-; RUN: llc < %s -march=arm -mattr=+neon | FileCheck %s
+; RUN: llc < %s -march=arm -mcpu=cortex-a8 | FileCheck %s
define <8 x i8> @vmuli8(<8 x i8>* %A, <8 x i8>* %B) nounwind {
;CHECK: vmuli8:
ret <8 x i16> %tmp5
}
+define <8 x i16> @vmulls8_int(<8 x i8>* %A, <8 x i8>* %B) nounwind {
+;CHECK: vmulls8_int:
+;CHECK: vmull.s8
+ %tmp1 = load <8 x i8>* %A
+ %tmp2 = load <8 x i8>* %B
+ %tmp3 = call <8 x i16> @llvm.arm.neon.vmulls.v8i16(<8 x i8> %tmp1, <8 x i8> %tmp2)
+ ret <8 x i16> %tmp3
+}
+
define <4 x i32> @vmulls16(<4 x i16>* %A, <4 x i16>* %B) nounwind {
;CHECK: vmulls16:
;CHECK: vmull.s16
ret <4 x i32> %tmp5
}
+define <4 x i32> @vmulls16_int(<4 x i16>* %A, <4 x i16>* %B) nounwind {
+;CHECK: vmulls16_int:
+;CHECK: vmull.s16
+ %tmp1 = load <4 x i16>* %A
+ %tmp2 = load <4 x i16>* %B
+ %tmp3 = call <4 x i32> @llvm.arm.neon.vmulls.v4i32(<4 x i16> %tmp1, <4 x i16> %tmp2)
+ ret <4 x i32> %tmp3
+}
+
define <2 x i64> @vmulls32(<2 x i32>* %A, <2 x i32>* %B) nounwind {
;CHECK: vmulls32:
;CHECK: vmull.s32
ret <2 x i64> %tmp5
}
+define <2 x i64> @vmulls32_int(<2 x i32>* %A, <2 x i32>* %B) nounwind {
+;CHECK: vmulls32_int:
+;CHECK: vmull.s32
+ %tmp1 = load <2 x i32>* %A
+ %tmp2 = load <2 x i32>* %B
+ %tmp3 = call <2 x i64> @llvm.arm.neon.vmulls.v2i64(<2 x i32> %tmp1, <2 x i32> %tmp2)
+ ret <2 x i64> %tmp3
+}
+
define <8 x i16> @vmullu8(<8 x i8>* %A, <8 x i8>* %B) nounwind {
;CHECK: vmullu8:
;CHECK: vmull.u8
ret <8 x i16> %tmp5
}
+define <8 x i16> @vmullu8_int(<8 x i8>* %A, <8 x i8>* %B) nounwind {
+;CHECK: vmullu8_int:
+;CHECK: vmull.u8
+ %tmp1 = load <8 x i8>* %A
+ %tmp2 = load <8 x i8>* %B
+ %tmp3 = call <8 x i16> @llvm.arm.neon.vmullu.v8i16(<8 x i8> %tmp1, <8 x i8> %tmp2)
+ ret <8 x i16> %tmp3
+}
+
define <4 x i32> @vmullu16(<4 x i16>* %A, <4 x i16>* %B) nounwind {
;CHECK: vmullu16:
;CHECK: vmull.u16
ret <4 x i32> %tmp5
}
+define <4 x i32> @vmullu16_int(<4 x i16>* %A, <4 x i16>* %B) nounwind {
+;CHECK: vmullu16_int:
+;CHECK: vmull.u16
+ %tmp1 = load <4 x i16>* %A
+ %tmp2 = load <4 x i16>* %B
+ %tmp3 = call <4 x i32> @llvm.arm.neon.vmullu.v4i32(<4 x i16> %tmp1, <4 x i16> %tmp2)
+ ret <4 x i32> %tmp3
+}
+
define <2 x i64> @vmullu32(<2 x i32>* %A, <2 x i32>* %B) nounwind {
;CHECK: vmullu32:
;CHECK: vmull.u32
ret <2 x i64> %tmp5
}
+define <2 x i64> @vmullu32_int(<2 x i32>* %A, <2 x i32>* %B) nounwind {
+;CHECK: vmullu32_int:
+;CHECK: vmull.u32
+ %tmp1 = load <2 x i32>* %A
+ %tmp2 = load <2 x i32>* %B
+ %tmp3 = call <2 x i64> @llvm.arm.neon.vmullu.v2i64(<2 x i32> %tmp1, <2 x i32> %tmp2)
+ ret <2 x i64> %tmp3
+}
+
define <8 x i16> @vmullp8(<8 x i8>* %A, <8 x i8>* %B) nounwind {
;CHECK: vmullp8:
;CHECK: vmull.p8
ret <4 x i32> %3
}
+define arm_aapcs_vfpcc <4 x i32> @test_vmull_lanes16_int(<4 x i16> %arg0_int16x4_t, <4 x i16> %arg1_int16x4_t) nounwind readnone {
+entry:
+; CHECK: test_vmull_lanes16_int
+; CHECK: vmull.s16 q0, d0, d1[1]
+ %0 = shufflevector <4 x i16> %arg1_int16x4_t, <4 x i16> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1> ; <<4 x i16>> [#uses=1]
+ %1 = tail call <4 x i32> @llvm.arm.neon.vmulls.v4i32(<4 x i16> %arg0_int16x4_t, <4 x i16> %0) ; <<4 x i32>> [#uses=1]
+ ret <4 x i32> %1
+}
+
define arm_aapcs_vfpcc <2 x i64> @test_vmull_lanes32(<2 x i32> %arg0_int32x2_t, <2 x i32> %arg1_int32x2_t) nounwind readnone {
entry:
; CHECK: test_vmull_lanes32
ret <2 x i64> %3
}
+define arm_aapcs_vfpcc <2 x i64> @test_vmull_lanes32_int(<2 x i32> %arg0_int32x2_t, <2 x i32> %arg1_int32x2_t) nounwind readnone {
+entry:
+; CHECK: test_vmull_lanes32_int
+; CHECK: vmull.s32 q0, d0, d1[1]
+ %0 = shufflevector <2 x i32> %arg1_int32x2_t, <2 x i32> undef, <2 x i32> <i32 1, i32 1> ; <<2 x i32>> [#uses=1]
+ %1 = tail call <2 x i64> @llvm.arm.neon.vmulls.v2i64(<2 x i32> %arg0_int32x2_t, <2 x i32> %0) ; <<2 x i64>> [#uses=1]
+ ret <2 x i64> %1
+}
+
define arm_aapcs_vfpcc <4 x i32> @test_vmull_laneu16(<4 x i16> %arg0_uint16x4_t, <4 x i16> %arg1_uint16x4_t) nounwind readnone {
entry:
; CHECK: test_vmull_laneu16
ret <4 x i32> %3
}
+define arm_aapcs_vfpcc <4 x i32> @test_vmull_laneu16_int(<4 x i16> %arg0_uint16x4_t, <4 x i16> %arg1_uint16x4_t) nounwind readnone {
+entry:
+; CHECK: test_vmull_laneu16_int
+; CHECK: vmull.u16 q0, d0, d1[1]
+ %0 = shufflevector <4 x i16> %arg1_uint16x4_t, <4 x i16> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1> ; <<4 x i16>> [#uses=1]
+ %1 = tail call <4 x i32> @llvm.arm.neon.vmullu.v4i32(<4 x i16> %arg0_uint16x4_t, <4 x i16> %0) ; <<4 x i32>> [#uses=1]
+ ret <4 x i32> %1
+}
+
define arm_aapcs_vfpcc <2 x i64> @test_vmull_laneu32(<2 x i32> %arg0_uint32x2_t, <2 x i32> %arg1_uint32x2_t) nounwind readnone {
entry:
; CHECK: test_vmull_laneu32
ret <2 x i64> %3
}
+define arm_aapcs_vfpcc <2 x i64> @test_vmull_laneu32_int(<2 x i32> %arg0_uint32x2_t, <2 x i32> %arg1_uint32x2_t) nounwind readnone {
+entry:
+; CHECK: test_vmull_laneu32_int
+; CHECK: vmull.u32 q0, d0, d1[1]
+ %0 = shufflevector <2 x i32> %arg1_uint32x2_t, <2 x i32> undef, <2 x i32> <i32 1, i32 1> ; <<2 x i32>> [#uses=1]
+ %1 = tail call <2 x i64> @llvm.arm.neon.vmullu.v2i64(<2 x i32> %arg0_uint32x2_t, <2 x i32> %0) ; <<2 x i64>> [#uses=1]
+ ret <2 x i64> %1
+}
+
+declare <8 x i16> @llvm.arm.neon.vmulls.v8i16(<8 x i8>, <8 x i8>) nounwind readnone
+declare <4 x i32> @llvm.arm.neon.vmulls.v4i32(<4 x i16>, <4 x i16>) nounwind readnone
+declare <2 x i64> @llvm.arm.neon.vmulls.v2i64(<2 x i32>, <2 x i32>) nounwind readnone
+
+declare <8 x i16> @llvm.arm.neon.vmullu.v8i16(<8 x i8>, <8 x i8>) nounwind readnone
+declare <4 x i32> @llvm.arm.neon.vmullu.v4i32(<4 x i16>, <4 x i16>) nounwind readnone
+declare <2 x i64> @llvm.arm.neon.vmullu.v2i64(<2 x i32>, <2 x i32>) nounwind readnone
+
declare <8 x i16> @llvm.arm.neon.vmullp.v8i16(<8 x i8>, <8 x i8>) nounwind readnone
%tmp4 = mul <2 x i64> %tmp3, <i64 1234, i64 1234>
ret <2 x i64> %tmp4
}
+
+; rdar://9197392
+define void @distribute(i16* %dst, i8* %src, i32 %mul) nounwind {
+entry:
+; CHECK: distribute:
+; CHECK: vmull.u8 [[REG1:(q[0-9]+)]], d{{.*}}, [[REG2:(d[0-9]+)]]
+; CHECK: vmlal.u8 [[REG1]], d{{.*}}, [[REG2]]
+ %0 = trunc i32 %mul to i8
+ %1 = insertelement <8 x i8> undef, i8 %0, i32 0
+ %2 = shufflevector <8 x i8> %1, <8 x i8> undef, <8 x i32> zeroinitializer
+ %3 = tail call <16 x i8> @llvm.arm.neon.vld1.v16i8(i8* %src, i32 1)
+ %4 = bitcast <16 x i8> %3 to <2 x double>
+ %5 = extractelement <2 x double> %4, i32 1
+ %6 = bitcast double %5 to <8 x i8>
+ %7 = zext <8 x i8> %6 to <8 x i16>
+ %8 = zext <8 x i8> %2 to <8 x i16>
+ %9 = extractelement <2 x double> %4, i32 0
+ %10 = bitcast double %9 to <8 x i8>
+ %11 = zext <8 x i8> %10 to <8 x i16>
+ %12 = add <8 x i16> %7, %11
+ %13 = mul <8 x i16> %12, %8
+ %14 = bitcast i16* %dst to i8*
+ tail call void @llvm.arm.neon.vst1.v8i16(i8* %14, <8 x i16> %13, i32 2)
+ ret void
+}
+
+declare <16 x i8> @llvm.arm.neon.vld1.v16i8(i8*, i32) nounwind readonly
+
+declare void @llvm.arm.neon.vst1.v8i16(i8*, <8 x i16>, i32) nounwind
+
+; Take advantage of the Cortex-A8 multiplier accumulator forward.
+
+%struct.uint8x8_t = type { <8 x i8> }
+
+define void @distribute2(%struct.uint8x8_t* nocapture %dst, i8* %src, i32 %mul) nounwind {
+entry:
+; CHECK: distribute2
+; CHECK-NOT: vadd.i8
+; CHECK: vmul.i8
+; CHECK: vmla.i8
+ %0 = trunc i32 %mul to i8
+ %1 = insertelement <8 x i8> undef, i8 %0, i32 0
+ %2 = shufflevector <8 x i8> %1, <8 x i8> undef, <8 x i32> zeroinitializer
+ %3 = tail call <16 x i8> @llvm.arm.neon.vld1.v16i8(i8* %src, i32 1)
+ %4 = bitcast <16 x i8> %3 to <2 x double>
+ %5 = extractelement <2 x double> %4, i32 1
+ %6 = bitcast double %5 to <8 x i8>
+ %7 = extractelement <2 x double> %4, i32 0
+ %8 = bitcast double %7 to <8 x i8>
+ %9 = add <8 x i8> %6, %8
+ %10 = mul <8 x i8> %9, %2
+ %11 = getelementptr inbounds %struct.uint8x8_t* %dst, i32 0, i32 0
+ store <8 x i8> %10, <8 x i8>* %11, align 8
+ ret void
+}
+
+define void @distribute2_commutative(%struct.uint8x8_t* nocapture %dst, i8* %src, i32 %mul) nounwind {
+entry:
+; CHECK: distribute2_commutative
+; CHECK-NOT: vadd.i8
+; CHECK: vmul.i8
+; CHECK: vmla.i8
+ %0 = trunc i32 %mul to i8
+ %1 = insertelement <8 x i8> undef, i8 %0, i32 0
+ %2 = shufflevector <8 x i8> %1, <8 x i8> undef, <8 x i32> zeroinitializer
+ %3 = tail call <16 x i8> @llvm.arm.neon.vld1.v16i8(i8* %src, i32 1)
+ %4 = bitcast <16 x i8> %3 to <2 x double>
+ %5 = extractelement <2 x double> %4, i32 1
+ %6 = bitcast double %5 to <8 x i8>
+ %7 = extractelement <2 x double> %4, i32 0
+ %8 = bitcast double %7 to <8 x i8>
+ %9 = add <8 x i8> %6, %8
+ %10 = mul <8 x i8> %2, %9
+ %11 = getelementptr inbounds %struct.uint8x8_t* %dst, i32 0, i32 0
+ store <8 x i8> %10, <8 x i8>* %11, align 8
+ ret void
+}
+
+; If one operand has a zero-extend and the other a sign-extend, vmull
+; cannot be used.
+define i16 @vmullWithInconsistentExtensions(<8 x i8> %vec) {
+; CHECK: vmullWithInconsistentExtensions
+; CHECK-NOT: vmull.s8
+ %1 = sext <8 x i8> %vec to <8 x i16>
+ %2 = mul <8 x i16> %1, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
+ %3 = extractelement <8 x i16> %2, i32 0
+ ret i16 %3
+}
+
+; A constant build_vector created for a vmull with half-width elements must
+; not introduce illegal types. <rdar://problem/11324364>
+define void @vmull_buildvector() nounwind optsize ssp align 2 {
+; CHECK: vmull_buildvector
+entry:
+ br i1 undef, label %for.end179, label %for.body.lr.ph
+
+for.body.lr.ph: ; preds = %entry
+ br label %for.body
+
+for.cond.loopexit: ; preds = %for.body33, %for.body
+ br i1 undef, label %for.end179, label %for.body
+
+for.body: ; preds = %for.cond.loopexit, %for.body.lr.ph
+ br i1 undef, label %for.cond.loopexit, label %for.body33.lr.ph
+
+for.body33.lr.ph: ; preds = %for.body
+ %.sub = select i1 undef, i32 0, i32 undef
+ br label %for.body33
+
+for.body33: ; preds = %for.body33, %for.body33.lr.ph
+ %add45 = add i32 undef, undef
+ %vld155 = tail call <16 x i8> @llvm.arm.neon.vld1.v16i8(i8* undef, i32 1)
+ %0 = load i32** undef, align 4
+ %shuffle.i250 = shufflevector <2 x i64> undef, <2 x i64> undef, <1 x i32> zeroinitializer
+ %1 = bitcast <1 x i64> %shuffle.i250 to <8 x i8>
+ %vmovl.i249 = zext <8 x i8> %1 to <8 x i16>
+ %shuffle.i246 = shufflevector <2 x i64> undef, <2 x i64> undef, <1 x i32> zeroinitializer
+ %shuffle.i240 = shufflevector <2 x i64> undef, <2 x i64> undef, <1 x i32> <i32 1>
+ %2 = bitcast <1 x i64> %shuffle.i240 to <8 x i8>
+ %3 = bitcast <16 x i8> undef to <2 x i64>
+ %vmovl.i237 = zext <8 x i8> undef to <8 x i16>
+ %shuffle.i234 = shufflevector <2 x i64> undef, <2 x i64> undef, <1 x i32> zeroinitializer
+ %shuffle.i226 = shufflevector <2 x i64> undef, <2 x i64> undef, <1 x i32> zeroinitializer
+ %vmovl.i225 = zext <8 x i8> undef to <8 x i16>
+ %mul.i223 = mul <8 x i16> %vmovl.i249, %vmovl.i249
+ %vshl_n = shl <8 x i16> %mul.i223, <i16 2, i16 2, i16 2, i16 2, i16 2, i16 2, i16 2, i16 2>
+ %vqsub2.i216 = tail call <8 x i16> @llvm.arm.neon.vqsubu.v8i16(<8 x i16> <i16 256, i16 256, i16 256, i16 256, i16 256, i16 256, i16 256, i16 256>, <8 x i16> %vshl_n) nounwind
+ %mul.i209 = mul <8 x i16> undef, <i16 80, i16 80, i16 80, i16 80, i16 80, i16 80, i16 80, i16 80>
+ %vshr_n130 = lshr <8 x i16> undef, <i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8>
+ %vshr_n134 = lshr <8 x i16> %mul.i209, <i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8>
+ %sub.i205 = sub <8 x i16> <i16 80, i16 80, i16 80, i16 80, i16 80, i16 80, i16 80, i16 80>, %vshr_n130
+ %sub.i203 = sub <8 x i16> <i16 80, i16 80, i16 80, i16 80, i16 80, i16 80, i16 80, i16 80>, %vshr_n134
+ %add.i200 = add <8 x i16> %sub.i205, <i16 96, i16 96, i16 96, i16 96, i16 96, i16 96, i16 96, i16 96>
+ %add.i198 = add <8 x i16> %add.i200, %sub.i203
+ %mul.i194 = mul <8 x i16> %add.i198, %vmovl.i237
+ %mul.i191 = mul <8 x i16> %vshr_n130, undef
+ %add.i192 = add <8 x i16> %mul.i191, %mul.i194
+ %mul.i187 = mul <8 x i16> %vshr_n134, undef
+ %add.i188 = add <8 x i16> %mul.i187, %add.i192
+ %mul.i185 = mul <8 x i16> undef, undef
+ %add.i186 = add <8 x i16> %mul.i185, undef
+ %vrshr_n160 = tail call <8 x i16> @llvm.arm.neon.vrshiftu.v8i16(<8 x i16> %add.i188, <8 x i16> <i16 -8, i16 -8, i16 -8, i16 -8, i16 -8, i16 -8, i16 -8, i16 -8>)
+ %vrshr_n163 = tail call <8 x i16> @llvm.arm.neon.vrshiftu.v8i16(<8 x i16> %add.i186, <8 x i16> <i16 -8, i16 -8, i16 -8, i16 -8, i16 -8, i16 -8, i16 -8, i16 -8>)
+ %mul.i184 = mul <8 x i16> undef, %vrshr_n160
+ %mul.i181 = mul <8 x i16> undef, %vmovl.i225
+ %add.i182 = add <8 x i16> %mul.i181, %mul.i184
+ %vrshr_n170 = tail call <8 x i16> @llvm.arm.neon.vrshiftu.v8i16(<8 x i16> %add.i182, <8 x i16> <i16 -7, i16 -7, i16 -7, i16 -7, i16 -7, i16 -7, i16 -7, i16 -7>)
+ %vqmovn1.i180 = tail call <8 x i8> @llvm.arm.neon.vqmovnu.v8i8(<8 x i16> %vrshr_n170) nounwind
+ %4 = bitcast <8 x i8> %vqmovn1.i180 to <1 x i64>
+ %shuffle.i = shufflevector <1 x i64> %4, <1 x i64> undef, <2 x i32> <i32 0, i32 1>
+ %5 = bitcast <2 x i64> %shuffle.i to <16 x i8>
+ store <16 x i8> %5, <16 x i8>* undef, align 16
+ %add177 = add nsw i32 undef, 16
+ br i1 undef, label %for.body33, label %for.cond.loopexit
+
+for.end179: ; preds = %for.cond.loopexit, %entry
+ ret void
+}
+
+declare <8 x i16> @llvm.arm.neon.vrshiftu.v8i16(<8 x i16>, <8 x i16>) nounwind readnone
+declare <8 x i16> @llvm.arm.neon.vqsubu.v8i16(<8 x i16>, <8 x i16>) nounwind readnone
+declare <8 x i8> @llvm.arm.neon.vqmovnu.v8i8(<8 x i16>) nounwind readnone