0.4h ; CHECK-FP16-NEXT: fcvtzu v0.4h, v0.4h
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <4 x i16> @llvm.fptoui.sat.v4f16.v4i16(<4 x half> %f) %x = call <4 x i16> @llvm.fptoui.sat.v4f16.v4i16(<4 x half> %f)
ret <4 x i16> %x ret <4 x i16> %x
} }
define <4 x i19> @test_unsigned_v4f16_v4i19(<4 x half> %f) { define <4 x i19> @test_unsigned_v4f16_v4i19(<4 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v4f16_v4i19: ; CHECK-LABEL: test_unsigned_v4f16_v4i19:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov h1, v0.h[1] ; CHECK-NEXT: movi v1.4s, #7, msl #16
; CHECK-CVT-NEXT: fcvt s2, h0 ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: mov w8, #524287 ; CHECK-NEXT: umin v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-NEXT: ret
; CHECK-CVT-NEXT: fcvtzu w10, s2
; CHECK-CVT-NEXT: mov h2, v0.h[3]
; CHECK-CVT-NEXT: fcvtzu w9, s1
; CHECK-CVT-NEXT: mov h1, v0.h[2]
; CHECK-CVT-NEXT: cmp w9, w8
; CHECK-CVT-NEXT: csel w9, w9, w8, lo
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: fcvt s1, h1
; CHECK-CVT-NEXT: csel w10, w10, w8, lo
; CHECK-CVT-NEXT: fmov s0, w10
; CHECK-CVT-NEXT: fcvtzu w10, s1
; CHECK-CVT-NEXT: fcvt s1, h2
; CHECK-CVT-NEXT: mov v0.s[1], w9
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w9, w10, w8, lo
; CHECK-CVT-NEXT: fcvtzu w10, s1
; CHECK-CVT-NEXT: mov v0.s[2], w9
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w8, w10, w8, lo
; CHECK-CVT-NEXT: mov v0.s[3], w8
; CHECK-CVT-NEXT: ret
;
; CHECK-FP16-LABEL: test_unsigned_v4f16_v4i19:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0
; CHECK-FP16-NEXT: mov h1, v0.h[1]
; CHECK-FP16-NEXT: fcvtzu w10, h0
; CHECK-FP16-NEXT: mov w8, #524287
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: fcvtzu w9, h1
; CHECK-FP16-NEXT: cmp w9, w8
; CHECK-FP16-NEXT: csel w9, w9, w8, lo
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w10, w10, w8, lo
; CHECK-FP16-NEXT: fmov s1, w10
; CHECK-FP16-NEXT: fcvtzu w10, h2
; CHECK-FP16-NEXT: mov v1.s[1], w9
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w9, w10, w8, lo
; CHECK-FP16-NEXT: fcvtzu w10, h0
; CHECK-FP16-NEXT: mov v1.s[2], w9
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w8, w10, w8, lo
; CHECK-FP16-NEXT: mov v1.s[3], w8
; CHECK-FP16-NEXT: mov v0.16b, v1.16b
; CHECK-FP16-NEXT: ret
%x = call <4 x i19> @llvm.fptoui.sat.v4f16.v4i19(<4 x half> %f) %x = call <4 x i19> @llvm.fptoui.sat.v4f16.v4i19(<4 x half> %f)
ret <4 x i19> %x ret <4 x i19> %x
} }
define <4 x i32> @test_unsigned_v4f16_v4i32_duplicate(<4 x half> %f) { define <4 x i32> @test_unsigned_v4f16_v4i32_duplicate(<4 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v4f16_v4i32_duplicate: ; CHECK-LABEL: test_unsigned_v4f16_v4i32_duplicate:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: fcvtzu v0.4s, v0.4s ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_unsigned_v4f16_v4i32_duplicate:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: // kill: def $d0 killed $d0 def $q0
; CHECK-FP16-NEXT: mov h2, v0.h[1]
; CHECK-FP16-NEXT: fcvtzu w8, h0
; CHECK-FP16-NEXT: fmov s1, w8
; CHECK-FP16-NEXT: fcvtzu w8, h2
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: mov v1.s[1], w8
; CHECK-FP16-NEXT: fcvtzu w8, h2
; CHECK-FP16-NEXT: mov v1.s[2], w8
; CHECK-FP16-NEXT: fcvtzu w8, h0
; CHECK-FP16-NEXT: mov v1.s[3], w8
; CHECK-FP16-NEXT: mov v0.16b, v1.16b
; CHECK-FP16-NEXT: ret
%x = call <4 x i32> @llvm.fptoui.sat.v4f16.v4i32(<4 x half> %f) %x = call <4 x i32> @llvm.fptoui.sat.v4f16.v4i32(<4 x half> %f)
ret <4 x i32> %x ret <4 x i32> %x
} }
define <4 x i50> @test_unsigned_v4f16_v4i50(<4 x half> %f) { define <4 x i50> @test_unsigned_v4f16_v4i50(<4 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v4f16_v4i50: ; CHECK-CVT-LABEL: test_unsigned_v4f16_v4i50:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 def $q0 ; CHECK-CVT-NEXT: // kill: def $d0 killed $d0 def $q0
▲ Show 20 Lines • Show All 267 Lines • ▼ Show 20 Lines
declare <8 x i50> @llvm.fptoui.sat.v8f16.v8i50 (<8 x half>) declare <8 x i50> @llvm.fptoui.sat.v8f16.v8i50 (<8 x half>)
declare <8 x i64> @llvm.fptoui.sat.v8f16.v8i64 (<8 x half>) declare <8 x i64> @llvm.fptoui.sat.v8f16.v8i64 (<8 x half>)
declare <8 x i100> @llvm.fptoui.sat.v8f16.v8i100(<8 x half>) declare <8 x i100> @llvm.fptoui.sat.v8f16.v8i100(<8 x half>)
declare <8 x i128> @llvm.fptoui.sat.v8f16.v8i128(<8 x half>) declare <8 x i128> @llvm.fptoui.sat.v8f16.v8i128(<8 x half>)
define <8 x i1> @test_unsigned_v8f16_v8i1(<8 x half> %f) { define <8 x i1> @test_unsigned_v8f16_v8i1(<8 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v8f16_v8i1: ; CHECK-CVT-LABEL: test_unsigned_v8f16_v8i1:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: mov h1, v0.h[1] ; CHECK-CVT-NEXT: fcvtl2 v1.4s, v0.8h
; CHECK-CVT-NEXT: fcvt s2, h0 ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov h3, v0.h[2] ; CHECK-CVT-NEXT: mov s2, v1.s[1]
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-CVT-NEXT: mov s3, v1.s[2]
; CHECK-CVT-NEXT: fcvtzu w9, s2 ; CHECK-CVT-NEXT: mov s4, v1.s[3]
; CHECK-CVT-NEXT: fcvt s2, h3 ; CHECK-CVT-NEXT: mov s5, v0.s[1]
; CHECK-CVT-NEXT: fcvtzu w8, s1 ; CHECK-CVT-NEXT: fcvtzu w9, s1
; CHECK-CVT-NEXT: mov h1, v0.h[3] ; CHECK-CVT-NEXT: fcvtzu w10, s0
; CHECK-CVT-NEXT: fcvtzu w10, s2 ; CHECK-CVT-NEXT: mov s1, v0.s[2]
; CHECK-CVT-NEXT: mov h2, v0.h[4] ; CHECK-CVT-NEXT: mov s0, v0.s[3]
; CHECK-CVT-NEXT: fcvtzu w8, s2
; CHECK-CVT-NEXT: fcvtzu w11, s3
; CHECK-CVT-NEXT: fcvtzu w12, s4
; CHECK-CVT-NEXT: fcvtzu w13, s5
; CHECK-CVT-NEXT: cmp w8, #1 ; CHECK-CVT-NEXT: cmp w8, #1
; CHECK-CVT-NEXT: csinc w8, w8, wzr, lo ; CHECK-CVT-NEXT: csinc w8, w8, wzr, lo
; CHECK-CVT-NEXT: cmp w9, #1 ; CHECK-CVT-NEXT: cmp w9, #1
; CHECK-CVT-NEXT: csinc w9, w9, wzr, lo ; CHECK-CVT-NEXT: csinc w9, w9, wzr, lo
; CHECK-CVT-NEXT: fcvt s3, h1 ; CHECK-CVT-NEXT: cmp w11, #1
; CHECK-CVT-NEXT: csinc w11, w11, wzr, lo
; CHECK-CVT-NEXT: cmp w12, #1
; CHECK-CVT-NEXT: csinc w12, w12, wzr, lo
; CHECK-CVT-NEXT: cmp w13, #1
; CHECK-CVT-NEXT: csinc w13, w13, wzr, lo
; CHECK-CVT-NEXT: cmp w10, #1 ; CHECK-CVT-NEXT: cmp w10, #1
; CHECK-CVT-NEXT: fcvt s2, h2 ; CHECK-CVT-NEXT: csinc w10, w10, wzr, lo
; CHECK-CVT-NEXT: fmov s2, w10
; CHECK-CVT-NEXT: fcvtzu w10, s1
; CHECK-CVT-NEXT: fmov s1, w9 ; CHECK-CVT-NEXT: fmov s1, w9
; CHECK-CVT-NEXT: fcvtzu w9, s3 ; CHECK-CVT-NEXT: mov v2.s[1], w13
; CHECK-CVT-NEXT: mov h3, v0.h[5] ; CHECK-CVT-NEXT: cmp w10, #1
; CHECK-CVT-NEXT: mov v1.b[1], w8 ; CHECK-CVT-NEXT: csinc w9, w10, wzr, lo
; CHECK-CVT-NEXT: csinc w8, w10, wzr, lo ; CHECK-CVT-NEXT: fcvtzu w10, s0
; CHECK-CVT-NEXT: cmp w9, #1 ; CHECK-CVT-NEXT: mov v1.s[1], w8
; CHECK-CVT-NEXT: fcvt s3, h3 ; CHECK-CVT-NEXT: mov v2.s[2], w9
; CHECK-CVT-NEXT: csinc w9, w9, wzr, lo
; CHECK-CVT-NEXT: mov v1.b[2], w8
; CHECK-CVT-NEXT: fcvtzu w8, s2
; CHECK-CVT-NEXT: mov h2, v0.h[6]
; CHECK-CVT-NEXT: fcvtzu w10, s3
; CHECK-CVT-NEXT: mov h0, v0.h[7]
; CHECK-CVT-NEXT: cmp w8, #1
; CHECK-CVT-NEXT: mov v1.b[3], w9
; CHECK-CVT-NEXT: csinc w8, w8, wzr, lo
; CHECK-CVT-NEXT: fcvt s2, h2
; CHECK-CVT-NEXT: cmp w10, #1 ; CHECK-CVT-NEXT: cmp w10, #1
; CHECK-CVT-NEXT: fcvt s0, h0
; CHECK-CVT-NEXT: mov v1.b[4], w8
; CHECK-CVT-NEXT: csinc w8, w10, wzr, lo ; CHECK-CVT-NEXT: csinc w8, w10, wzr, lo
; CHECK-CVT-NEXT: fcvtzu w9, s2 ; CHECK-CVT-NEXT: mov v1.s[2], w11
; CHECK-CVT-NEXT: mov v1.b[5], w8 ; CHECK-CVT-NEXT: mov v2.s[3], w8
; CHECK-CVT-NEXT: cmp w9, #1 ; CHECK-CVT-NEXT: mov v1.s[3], w12
; CHECK-CVT-NEXT: csinc w8, w9, wzr, lo ; CHECK-CVT-NEXT: xtn v0.4h, v2.4s
; CHECK-CVT-NEXT: fcvtzu w9, s0 ; CHECK-CVT-NEXT: xtn2 v0.8h, v1.4s
; CHECK-CVT-NEXT: mov v1.b[6], w8 ; CHECK-CVT-NEXT: xtn v0.8b, v0.8h
; CHECK-CVT-NEXT: cmp w9, #1
; CHECK-CVT-NEXT: csinc w8, w9, wzr, lo
; CHECK-CVT-NEXT: mov v1.b[7], w8
; CHECK-CVT-NEXT: fmov d0, d1
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_unsigned_v8f16_v8i1: ; CHECK-FP16-LABEL: test_unsigned_v8f16_v8i1:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: mov h1, v0.h[1] ; CHECK-FP16-NEXT: movi v1.8h, #1
; CHECK-FP16-NEXT: fcvtzu w9, h0 ; CHECK-FP16-NEXT: fcvtzu v0.8h, v0.8h
; CHECK-FP16-NEXT: mov h2, v0.h[2] ; CHECK-FP16-NEXT: umin v0.8h, v0.8h, v1.8h
; CHECK-FP16-NEXT: mov h3, v0.h[3] ; CHECK-FP16-NEXT: xtn v0.8b, v0.8h
; CHECK-FP16-NEXT: fcvtzu w8, h1
; CHECK-FP16-NEXT: cmp w8, #1
; CHECK-FP16-NEXT: csinc w8, w8, wzr, lo
; CHECK-FP16-NEXT: cmp w9, #1
; CHECK-FP16-NEXT: csinc w9, w9, wzr, lo
; CHECK-FP16-NEXT: fmov s1, w9
; CHECK-FP16-NEXT: fcvtzu w9, h2
; CHECK-FP16-NEXT: mov h2, v0.h[4]
; CHECK-FP16-NEXT: mov v1.b[1], w8
; CHECK-FP16-NEXT: cmp w9, #1
; CHECK-FP16-NEXT: fcvtzu w8, h3
; CHECK-FP16-NEXT: csinc w9, w9, wzr, lo
; CHECK-FP16-NEXT: mov h3, v0.h[5]
; CHECK-FP16-NEXT: cmp w8, #1
; CHECK-FP16-NEXT: mov v1.b[2], w9
; CHECK-FP16-NEXT: csinc w8, w8, wzr, lo
; CHECK-FP16-NEXT: fcvtzu w9, h2
; CHECK-FP16-NEXT: mov h2, v0.h[6]
; CHECK-FP16-NEXT: mov h0, v0.h[7]
; CHECK-FP16-NEXT: mov v1.b[3], w8
; CHECK-FP16-NEXT: cmp w9, #1
; CHECK-FP16-NEXT: fcvtzu w8, h3
; CHECK-FP16-NEXT: csinc w9, w9, wzr, lo
; CHECK-FP16-NEXT: cmp w8, #1
; CHECK-FP16-NEXT: mov v1.b[4], w9
; CHECK-FP16-NEXT: csinc w8, w8, wzr, lo
; CHECK-FP16-NEXT: fcvtzu w9, h2
; CHECK-FP16-NEXT: mov v1.b[5], w8
; CHECK-FP16-NEXT: cmp w9, #1
; CHECK-FP16-NEXT: csinc w8, w9, wzr, lo
; CHECK-FP16-NEXT: fcvtzu w9, h0
; CHECK-FP16-NEXT: mov v1.b[6], w8
; CHECK-FP16-NEXT: cmp w9, #1
; CHECK-FP16-NEXT: csinc w8, w9, wzr, lo
; CHECK-FP16-NEXT: mov v1.b[7], w8
; CHECK-FP16-NEXT: fmov d0, d1
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <8 x i1> @llvm.fptoui.sat.v8f16.v8i1(<8 x half> %f) %x = call <8 x i1> @llvm.fptoui.sat.v8f16.v8i1(<8 x half> %f)
ret <8 x i1> %x ret <8 x i1> %x
} }
define <8 x i8> @test_unsigned_v8f16_v8i8(<8 x half> %f) { define <8 x i8> @test_unsigned_v8f16_v8i8(<8 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v8f16_v8i8: ; CHECK-CVT-LABEL: test_unsigned_v8f16_v8i8:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v1.4s, v0.4h ; CHECK-CVT-NEXT: fcvtl2 v1.4s, v0.8h
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov w8, #255 ; CHECK-CVT-NEXT: mov w8, #255
; CHECK-CVT-NEXT: fcvtl2 v0.4s, v0.8h
; CHECK-CVT-NEXT: mov s2, v1.s[1] ; CHECK-CVT-NEXT: mov s2, v1.s[1]
; CHECK-CVT-NEXT: mov s3, v1.s[2]
; CHECK-CVT-NEXT: mov s4, v1.s[3]
; CHECK-CVT-NEXT: mov s5, v0.s[1]
; CHECK-CVT-NEXT: fcvtzu w10, s1 ; CHECK-CVT-NEXT: fcvtzu w10, s1
; CHECK-CVT-NEXT: mov s3, v1.s[3] ; CHECK-CVT-NEXT: fcvtzu w11, s0
; CHECK-CVT-NEXT: mov s1, v0.s[2]
; CHECK-CVT-NEXT: mov s0, v0.s[3]
; CHECK-CVT-NEXT: fcvtzu w9, s2 ; CHECK-CVT-NEXT: fcvtzu w9, s2
; CHECK-CVT-NEXT: mov s2, v1.s[2] ; CHECK-CVT-NEXT: fcvtzu w12, s3
; CHECK-CVT-NEXT: fcvtzu w13, s4
; CHECK-CVT-NEXT: fcvtzu w14, s5
; CHECK-CVT-NEXT: cmp w9, #255 ; CHECK-CVT-NEXT: cmp w9, #255
; CHECK-CVT-NEXT: csel w9, w9, w8, lo ; CHECK-CVT-NEXT: csel w9, w9, w8, lo
; CHECK-CVT-NEXT: cmp w10, #255 ; CHECK-CVT-NEXT: cmp w10, #255
; CHECK-CVT-NEXT: csel w10, w10, w8, lo ; CHECK-CVT-NEXT: csel w10, w10, w8, lo
; CHECK-CVT-NEXT: cmp w12, #255
; CHECK-CVT-NEXT: csel w12, w12, w8, lo
; CHECK-CVT-NEXT: cmp w13, #255
; CHECK-CVT-NEXT: csel w13, w13, w8, lo
; CHECK-CVT-NEXT: cmp w14, #255
; CHECK-CVT-NEXT: csel w14, w14, w8, lo
; CHECK-CVT-NEXT: cmp w11, #255
; CHECK-CVT-NEXT: csel w11, w11, w8, lo
; CHECK-CVT-NEXT: fmov s2, w11
; CHECK-CVT-NEXT: fcvtzu w11, s1
; CHECK-CVT-NEXT: fmov s1, w10 ; CHECK-CVT-NEXT: fmov s1, w10
; CHECK-CVT-NEXT: fcvtzu w10, s2 ; CHECK-CVT-NEXT: mov v2.s[1], w14
; CHECK-CVT-NEXT: mov s2, v0.s[1] ; CHECK-CVT-NEXT: cmp w11, #255
; CHECK-CVT-NEXT: mov v1.b[1], w9 ; CHECK-CVT-NEXT: csel w10, w11, w8, lo
; CHECK-CVT-NEXT: cmp w10, #255 ; CHECK-CVT-NEXT: fcvtzu w11, s0
; CHECK-CVT-NEXT: fcvtzu w9, s3 ; CHECK-CVT-NEXT: mov v1.s[1], w9
; CHECK-CVT-NEXT: csel w10, w10, w8, lo ; CHECK-CVT-NEXT: mov v2.s[2], w10
; CHECK-CVT-NEXT: cmp w9, #255 ; CHECK-CVT-NEXT: cmp w11, #255
; CHECK-CVT-NEXT: mov v1.b[2], w10 ; CHECK-CVT-NEXT: csel w8, w11, w8, lo
; CHECK-CVT-NEXT: csel w9, w9, w8, lo ; CHECK-CVT-NEXT: mov v1.s[2], w12
; CHECK-CVT-NEXT: fcvtzu w10, s0 ; CHECK-CVT-NEXT: mov v2.s[3], w8
; CHECK-CVT-NEXT: mov v1.b[3], w9 ; CHECK-CVT-NEXT: mov v1.s[3], w13
; CHECK-CVT-NEXT: cmp w10, #255 ; CHECK-CVT-NEXT: xtn v0.4h, v2.4s
; CHECK-CVT-NEXT: fcvtzu w9, s2 ; CHECK-CVT-NEXT: xtn2 v0.8h, v1.4s
; CHECK-CVT-NEXT: csel w10, w10, w8, lo ; CHECK-CVT-NEXT: xtn v0.8b, v0.8h
; CHECK-CVT-NEXT: mov s2, v0.s[2]
; CHECK-CVT-NEXT: mov s0, v0.s[3]
; CHECK-CVT-NEXT: cmp w9, #255
; CHECK-CVT-NEXT: mov v1.b[4], w10
; CHECK-CVT-NEXT: csel w9, w9, w8, lo
; CHECK-CVT-NEXT: fcvtzu w10, s2
; CHECK-CVT-NEXT: mov v1.b[5], w9
; CHECK-CVT-NEXT: cmp w10, #255
; CHECK-CVT-NEXT: csel w9, w10, w8, lo
; CHECK-CVT-NEXT: fcvtzu w10, s0
; CHECK-CVT-NEXT: mov v1.b[6], w9
; CHECK-CVT-NEXT: cmp w10, #255
; CHECK-CVT-NEXT: csel w8, w10, w8, lo
; CHECK-CVT-NEXT: mov v1.b[7], w8
; CHECK-CVT-NEXT: fmov d0, d1
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_unsigned_v8f16_v8i8: ; CHECK-FP16-LABEL: test_unsigned_v8f16_v8i8:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: mov h1, v0.h[1] ; CHECK-FP16-NEXT: fcvtzu v0.8h, v0.8h
; CHECK-FP16-NEXT: fcvtzu w10, h0 ; CHECK-FP16-NEXT: uqxtn v0.8b, v0.8h
; CHECK-FP16-NEXT: mov w8, #255
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: mov h3, v0.h[3]
; CHECK-FP16-NEXT: fcvtzu w9, h1
; CHECK-FP16-NEXT: cmp w9, #255
; CHECK-FP16-NEXT: csel w9, w9, w8, lo
; CHECK-FP16-NEXT: cmp w10, #255
; CHECK-FP16-NEXT: csel w10, w10, w8, lo
; CHECK-FP16-NEXT: fmov s1, w10
; CHECK-FP16-NEXT: fcvtzu w10, h2
; CHECK-FP16-NEXT: mov h2, v0.h[4]
; CHECK-FP16-NEXT: mov v1.b[1], w9
; CHECK-FP16-NEXT: cmp w10, #255
; CHECK-FP16-NEXT: fcvtzu w9, h3
; CHECK-FP16-NEXT: csel w10, w10, w8, lo
; CHECK-FP16-NEXT: mov h3, v0.h[5]
; CHECK-FP16-NEXT: cmp w9, #255
; CHECK-FP16-NEXT: mov v1.b[2], w10
; CHECK-FP16-NEXT: csel w9, w9, w8, lo
; CHECK-FP16-NEXT: fcvtzu w10, h2
; CHECK-FP16-NEXT: mov h2, v0.h[6]
; CHECK-FP16-NEXT: mov h0, v0.h[7]
; CHECK-FP16-NEXT: mov v1.b[3], w9
; CHECK-FP16-NEXT: cmp w10, #255
; CHECK-FP16-NEXT: fcvtzu w9, h3
; CHECK-FP16-NEXT: csel w10, w10, w8, lo
; CHECK-FP16-NEXT: cmp w9, #255
; CHECK-FP16-NEXT: mov v1.b[4], w10
; CHECK-FP16-NEXT: csel w9, w9, w8, lo
; CHECK-FP16-NEXT: fcvtzu w10, h2
; CHECK-FP16-NEXT: mov v1.b[5], w9
; CHECK-FP16-NEXT: cmp w10, #255
; CHECK-FP16-NEXT: csel w9, w10, w8, lo
; CHECK-FP16-NEXT: fcvtzu w10, h0
; CHECK-FP16-NEXT: mov v1.b[6], w9
; CHECK-FP16-NEXT: cmp w10, #255
; CHECK-FP16-NEXT: csel w8, w10, w8, lo
; CHECK-FP16-NEXT: mov v1.b[7], w8
; CHECK-FP16-NEXT: fmov d0, d1
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <8 x i8> @llvm.fptoui.sat.v8f16.v8i8(<8 x half> %f) %x = call <8 x i8> @llvm.fptoui.sat.v8f16.v8i8(<8 x half> %f)
ret <8 x i8> %x ret <8 x i8> %x
} }
define <8 x i13> @test_unsigned_v8f16_v8i13(<8 x half> %f) { define <8 x i13> @test_unsigned_v8f16_v8i13(<8 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v8f16_v8i13: ; CHECK-CVT-LABEL: test_unsigned_v8f16_v8i13:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: mov h1, v0.h[1] ; CHECK-CVT-NEXT: fcvtl2 v1.4s, v0.8h
; CHECK-CVT-NEXT: fcvt s2, h0 ; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov h3, v0.h[2]
; CHECK-CVT-NEXT: mov w8, #8191 ; CHECK-CVT-NEXT: mov w8, #8191
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-CVT-NEXT: mov s2, v1.s[1]
; CHECK-CVT-NEXT: fcvtzu w10, s2 ; CHECK-CVT-NEXT: mov s3, v1.s[2]
; CHECK-CVT-NEXT: fcvt s2, h3 ; CHECK-CVT-NEXT: mov s4, v1.s[3]
; CHECK-CVT-NEXT: fcvtzu w9, s1 ; CHECK-CVT-NEXT: mov s5, v0.s[1]
; CHECK-CVT-NEXT: mov h1, v0.h[3] ; CHECK-CVT-NEXT: fcvtzu w10, s1
; CHECK-CVT-NEXT: fcvtzu w11, s2 ; CHECK-CVT-NEXT: fcvtzu w11, s0
; CHECK-CVT-NEXT: mov h2, v0.h[4] ; CHECK-CVT-NEXT: mov s1, v0.s[2]
; CHECK-CVT-NEXT: mov s0, v0.s[3]
; CHECK-CVT-NEXT: fcvtzu w9, s2
; CHECK-CVT-NEXT: fcvtzu w12, s3
; CHECK-CVT-NEXT: fcvtzu w13, s4
; CHECK-CVT-NEXT: fcvtzu w14, s5
; CHECK-CVT-NEXT: cmp w9, w8 ; CHECK-CVT-NEXT: cmp w9, w8
; CHECK-CVT-NEXT: csel w9, w9, w8, lo ; CHECK-CVT-NEXT: csel w9, w9, w8, lo
; CHECK-CVT-NEXT: cmp w10, w8 ; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w10, w10, w8, lo ; CHECK-CVT-NEXT: csel w10, w10, w8, lo
; CHECK-CVT-NEXT: fcvt s3, h1 ; CHECK-CVT-NEXT: cmp w12, w8
; CHECK-CVT-NEXT: csel w12, w12, w8, lo
; CHECK-CVT-NEXT: cmp w13, w8
; CHECK-CVT-NEXT: csel w13, w13, w8, lo
; CHECK-CVT-NEXT: cmp w14, w8
; CHECK-CVT-NEXT: csel w14, w14, w8, lo
; CHECK-CVT-NEXT: cmp w11, w8 ; CHECK-CVT-NEXT: cmp w11, w8
; CHECK-CVT-NEXT: fcvt s2, h2 ; CHECK-CVT-NEXT: csel w11, w11, w8, lo
; CHECK-CVT-NEXT: fmov s2, w11
; CHECK-CVT-NEXT: fcvtzu w11, s1
; CHECK-CVT-NEXT: fmov s1, w10 ; CHECK-CVT-NEXT: fmov s1, w10
; CHECK-CVT-NEXT: fcvtzu w10, s3 ; CHECK-CVT-NEXT: mov v2.s[1], w14
; CHECK-CVT-NEXT: mov h3, v0.h[5]
; CHECK-CVT-NEXT: mov v1.h[1], w9
; CHECK-CVT-NEXT: csel w9, w11, w8, lo
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: fcvt s3, h3
; CHECK-CVT-NEXT: csel w10, w10, w8, lo
; CHECK-CVT-NEXT: mov v1.h[2], w9
; CHECK-CVT-NEXT: fcvtzu w9, s2
; CHECK-CVT-NEXT: mov h2, v0.h[6]
; CHECK-CVT-NEXT: fcvtzu w11, s3
; CHECK-CVT-NEXT: mov h0, v0.h[7]
; CHECK-CVT-NEXT: cmp w9, w8
; CHECK-CVT-NEXT: mov v1.h[3], w10
; CHECK-CVT-NEXT: csel w9, w9, w8, lo
; CHECK-CVT-NEXT: fcvt s2, h2
; CHECK-CVT-NEXT: cmp w11, w8 ; CHECK-CVT-NEXT: cmp w11, w8
; CHECK-CVT-NEXT: fcvt s0, h0 ; CHECK-CVT-NEXT: csel w10, w11, w8, lo
; CHECK-CVT-NEXT: mov v1.h[4], w9 ; CHECK-CVT-NEXT: fcvtzu w11, s0
; CHECK-CVT-NEXT: csel w9, w11, w8, lo ; CHECK-CVT-NEXT: mov v1.s[1], w9
; CHECK-CVT-NEXT: fcvtzu w10, s2 ; CHECK-CVT-NEXT: mov v2.s[2], w10
; CHECK-CVT-NEXT: mov v1.h[5], w9 ; CHECK-CVT-NEXT: cmp w11, w8
; CHECK-CVT-NEXT: cmp w10, w8 ; CHECK-CVT-NEXT: csel w8, w11, w8, lo
; CHECK-CVT-NEXT: csel w9, w10, w8, lo ; CHECK-CVT-NEXT: mov v1.s[2], w12
; CHECK-CVT-NEXT: fcvtzu w10, s0 ; CHECK-CVT-NEXT: mov v2.s[3], w8
; CHECK-CVT-NEXT: mov v1.h[6], w9 ; CHECK-CVT-NEXT: mov v1.s[3], w13
; CHECK-CVT-NEXT: cmp w10, w8 ; CHECK-CVT-NEXT: xtn v0.4h, v2.4s
; CHECK-CVT-NEXT: csel w8, w10, w8, lo ; CHECK-CVT-NEXT: xtn2 v0.8h, v1.4s
; CHECK-CVT-NEXT: mov v1.h[7], w8
; CHECK-CVT-NEXT: mov v0.16b, v1.16b
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_unsigned_v8f16_v8i13: ; CHECK-FP16-LABEL: test_unsigned_v8f16_v8i13:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: mov h1, v0.h[1] ; CHECK-FP16-NEXT: mvni v1.8h, #224, lsl #8
; CHECK-FP16-NEXT: fcvtzu w10, h0 ; CHECK-FP16-NEXT: fcvtzu v0.8h, v0.8h
; CHECK-FP16-NEXT: mov w8, #8191 ; CHECK-FP16-NEXT: umin v0.8h, v0.8h, v1.8h
; CHECK-FP16-NEXT: mov h2, v0.h[2]
; CHECK-FP16-NEXT: mov h3, v0.h[3]
; CHECK-FP16-NEXT: fcvtzu w9, h1
; CHECK-FP16-NEXT: cmp w9, w8
; CHECK-FP16-NEXT: csel w9, w9, w8, lo
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w10, w10, w8, lo
; CHECK-FP16-NEXT: fmov s1, w10
; CHECK-FP16-NEXT: fcvtzu w10, h2
; CHECK-FP16-NEXT: mov h2, v0.h[4]
; CHECK-FP16-NEXT: mov v1.h[1], w9
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: fcvtzu w9, h3
; CHECK-FP16-NEXT: csel w10, w10, w8, lo
; CHECK-FP16-NEXT: mov h3, v0.h[5]
; CHECK-FP16-NEXT: cmp w9, w8
; CHECK-FP16-NEXT: mov v1.h[2], w10
; CHECK-FP16-NEXT: csel w9, w9, w8, lo
; CHECK-FP16-NEXT: fcvtzu w10, h2
; CHECK-FP16-NEXT: mov h2, v0.h[6]
; CHECK-FP16-NEXT: mov h0, v0.h[7]
; CHECK-FP16-NEXT: mov v1.h[3], w9
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: fcvtzu w9, h3
; CHECK-FP16-NEXT: csel w10, w10, w8, lo
; CHECK-FP16-NEXT: cmp w9, w8
; CHECK-FP16-NEXT: mov v1.h[4], w10
; CHECK-FP16-NEXT: csel w9, w9, w8, lo
; CHECK-FP16-NEXT: fcvtzu w10, h2
; CHECK-FP16-NEXT: mov v1.h[5], w9
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w9, w10, w8, lo
; CHECK-FP16-NEXT: fcvtzu w10, h0
; CHECK-FP16-NEXT: mov v1.h[6], w9
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: csel w8, w10, w8, lo
; CHECK-FP16-NEXT: mov v1.h[7], w8
; CHECK-FP16-NEXT: mov v0.16b, v1.16b
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <8 x i13> @llvm.fptoui.sat.v8f16.v8i13(<8 x half> %f) %x = call <8 x i13> @llvm.fptoui.sat.v8f16.v8i13(<8 x half> %f)
ret <8 x i13> %x ret <8 x i13> %x
} }
define <8 x i16> @test_unsigned_v8f16_v8i16(<8 x half> %f) { define <8 x i16> @test_unsigned_v8f16_v8i16(<8 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v8f16_v8i16: ; CHECK-CVT-LABEL: test_unsigned_v8f16_v8i16:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: fcvtl v1.4s, v0.4h ; CHECK-CVT-NEXT: fcvtl2 v1.4s, v0.8h
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: mov w8, #65535 ; CHECK-CVT-NEXT: mov w8, #65535
; CHECK-CVT-NEXT: fcvtl2 v0.4s, v0.8h
; CHECK-CVT-NEXT: mov s2, v1.s[1] ; CHECK-CVT-NEXT: mov s2, v1.s[1]
; CHECK-CVT-NEXT: mov s3, v1.s[2]
; CHECK-CVT-NEXT: mov s4, v1.s[3]
; CHECK-CVT-NEXT: mov s5, v0.s[1]
; CHECK-CVT-NEXT: fcvtzu w10, s1 ; CHECK-CVT-NEXT: fcvtzu w10, s1
; CHECK-CVT-NEXT: mov s3, v1.s[3] ; CHECK-CVT-NEXT: fcvtzu w11, s0
; CHECK-CVT-NEXT: mov s1, v0.s[2]
; CHECK-CVT-NEXT: mov s0, v0.s[3]
; CHECK-CVT-NEXT: fcvtzu w9, s2 ; CHECK-CVT-NEXT: fcvtzu w9, s2
; CHECK-CVT-NEXT: mov s2, v1.s[2] ; CHECK-CVT-NEXT: fcvtzu w12, s3
; CHECK-CVT-NEXT: fcvtzu w13, s4
; CHECK-CVT-NEXT: fcvtzu w14, s5
; CHECK-CVT-NEXT: cmp w9, w8 ; CHECK-CVT-NEXT: cmp w9, w8
; CHECK-CVT-NEXT: csel w9, w9, w8, lo ; CHECK-CVT-NEXT: csel w9, w9, w8, lo
; CHECK-CVT-NEXT: cmp w10, w8 ; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w10, w10, w8, lo ; CHECK-CVT-NEXT: csel w10, w10, w8, lo
; CHECK-CVT-NEXT: cmp w12, w8
; CHECK-CVT-NEXT: csel w12, w12, w8, lo
; CHECK-CVT-NEXT: cmp w13, w8
; CHECK-CVT-NEXT: csel w13, w13, w8, lo
; CHECK-CVT-NEXT: cmp w14, w8
; CHECK-CVT-NEXT: csel w14, w14, w8, lo
; CHECK-CVT-NEXT: cmp w11, w8
; CHECK-CVT-NEXT: csel w11, w11, w8, lo
; CHECK-CVT-NEXT: fmov s2, w11
; CHECK-CVT-NEXT: fcvtzu w11, s1
; CHECK-CVT-NEXT: fmov s1, w10 ; CHECK-CVT-NEXT: fmov s1, w10
; CHECK-CVT-NEXT: fcvtzu w10, s2 ; CHECK-CVT-NEXT: mov v2.s[1], w14
; CHECK-CVT-NEXT: mov s2, v0.s[1] ; CHECK-CVT-NEXT: cmp w11, w8
; CHECK-CVT-NEXT: mov v1.h[1], w9 ; CHECK-CVT-NEXT: csel w10, w11, w8, lo
; CHECK-CVT-NEXT: cmp w10, w8 ; CHECK-CVT-NEXT: fcvtzu w11, s0
; CHECK-CVT-NEXT: fcvtzu w9, s3 ; CHECK-CVT-NEXT: mov v1.s[1], w9
; CHECK-CVT-NEXT: csel w10, w10, w8, lo ; CHECK-CVT-NEXT: mov v2.s[2], w10
; CHECK-CVT-NEXT: cmp w9, w8 ; CHECK-CVT-NEXT: cmp w11, w8
; CHECK-CVT-NEXT: mov v1.h[2], w10 ; CHECK-CVT-NEXT: csel w8, w11, w8, lo
; CHECK-CVT-NEXT: csel w9, w9, w8, lo ; CHECK-CVT-NEXT: mov v1.s[2], w12
; CHECK-CVT-NEXT: fcvtzu w10, s0 ; CHECK-CVT-NEXT: mov v2.s[3], w8
; CHECK-CVT-NEXT: mov v1.h[3], w9 ; CHECK-CVT-NEXT: mov v1.s[3], w13
; CHECK-CVT-NEXT: cmp w10, w8 ; CHECK-CVT-NEXT: xtn v0.4h, v2.4s
; CHECK-CVT-NEXT: fcvtzu w9, s2 ; CHECK-CVT-NEXT: xtn2 v0.8h, v1.4s
; CHECK-CVT-NEXT: csel w10, w10, w8, lo
; CHECK-CVT-NEXT: mov s2, v0.s[2]
; CHECK-CVT-NEXT: mov s0, v0.s[3]
; CHECK-CVT-NEXT: cmp w9, w8
; CHECK-CVT-NEXT: mov v1.h[4], w10
; CHECK-CVT-NEXT: csel w9, w9, w8, lo
; CHECK-CVT-NEXT: fcvtzu w10, s2
; CHECK-CVT-NEXT: mov v1.h[5], w9
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w9, w10, w8, lo
; CHECK-CVT-NEXT: fcvtzu w10, s0
; CHECK-CVT-NEXT: mov v1.h[6], w9
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: csel w8, w10, w8, lo
; CHECK-CVT-NEXT: mov v1.h[7], w8
; CHECK-CVT-NEXT: mov v0.16b, v1.16b
; CHECK-CVT-NEXT: ret ; CHECK-CVT-NEXT: ret
; ;
; CHECK-FP16-LABEL: test_unsigned_v8f16_v8i16: ; CHECK-FP16-LABEL: test_unsigned_v8f16_v8i16:
; CHECK-FP16: // %bb.0: ; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: fcvtzu v0.8h, v0.8h ; CHECK-FP16-NEXT: fcvtzu v0.8h, v0.8h
; CHECK-FP16-NEXT: ret ; CHECK-FP16-NEXT: ret
%x = call <8 x i16> @llvm.fptoui.sat.v8f16.v8i16(<8 x half> %f) %x = call <8 x i16> @llvm.fptoui.sat.v8f16.v8i16(<8 x half> %f)
ret <8 x i16> %x ret <8 x i16> %x
} }
define <8 x i19> @test_unsigned_v8f16_v8i19(<8 x half> %f) { define <8 x i19> @test_unsigned_v8f16_v8i19(<8 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v8f16_v8i19: ; CHECK-LABEL: test_unsigned_v8f16_v8i19:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: ext v1.16b, v0.16b, v0.16b, #8 ; CHECK-NEXT: fcvtl v2.4s, v0.4h
; CHECK-CVT-NEXT: mov w8, #524287 ; CHECK-NEXT: fcvtl2 v0.4s, v0.8h
; CHECK-CVT-NEXT: mov h2, v0.h[1] ; CHECK-NEXT: movi v1.4s, #7, msl #16
; CHECK-CVT-NEXT: mov h3, v0.h[2] ; CHECK-NEXT: fcvtzu v2.4s, v2.4s
; CHECK-CVT-NEXT: mov h5, v0.h[3] ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: fcvt s0, h0 ; CHECK-NEXT: umin v2.4s, v2.4s, v1.4s
; CHECK-CVT-NEXT: mov h4, v1.h[1] ; CHECK-NEXT: umin v0.4s, v0.4s, v1.4s
; CHECK-CVT-NEXT: mov h6, v1.h[2] ; CHECK-NEXT: mov w1, v2.s[1]
; CHECK-CVT-NEXT: mov h7, v1.h[3] ; CHECK-NEXT: mov w2, v2.s[2]
; CHECK-CVT-NEXT: fcvt s1, h1 ; CHECK-NEXT: mov w5, v0.s[1]
; CHECK-CVT-NEXT: fcvt s2, h2 ; CHECK-NEXT: mov w3, v2.s[3]
; CHECK-CVT-NEXT: fcvt s3, h3 ; CHECK-NEXT: mov w6, v0.s[2]
; CHECK-CVT-NEXT: fcvtzu w9, s0 ; CHECK-NEXT: mov w7, v0.s[3]
; CHECK-CVT-NEXT: fcvt s5, h5 ; CHECK-NEXT: fmov w4, s0
; CHECK-CVT-NEXT: fcvt s4, h4 ; CHECK-NEXT: fmov w0, s2
; CHECK-CVT-NEXT: fcvt s6, h6 ; CHECK-NEXT: ret
; CHECK-CVT-NEXT: fcvt s0, h7
; CHECK-CVT-NEXT: fcvtzu w10, s1
; CHECK-CVT-NEXT: fcvtzu w11, s2
; CHECK-CVT-NEXT: fcvtzu w12, s3
; CHECK-CVT-NEXT: fcvtzu w14, s5
; CHECK-CVT-NEXT: fcvtzu w13, s4
; CHECK-CVT-NEXT: fcvtzu w15, s6
; CHECK-CVT-NEXT: cmp w10, w8
; CHECK-CVT-NEXT: fcvtzu w16, s0
; CHECK-CVT-NEXT: csel w4, w10, w8, lo
; CHECK-CVT-NEXT: cmp w13, w8
; CHECK-CVT-NEXT: csel w5, w13, w8, lo
; CHECK-CVT-NEXT: cmp w15, w8
; CHECK-CVT-NEXT: csel w6, w15, w8, lo
; CHECK-CVT-NEXT: cmp w16, w8
; CHECK-CVT-NEXT: csel w7, w16, w8, lo
; CHECK-CVT-NEXT: cmp w9, w8
; CHECK-CVT-NEXT: csel w0, w9, w8, lo
; CHECK-CVT-NEXT: cmp w11, w8
; CHECK-CVT-NEXT: csel w1, w11, w8, lo
; CHECK-CVT-NEXT: cmp w12, w8
; CHECK-CVT-NEXT: csel w2, w12, w8, lo
; CHECK-CVT-NEXT: cmp w14, w8
; CHECK-CVT-NEXT: csel w3, w14, w8, lo
; CHECK-CVT-NEXT: ret
;
; CHECK-FP16-LABEL: test_unsigned_v8f16_v8i19:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: ext v1.16b, v0.16b, v0.16b, #8
; CHECK-FP16-NEXT: mov w8, #524287
; CHECK-FP16-NEXT: mov h2, v0.h[1]
; CHECK-FP16-NEXT: mov h3, v0.h[2]
; CHECK-FP16-NEXT: mov h5, v0.h[3]
; CHECK-FP16-NEXT: fcvtzu w9, h0
; CHECK-FP16-NEXT: mov h4, v1.h[1]
; CHECK-FP16-NEXT: mov h6, v1.h[2]
; CHECK-FP16-NEXT: mov h0, v1.h[3]
; CHECK-FP16-NEXT: fcvtzu w10, h1
; CHECK-FP16-NEXT: fcvtzu w11, h2
; CHECK-FP16-NEXT: fcvtzu w12, h3
; CHECK-FP16-NEXT: fcvtzu w14, h5
; CHECK-FP16-NEXT: fcvtzu w13, h4
; CHECK-FP16-NEXT: fcvtzu w15, h6
; CHECK-FP16-NEXT: cmp w10, w8
; CHECK-FP16-NEXT: fcvtzu w16, h0
; CHECK-FP16-NEXT: csel w4, w10, w8, lo
; CHECK-FP16-NEXT: cmp w13, w8
; CHECK-FP16-NEXT: csel w5, w13, w8, lo
; CHECK-FP16-NEXT: cmp w15, w8
; CHECK-FP16-NEXT: csel w6, w15, w8, lo
; CHECK-FP16-NEXT: cmp w16, w8
; CHECK-FP16-NEXT: csel w7, w16, w8, lo
; CHECK-FP16-NEXT: cmp w9, w8
; CHECK-FP16-NEXT: csel w0, w9, w8, lo
; CHECK-FP16-NEXT: cmp w11, w8
; CHECK-FP16-NEXT: csel w1, w11, w8, lo
; CHECK-FP16-NEXT: cmp w12, w8
; CHECK-FP16-NEXT: csel w2, w12, w8, lo
; CHECK-FP16-NEXT: cmp w14, w8
; CHECK-FP16-NEXT: csel w3, w14, w8, lo
; CHECK-FP16-NEXT: ret
%x = call <8 x i19> @llvm.fptoui.sat.v8f16.v8i19(<8 x half> %f) %x = call <8 x i19> @llvm.fptoui.sat.v8f16.v8i19(<8 x half> %f)
ret <8 x i19> %x ret <8 x i19> %x
} }
define <8 x i32> @test_unsigned_v8f16_v8i32_duplicate(<8 x half> %f) { define <8 x i32> @test_unsigned_v8f16_v8i32_duplicate(<8 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v8f16_v8i32_duplicate: ; CHECK-LABEL: test_unsigned_v8f16_v8i32_duplicate:
; CHECK-CVT: // %bb.0: ; CHECK: // %bb.0:
; CHECK-CVT-NEXT: fcvtl2 v1.4s, v0.8h ; CHECK-NEXT: fcvtl2 v1.4s, v0.8h
; CHECK-CVT-NEXT: fcvtl v0.4s, v0.4h ; CHECK-NEXT: fcvtl v0.4s, v0.4h
; CHECK-CVT-NEXT: fcvtzu v1.4s, v1.4s ; CHECK-NEXT: fcvtzu v1.4s, v1.4s
; CHECK-CVT-NEXT: fcvtzu v0.4s, v0.4s ; CHECK-NEXT: fcvtzu v0.4s, v0.4s
; CHECK-CVT-NEXT: ret ; CHECK-NEXT: ret
;
; CHECK-FP16-LABEL: test_unsigned_v8f16_v8i32_duplicate:
; CHECK-FP16: // %bb.0:
; CHECK-FP16-NEXT: ext v3.16b, v0.16b, v0.16b, #8
; CHECK-FP16-NEXT: mov h4, v0.h[1]
; CHECK-FP16-NEXT: fcvtzu w9, h0
; CHECK-FP16-NEXT: mov h2, v3.h[1]
; CHECK-FP16-NEXT: fcvtzu w8, h3
; CHECK-FP16-NEXT: mov h5, v3.h[2]
; CHECK-FP16-NEXT: mov h3, v3.h[3]
; CHECK-FP16-NEXT: fmov s1, w8
; CHECK-FP16-NEXT: fcvtzu w8, h2
; CHECK-FP16-NEXT: fmov s2, w9
; CHECK-FP16-NEXT: fcvtzu w9, h4
; CHECK-FP16-NEXT: mov h4, v0.h[2]
; CHECK-FP16-NEXT: mov h0, v0.h[3]
; CHECK-FP16-NEXT: mov v1.s[1], w8
; CHECK-FP16-NEXT: fcvtzu w8, h5
; CHECK-FP16-NEXT: mov v2.s[1], w9
; CHECK-FP16-NEXT: fcvtzu w9, h4
; CHECK-FP16-NEXT: mov v1.s[2], w8
; CHECK-FP16-NEXT: fcvtzu w8, h3
; CHECK-FP16-NEXT: mov v2.s[2], w9
; CHECK-FP16-NEXT: fcvtzu w9, h0
; CHECK-FP16-NEXT: mov v1.s[3], w8
; CHECK-FP16-NEXT: mov v2.s[3], w9
; CHECK-FP16-NEXT: mov v0.16b, v2.16b
; CHECK-FP16-NEXT: ret
%x = call <8 x i32> @llvm.fptoui.sat.v8f16.v8i32(<8 x half> %f) %x = call <8 x i32> @llvm.fptoui.sat.v8f16.v8i32(<8 x half> %f)
ret <8 x i32> %x ret <8 x i32> %x
} }
define <8 x i50> @test_unsigned_v8f16_v8i50(<8 x half> %f) { define <8 x i50> @test_unsigned_v8f16_v8i50(<8 x half> %f) {
; CHECK-CVT-LABEL: test_unsigned_v8f16_v8i50: ; CHECK-CVT-LABEL: test_unsigned_v8f16_v8i50:
; CHECK-CVT: // %bb.0: ; CHECK-CVT: // %bb.0:
; CHECK-CVT-NEXT: ext v1.16b, v0.16b, v0.16b, #8 ; CHECK-CVT-NEXT: ext v1.16b, v0.16b, v0.16b, #8
▲ Show 20 Lines • Show All 468 Lines • Show Last 20 Lines