--- /dev/null
+; RUN: llc -march=hexagon -filetype=obj < %s -o - | llvm-objdump -d - | FileCheck %s
+
+; CHECK-LABEL: f0:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = abs(r{{[1-9]}}:{{[0-9]}})
+define double @f0(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.A2.absp(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.A2.absp(i64) #1
+
+; CHECK-LABEL: f1:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = neg(r{{[1-9]}}:{{[0-9]}})
+define double @f1(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.A2.negp(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.A2.negp(i64) #1
+
+; CHECK-LABEL: f2:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = not(r{{[1-9]}}:{{[0-9]}})
+define double @f2(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.A2.notp(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.A2.notp(i64) #1
+
+; CHECK-LABEL: f3:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = interleave(r{{[1-9]}}:{{[0-9]}})
+define double @f3(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.interleave(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.interleave(i64) #1
+
+; CHECK-LABEL: f4:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = deinterleave(r{{[1-9]}}:{{[0-9]}})
+define double @f4(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.deinterleave(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.deinterleave(i64) #1
+
+; CHECK-LABEL: f5:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vconj(r{{[1-9]}}:{{[0-9]}}):sat
+define double @f5(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.A2.vconj(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.A2.vconj(i64) #1
+
+; CHECK-LABEL: f6:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vsathb(r{{[1-9]}}:{{[0-9]}})
+define double @f6(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.vsathb.nopack(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.vsathb.nopack(i64) #1
+
+; CHECK-LABEL: f7:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vsathub(r{{[1-9]}}:{{[0-9]}})
+define double @f7(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.vsathub.nopack(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.vsathub.nopack(i64) #1
+
+; CHECK-LABEL: f8:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vsatwh(r{{[1-9]}}:{{[0-9]}})
+define double @f8(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.vsatwh.nopack(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.vsatwh.nopack(i64) #1
+
+; CHECK-LABEL: f9:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vsatwuh(r{{[1-9]}}:{{[0-9]}})
+define double @f9(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.vsatwuh.nopack(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.vsatwuh.nopack(i64) #1
+
+; CHECK-LABEL: f10:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = asr(r{{[1-9]}}:{{[0-9]}},#1)
+define double @f10(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.asr.i.p(i64 %v2, i32 1)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.asr.i.p(i64, i32) #1
+
+; CHECK-LABEL: f11:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = lsr(r{{[1-9]}}:{{[0-9]}},#1)
+define double @f11(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.lsr.i.p(i64 %v2, i32 1)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.lsr.i.p(i64, i32) #1
+
+; CHECK-LABEL: f12:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = asl(r{{[1-9]}}:{{[0-9]}},#1)
+define double @f12(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.asl.i.p(i64 %v2, i32 1)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.asl.i.p(i64, i32) #1
+
+; CHECK-LABEL: f13:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vabsh(r{{[1-9]}}:{{[0-9]}})
+define double @f13(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.A2.vabsh(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.A2.vabsh(i64) #1
+
+; CHECK-LABEL: f14:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vabsh(r{{[1-9]}}:{{[0-9]}}):sat
+define double @f14(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.A2.vabshsat(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.A2.vabshsat(i64) #1
+
+; CHECK-LABEL: f15:
+; CHECK: r{{[0-9]}}:{{[0-9]}} = vasrh(r{{[1-9]}}:{{[0-9]}},#1)
+define double @f15(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.asr.i.vh(i64 %v2, i32 1)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.asr.i.vh(i64, i32) #1
+
+; CHECK-LABEL: f16:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vlsrh(r{{[1-9]}}:{{[0-9]}},#1)
+define double @f16(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.lsr.i.vh(i64 %v2, i32 1)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.lsr.i.vh(i64, i32) #1
+
+; CHECK-LABEL: f17:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vaslh(r{{[1-9]}}:{{[0-9]}},#1)
+define double @f17(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.asl.i.vh(i64 %v2, i32 1)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.asl.i.vh(i64, i32) #1
+
+; CHECK-LABEL: f18:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vabsw(r{{[1-9]}}:{{[0-9]}})
+define double @f18(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.A2.vabsw(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.A2.vabsw(i64) #1
+
+; CHECK-LABEL: f19:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vabsw(r{{[1-9]}}:{{[0-9]}}):sat
+define double @f19(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.A2.vabswsat(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.A2.vabswsat(i64) #1
+
+; CHECK-LABEL: f20:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vasrw(r{{[1-9]}}:{{[0-9]}},#1)
+define double @f20(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.asr.i.vw(i64 %v2, i32 1)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.asr.i.vw(i64, i32) #1
+
+; CHECK-LABEL: f21:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vlsrw(r{{[1-9]}}:{{[0-9]}},#1)
+define double @f21(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.lsr.i.vw(i64 %v2, i32 1)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.lsr.i.vw(i64, i32) #1
+
+; CHECK-LABEL: f22:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = vaslw(r{{[1-9]}}:{{[0-9]}},#1)
+define double @f22(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.asl.i.vw(i64 %v2, i32 1)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.asl.i.vw(i64, i32) #1
+
+; CHECK-LABEL: f23:
+; CHECK: r{{[1-9]}}:{{[0-9]}} = brev(r{{[1-9]}}:{{[0-9]}})
+define double @f23(double %a0) #0 {
+b0:
+ %v0 = alloca double, align 8
+ store double %a0, double* %v0, align 8
+ %v1 = load double, double* %v0, align 8
+ %v2 = fptosi double %v1 to i64
+ %v3 = call i64 @llvm.hexagon.S2.brevp(i64 %v2)
+ %v4 = sitofp i64 %v3 to double
+ ret double %v4
+}
+
+declare i64 @llvm.hexagon.S2.brevp(i64) #1
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }
--- /dev/null
+; RUN: llc -O2 -march=hexagon -fp-contract=fast -pipeliner-prune-loop-carried=false < %s | FileCheck %s
+
+; Test that there is 1 packet between the FP result and its use.
+
+; CHECK: loop0([[LOOP0:.LBB[0-9_]+]],
+; CHECK: [[LOOP0]]
+; CHECK: [[REG0:(r[0-9]+)]] += sfmpy(r{{[0-9]+}},r{{[0-9]+}})
+; CHECK: }
+; CHECK: }
+; CHECK: r{{[0-9]+}} = {{.*}}[[REG0]]
+
+; Function Attrs: nounwind readnone
+define void @f0(i32 %a0, i32 %a1) #0 {
+b0:
+ %v0 = alloca [1000 x float], align 64
+ %v1 = alloca [1000 x float], align 64
+ %v2 = alloca [1000 x float], align 64
+ %v3 = alloca [1000 x float], align 64
+ %v4 = bitcast [1000 x float]* %v0 to i8*
+ call void @llvm.lifetime.start.p0i8(i64 4000, i8* %v4) #2
+ %v5 = bitcast [1000 x float]* %v1 to i8*
+ call void @llvm.lifetime.start.p0i8(i64 4000, i8* %v5) #2
+ %v6 = bitcast [1000 x float]* %v2 to i8*
+ call void @llvm.lifetime.start.p0i8(i64 4000, i8* %v6) #2
+ %v7 = bitcast [1000 x float]* %v3 to i8*
+ call void @llvm.lifetime.start.p0i8(i64 4000, i8* %v7) #2
+ %v8 = icmp sgt i32 %a1, 0
+ %v9 = add i32 %a1, -1
+ %v10 = getelementptr [1000 x float], [1000 x float]* %v3, i32 0, i32 0
+ br label %b1
+
+b1: ; preds = %b3, %b0
+ %v11 = phi i32 [ 0, %b0 ], [ %v34, %b3 ]
+ br i1 %v8, label %b2, label %b3
+
+b2: ; preds = %b2, %b1
+ %v12 = phi float* [ %v33, %b2 ], [ %v10, %b1 ]
+ %v13 = phi i32 [ %v31, %b2 ], [ 0, %b1 ]
+ %v14 = mul nsw i32 %v13, %a1
+ %v15 = add nsw i32 %v14, %v11
+ %v16 = getelementptr inbounds [1000 x float], [1000 x float]* %v1, i32 0, i32 %v15
+ %v17 = load float, float* %v16, align 4, !tbaa !0
+ %v18 = fmul float %v17, undef
+ %v19 = mul nsw i32 %v13, 25
+ %v20 = add nsw i32 %v19, %v11
+ %v21 = getelementptr inbounds [1000 x float], [1000 x float]* %v2, i32 0, i32 %v20
+ %v22 = load float, float* %v21, align 4, !tbaa !0
+ %v23 = fmul float %v22, undef
+ %v24 = fadd float %v18, %v23
+ %v25 = load float, float* %v12, align 4, !tbaa !0
+ %v26 = fmul float %v25, undef
+ %v27 = fadd float %v24, %v26
+ %v28 = getelementptr inbounds [1000 x float], [1000 x float]* %v0, i32 0, i32 %v20
+ %v29 = load float, float* %v28, align 4, !tbaa !0
+ %v30 = fadd float %v29, %v27
+ store float %v30, float* %v28, align 4, !tbaa !0
+ %v31 = add nuw nsw i32 %v13, 1
+ %v32 = icmp eq i32 %v13, %v9
+ %v33 = getelementptr float, float* %v12, i32 1
+ br i1 %v32, label %b3, label %b2
+
+b3: ; preds = %b2, %b1
+ %v34 = add nuw nsw i32 %v11, 1
+ %v35 = icmp eq i32 %v34, 25
+ br i1 %v35, label %b4, label %b1
+
+b4: ; preds = %b3
+ call void @llvm.lifetime.end.p0i8(i64 4000, i8* %v7) #2
+ call void @llvm.lifetime.end.p0i8(i64 4000, i8* %v6) #2
+ call void @llvm.lifetime.end.p0i8(i64 4000, i8* %v5) #2
+ call void @llvm.lifetime.end.p0i8(i64 4000, i8* %v4) #2
+ ret void
+}
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64, i8* nocapture) #1
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64, i8* nocapture) #1
+
+attributes #0 = { nounwind readnone "target-cpu"="hexagonv60" "target-features"="+hvxv60,+hvx-length64b" }
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { nounwind }
+
+!0 = !{!1, !1, i64 0}
+!1 = !{!"float", !2, i64 0}
+!2 = !{!"omnipotent char", !3, i64 0}
+!3 = !{!"Simple C/C++ TBAA"}
--- /dev/null
+; RUN: llc -O2 -march=hexagon < %s | FileCheck %s
+; Test that we do generate max #u5 in memops.
+; CHECK: memb(r{{[0-9]+}}+#0) -= #31
+
+@g0 = unnamed_addr global i8 112, align 1
+
+; Function Attrs: norecurse nounwind
+define fastcc void @f0() unnamed_addr #0 {
+b0:
+ %v0 = load i8, i8* @g0, align 1, !tbaa !4
+ %v1 = zext i8 %v0 to i32
+ %v2 = mul nuw nsw i32 %v1, 9625
+ %v3 = and i32 %v2, 255
+ %v4 = mul nuw nsw i32 %v3, 9625
+ %v5 = and i32 %v4, 255
+ %v6 = trunc i32 %v5 to i8
+ store i8 %v6, i8* @g0, align 1, !tbaa !4
+ ret void
+}
+
+define i32 @f1() {
+b0:
+ %v0 = load i8, i8* @g0, align 1, !tbaa !4
+ %v1 = zext i8 %v0 to i32
+ %v2 = add nuw nsw i32 %v1, 225
+ %v3 = trunc i32 %v2 to i8
+ store i8 %v3, i8* @g0, align 1, !tbaa !4
+ tail call fastcc void @f0()
+ %v4 = load i8, i8* @g0, align 1, !tbaa !4
+ %v5 = zext i8 %v4 to i32
+ ret i32 %v5
+}
+
+attributes #0 = { norecurse nounwind "target-cpu"="hexagonv55" }
+
+!llvm.module.flags = !{!0, !2}
+
+!0 = !{i32 6, !"Target CPU", !1}
+!1 = !{!"hexagonv55"}
+!2 = !{i32 6, !"Target Features", !3}
+!3 = !{!"-hvx"}
+!4 = !{!5, !5, i64 0}
+!5 = !{!"omnipotent char", !6, i64 0}
+!6 = !{!"Simple C/C++ TBAA"}
--- /dev/null
+; RUN: llc -O2 -march=hexagon < %s | FileCheck %s
+; Test that we do generate max #u5 in memops.
+; CHECK: memh(r{{[0-9]+}}+#0) -= #31
+
+@g0 = unnamed_addr global i16 -32, align 2
+
+; Function Attrs: norecurse nounwind
+define fastcc void @f0() unnamed_addr #0 {
+b0:
+ %v0 = load i16, i16* @g0, align 1, !tbaa !4
+ %v1 = zext i16 %v0 to i32
+ %v2 = mul nuw nsw i32 %v1, 9625
+ %v3 = and i32 %v2, 255
+ %v4 = mul nuw nsw i32 %v3, 9625
+ %v5 = and i32 %v4, 255
+ %v6 = trunc i32 %v5 to i16
+ store i16 %v6, i16* @g0, align 2, !tbaa !4
+ ret void
+}
+
+define i32 @f1() {
+b0:
+ %v0 = load i16, i16* @g0, align 2, !tbaa !4
+ %v1 = zext i16 %v0 to i32
+ %v2 = add nuw nsw i32 %v1, 65505
+ %v3 = trunc i32 %v2 to i16
+ store i16 %v3, i16* @g0, align 2, !tbaa !4
+ tail call fastcc void @f0()
+ %v4 = load i16, i16* @g0, align 2, !tbaa !4
+ %v5 = zext i16 %v4 to i32
+ ret i32 %v5
+}
+
+attributes #0 = { norecurse nounwind "target-cpu"="hexagonv55" }
+
+!llvm.module.flags = !{!0, !2}
+
+!0 = !{i32 6, !"Target CPU", !1}
+!1 = !{!"hexagonv55"}
+!2 = !{i32 6, !"Target Features", !3}
+!3 = !{!"-hvx"}
+!4 = !{!5, !5, i64 0}
+!5 = !{!"omnipotent char", !6, i64 0}
+!6 = !{!"Simple C/C++ TBAA"}
ret i32 %v5
}
-attributes #0 = { norecurse nounwind }
+attributes #0 = { norecurse nounwind "target-cpu"="hexagonv55" }
!llvm.module.flags = !{!0, !2}
--- /dev/null
+; RUN: llc -march=hexagon < %s | FileCheck %s
+
+; Check that store is post-incremented.
+; CHECK: memd(r{{[0-9]+}}++#8) = r
+
+; Function Attrs: nounwind
+define void @f0(i32 %a0, i16* nocapture %a1, i16 signext %a2) #0 {
+b0:
+ %v0 = icmp eq i32 %a0, 0
+ br i1 %v0, label %b2, label %b3
+
+b1: ; preds = %b10
+ br label %b2
+
+b2: ; preds = %b7, %b1, %b0
+ ret void
+
+b3: ; preds = %b0
+ %v1 = icmp sgt i32 %a0, 3
+ br i1 %v1, label %b4, label %b7
+
+b4: ; preds = %b3
+ %v2 = add i32 %a0, -1
+ %v3 = and i32 %v2, -4
+ %v4 = icmp sgt i32 %v3, 0
+ br i1 %v4, label %b5, label %b7
+
+b5: ; preds = %b4
+ %v5 = insertelement <4 x i16> undef, i16 %a2, i32 0
+ %v6 = insertelement <4 x i16> %v5, i16 %a2, i32 1
+ %v7 = insertelement <4 x i16> %v6, i16 %a2, i32 2
+ %v8 = insertelement <4 x i16> %v7, i16 %a2, i32 3
+ br label %b9
+
+b6: ; preds = %b9
+ br label %b7
+
+b7: ; preds = %b6, %b4, %b3
+ %v9 = phi i32 [ 0, %b3 ], [ %v3, %b4 ], [ %v3, %b6 ]
+ %v10 = icmp slt i32 %v9, %a0
+ br i1 %v10, label %b8, label %b2
+
+b8: ; preds = %b7
+ br label %b10
+
+b9: ; preds = %b9, %b5
+ %v11 = phi i32 [ 0, %b5 ], [ %v12, %b9 ]
+ %v12 = add nsw i32 %v11, 4
+ %v13 = getelementptr i16, i16* %a1, i32 %v11
+ %v14 = bitcast i16* %v13 to <4 x i16>*
+ %v15 = load <4 x i16>, <4 x i16>* %v14, align 16
+ %v16 = add <4 x i16> %v15, %v8
+ store <4 x i16> %v16, <4 x i16>* %v14, align 16
+ %v17 = icmp slt i32 %v12, %v3
+ br i1 %v17, label %b9, label %b6
+
+b10: ; preds = %b10, %b8
+ %v18 = phi i32 [ %v19, %b10 ], [ %v9, %b8 ]
+ %v19 = add nsw i32 %v18, 1
+ %v20 = getelementptr i16, i16* %a1, i32 %v18
+ %v21 = load i16, i16* %v20, align 2
+ %v22 = add i16 %v21, %a2
+ store i16 %v22, i16* %v20, align 2
+ %v23 = icmp eq i32 %v19, %a0
+ br i1 %v23, label %b1, label %b10
+}
+
+attributes #0 = { nounwind "target-cpu"="hexagonv55" }
--- /dev/null
+; RUN: llc -march=hexagon -enable-pipeliner -pipeliner-max-stages=1 < %s | FileCheck %s
+
+; Test that we update the offset correctly for loads that are
+; moved past stores. In these cases, we change the dependences
+; to make it easier to move the instructions, and we have to update
+; the register/offsets correctly after the schedule is finalized.
+
+@g0 = common global [400 x i32] zeroinitializer, align 8
+@g1 = common global [400 x i32] zeroinitializer, align 8
+
+; Function Attrs: nounwind
+define void @f0() #0 {
+b0:
+ br label %b2
+
+b1: ; preds = %b2
+ ret void
+
+; CHECK: loop0(.LBB0_[[LOOP:.]],
+; CHECK: .LBB0_[[LOOP]]:
+; CHECK: = memd([[REG1:(r[0-9]+)]]+#8)
+; CHECK: memd([[REG1]]++#8) =
+; CHECK: }{{[ \t]*}}:endloop
+
+b2: ; preds = %b2, %b0
+ %v0 = phi i32* [ getelementptr inbounds ([400 x i32], [400 x i32]* @g0, i32 0, i32 0), %b0 ], [ %v11, %b2 ]
+ %v1 = phi i32* [ getelementptr inbounds ([400 x i32], [400 x i32]* @g1, i32 0, i32 0), %b0 ], [ %v12, %b2 ]
+ %v2 = phi i32 [ 0, %b0 ], [ %v9, %b2 ]
+ %v3 = bitcast i32* %v0 to <2 x i32>*
+ %v4 = load <2 x i32>, <2 x i32>* %v3, align 8
+ %v5 = mul <2 x i32> %v4, <i32 7, i32 7>
+ %v6 = bitcast i32* %v1 to <2 x i32>*
+ %v7 = load <2 x i32>, <2 x i32>* %v6, align 8
+ %v8 = add <2 x i32> %v7, %v5
+ store <2 x i32> %v8, <2 x i32>* %v6, align 8
+ %v9 = add nsw i32 %v2, 2
+ %v10 = icmp slt i32 %v2, 398
+ %v11 = getelementptr i32, i32* %v0, i32 2
+ %v12 = getelementptr i32, i32* %v1, i32 2
+ br i1 %v10, label %b2, label %b1
+}
+
+attributes #0 = { nounwind "target-cpu"="hexagonv55" }
--- /dev/null
+; RUN: llc -march=hexagon < %s | FileCheck %s
+
+; This version of the conv3x3 test has both loops. This test checks that the
+; inner loop has 13 packets.
+
+; CHECK: loop0(.LBB0_[[LOOP:.]],
+; CHECK: .LBB0_[[LOOP]]:
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK-NOT: }
+; CHECK: }{{[ \t]*}}:endloop0
+
+declare <16 x i32> @llvm.hexagon.V6.vd0() #0
+declare <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32>, <16 x i32>, i32) #0
+declare <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32>, <16 x i32>) #0
+declare <32 x i32> @llvm.hexagon.V6.vrmpybusi(<32 x i32>, i32, i32) #0
+declare <32 x i32> @llvm.hexagon.V6.vrmpybusi.acc(<32 x i32>, <32 x i32>, i32, i32) #0
+declare <16 x i32> @llvm.hexagon.V6.vasrwhsat(<16 x i32>, <16 x i32>, i32) #0
+declare <16 x i32> @llvm.hexagon.V6.hi(<32 x i32>) #0
+declare <16 x i32> @llvm.hexagon.V6.lo(<32 x i32>) #0
+declare <16 x i32> @llvm.hexagon.V6.vsathub(<16 x i32>, <16 x i32>) #0
+declare <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32>, <16 x i32>, i32) #0
+
+define void @f0(i8* noalias nocapture readonly %a0, i32 %a1, i32 %a2, i32 %a3, i8* noalias nocapture readonly %a4, i32 %a5, i8* noalias nocapture %a6) local_unnamed_addr #1 {
+b0:
+ %v0 = add nsw i32 %a3, -1
+ %v1 = icmp sgt i32 %a3, 2
+ br i1 %v1, label %b1, label %b6
+
+b1: ; preds = %b0
+ %v2 = getelementptr inbounds i8, i8* %a6, i32 %a1
+ %v3 = getelementptr inbounds i8, i8* %a0, i32 %a1
+ %v4 = bitcast i8* %a4 to i32*
+ %v5 = load i32, i32* %v4, align 4, !tbaa !1, !alias.scope !5, !noalias !8
+ %v6 = getelementptr inbounds i8, i8* %a4, i32 4
+ %v7 = bitcast i8* %v6 to i32*
+ %v8 = load i32, i32* %v7, align 4, !tbaa !1, !alias.scope !5, !noalias !8
+ %v9 = getelementptr inbounds i8, i8* %a4, i32 8
+ %v10 = bitcast i8* %v9 to i32*
+ %v11 = load i32, i32* %v10, align 4, !tbaa !1, !alias.scope !5, !noalias !8
+ %v12 = sub i32 0, %a1
+ %v13 = shl nsw i32 %a1, 1
+ %v14 = tail call <16 x i32> @llvm.hexagon.V6.vd0() #2
+ %v15 = icmp sgt i32 %a2, 0
+ br label %b2
+
+b2: ; preds = %b5, %b1
+ %v16 = phi i8* [ %v2, %b1 ], [ %v102, %b5 ]
+ %v17 = phi i8* [ %v3, %b1 ], [ %v21, %b5 ]
+ %v18 = phi i32 [ 1, %b1 ], [ %v103, %b5 ]
+ %v19 = getelementptr inbounds i8, i8* %v17, i32 %v12
+ %v20 = getelementptr inbounds i8, i8* %v17, i32 %a1
+ %v21 = getelementptr inbounds i8, i8* %v17, i32 %v13
+ br i1 %v15, label %b3, label %b5
+
+b3: ; preds = %b2
+ %v22 = bitcast i8* %v21 to <16 x i32>*
+ %v23 = load <16 x i32>, <16 x i32>* %v22, align 64, !tbaa !11, !alias.scope !12, !noalias !13
+ %v24 = getelementptr inbounds i8, i8* %v21, i32 64
+ %v25 = bitcast i8* %v24 to <16 x i32>*
+ %v26 = bitcast i8* %v20 to <16 x i32>*
+ %v27 = load <16 x i32>, <16 x i32>* %v26, align 64, !tbaa !11, !alias.scope !12, !noalias !13
+ %v28 = getelementptr inbounds i8, i8* %v20, i32 64
+ %v29 = bitcast i8* %v28 to <16 x i32>*
+ %v30 = bitcast i8* %v17 to <16 x i32>*
+ %v31 = load <16 x i32>, <16 x i32>* %v30, align 64, !tbaa !11, !alias.scope !12, !noalias !13
+ %v32 = getelementptr inbounds i8, i8* %v17, i32 64
+ %v33 = bitcast i8* %v32 to <16 x i32>*
+ %v34 = bitcast i8* %v19 to <16 x i32>*
+ %v35 = load <16 x i32>, <16 x i32>* %v34, align 64, !tbaa !11, !alias.scope !12, !noalias !13
+ %v36 = getelementptr inbounds i8, i8* %v19, i32 64
+ %v37 = bitcast i8* %v36 to <16 x i32>*
+ %v38 = getelementptr inbounds i8, i8* %v16, i32 %a1
+ %v39 = bitcast i8* %v38 to <16 x i32>*
+ %v40 = bitcast i8* %v16 to <16 x i32>*
+ br label %b4
+
+b4: ; preds = %b4, %b3
+ %v41 = phi <16 x i32>* [ %v39, %b3 ], [ %v99, %b4 ]
+ %v42 = phi <16 x i32>* [ %v40, %b3 ], [ %v84, %b4 ]
+ %v43 = phi <16 x i32>* [ %v25, %b3 ], [ %v60, %b4 ]
+ %v44 = phi <16 x i32>* [ %v29, %b3 ], [ %v58, %b4 ]
+ %v45 = phi <16 x i32>* [ %v33, %b3 ], [ %v56, %b4 ]
+ %v46 = phi <16 x i32>* [ %v37, %b3 ], [ %v54, %b4 ]
+ %v47 = phi i32 [ %a2, %b3 ], [ %v100, %b4 ]
+ %v48 = phi <16 x i32> [ %v35, %b3 ], [ %v55, %b4 ]
+ %v49 = phi <16 x i32> [ %v31, %b3 ], [ %v57, %b4 ]
+ %v50 = phi <16 x i32> [ %v27, %b3 ], [ %v59, %b4 ]
+ %v51 = phi <16 x i32> [ %v23, %b3 ], [ %v61, %b4 ]
+ %v52 = phi <16 x i32> [ %v14, %b3 ], [ %v82, %b4 ]
+ %v53 = phi <16 x i32> [ %v14, %b3 ], [ %v97, %b4 ]
+ %v54 = getelementptr inbounds <16 x i32>, <16 x i32>* %v46, i32 1
+ %v55 = load <16 x i32>, <16 x i32>* %v46, align 64, !tbaa !11, !alias.scope !12, !noalias !13
+ %v56 = getelementptr inbounds <16 x i32>, <16 x i32>* %v45, i32 1
+ %v57 = load <16 x i32>, <16 x i32>* %v45, align 64, !tbaa !11, !alias.scope !12, !noalias !13
+ %v58 = getelementptr inbounds <16 x i32>, <16 x i32>* %v44, i32 1
+ %v59 = load <16 x i32>, <16 x i32>* %v44, align 64, !tbaa !11, !alias.scope !12, !noalias !13
+ %v60 = getelementptr inbounds <16 x i32>, <16 x i32>* %v43, i32 1
+ %v61 = load <16 x i32>, <16 x i32>* %v43, align 64, !tbaa !11, !alias.scope !12, !noalias !13
+ %v62 = tail call <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32> %v55, <16 x i32> %v48, i32 4) #2
+ %v63 = tail call <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32> %v57, <16 x i32> %v49, i32 4) #2
+ %v64 = tail call <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32> %v59, <16 x i32> %v50, i32 4) #2
+ %v65 = tail call <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32> %v61, <16 x i32> %v51, i32 4) #2
+ %v66 = tail call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %v62, <16 x i32> %v48) #2
+ %v67 = tail call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %v63, <16 x i32> %v49) #2
+ %v68 = tail call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %v64, <16 x i32> %v50) #2
+ %v69 = tail call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %v65, <16 x i32> %v51) #2
+ %v70 = tail call <32 x i32> @llvm.hexagon.V6.vrmpybusi(<32 x i32> %v66, i32 %v5, i32 0) #2
+ %v71 = tail call <32 x i32> @llvm.hexagon.V6.vrmpybusi(<32 x i32> %v66, i32 %v5, i32 1) #2
+ %v72 = tail call <32 x i32> @llvm.hexagon.V6.vrmpybusi.acc(<32 x i32> %v70, <32 x i32> %v67, i32 %v8, i32 0) #2
+ %v73 = tail call <32 x i32> @llvm.hexagon.V6.vrmpybusi.acc(<32 x i32> %v71, <32 x i32> %v67, i32 %v8, i32 1) #2
+ %v74 = tail call <32 x i32> @llvm.hexagon.V6.vrmpybusi.acc(<32 x i32> %v72, <32 x i32> %v68, i32 %v11, i32 0) #2
+ %v75 = tail call <32 x i32> @llvm.hexagon.V6.vrmpybusi.acc(<32 x i32> %v73, <32 x i32> %v68, i32 %v11, i32 1) #2
+ %v76 = tail call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %v75) #2
+ %v77 = tail call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %v75) #2
+ %v78 = tail call <16 x i32> @llvm.hexagon.V6.vasrwhsat(<16 x i32> %v76, <16 x i32> %v77, i32 %a5) #2
+ %v79 = tail call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %v74) #2
+ %v80 = tail call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %v74) #2
+ %v81 = tail call <16 x i32> @llvm.hexagon.V6.vasrwhsat(<16 x i32> %v79, <16 x i32> %v80, i32 %a5) #2
+ %v82 = tail call <16 x i32> @llvm.hexagon.V6.vsathub(<16 x i32> %v78, <16 x i32> %v81) #2
+ %v83 = tail call <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32> %v82, <16 x i32> %v52, i32 1) #2
+ %v84 = getelementptr inbounds <16 x i32>, <16 x i32>* %v42, i32 1
+ store <16 x i32> %v83, <16 x i32>* %v42, align 64, !tbaa !11, !alias.scope !14, !noalias !15
+ %v85 = tail call <32 x i32> @llvm.hexagon.V6.vrmpybusi(<32 x i32> %v67, i32 %v5, i32 0) #2
+ %v86 = tail call <32 x i32> @llvm.hexagon.V6.vrmpybusi(<32 x i32> %v67, i32 %v5, i32 1) #2
+ %v87 = tail call <32 x i32> @llvm.hexagon.V6.vrmpybusi.acc(<32 x i32> %v85, <32 x i32> %v68, i32 %v8, i32 0) #2
+ %v88 = tail call <32 x i32> @llvm.hexagon.V6.vrmpybusi.acc(<32 x i32> %v86, <32 x i32> %v68, i32 %v8, i32 1) #2
+ %v89 = tail call <32 x i32> @llvm.hexagon.V6.vrmpybusi.acc(<32 x i32> %v87, <32 x i32> %v69, i32 %v11, i32 0) #2
+ %v90 = tail call <32 x i32> @llvm.hexagon.V6.vrmpybusi.acc(<32 x i32> %v88, <32 x i32> %v69, i32 %v11, i32 1) #2
+ %v91 = tail call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %v90) #2
+ %v92 = tail call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %v90) #2
+ %v93 = tail call <16 x i32> @llvm.hexagon.V6.vasrwhsat(<16 x i32> %v91, <16 x i32> %v92, i32 %a5) #2
+ %v94 = tail call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %v89) #2
+ %v95 = tail call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %v89) #2
+ %v96 = tail call <16 x i32> @llvm.hexagon.V6.vasrwhsat(<16 x i32> %v94, <16 x i32> %v95, i32 %a5) #2
+ %v97 = tail call <16 x i32> @llvm.hexagon.V6.vsathub(<16 x i32> %v93, <16 x i32> %v96) #2
+ %v98 = tail call <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32> %v97, <16 x i32> %v53, i32 1) #2
+ %v99 = getelementptr inbounds <16 x i32>, <16 x i32>* %v41, i32 1
+ store <16 x i32> %v98, <16 x i32>* %v41, align 64, !tbaa !11, !alias.scope !14, !noalias !15
+ %v100 = add nsw i32 %v47, -64
+ %v101 = icmp sgt i32 %v47, 64
+ br i1 %v101, label %b4, label %b5
+
+b5: ; preds = %b4, %b2
+ %v102 = getelementptr inbounds i8, i8* %v16, i32 %v13
+ %v103 = add nuw nsw i32 %v18, 2
+ %v104 = icmp slt i32 %v103, %v0
+ br i1 %v104, label %b2, label %b6
+
+b6: ; preds = %b5, %b0
+ ret void
+}
+
+attributes #0 = { nounwind readnone }
+attributes #1 = { nounwind "target-cpu"="hexagonv62" "target-features"="+hvx-length64b,+hvxv62" }
+attributes #2 = { nounwind }
+
+!llvm.module.flags = !{!0}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!2, !2, i64 0}
+!2 = !{!"int", !3, i64 0}
+!3 = !{!"omnipotent char", !4, i64 0}
+!4 = !{!"Simple C/C++ TBAA"}
+!5 = !{!6}
+!6 = distinct !{!6, !7, !"x: %a"}
+!7 = distinct !{!7, !"x"}
+!8 = !{!9, !10}
+!9 = distinct !{!9, !7, !"x: %b"}
+!10 = distinct !{!10, !7, !"x: %c"}
+!11 = !{!3, !3, i64 0}
+!12 = !{!9}
+!13 = !{!6, !10}
+!14 = !{!10}
+!15 = !{!9, !6}
--- /dev/null
+; RUN: llc -march=hexagon -enable-pipeliner < %s | FileCheck %s
+
+; Test that the code that changes the dependences does not allow
+; a load with a negative offset to be overlapped with the post
+; increment store that generates the base register.
+
+; CHECK: loop0(.LBB0_[[LOOP:.]],
+; CHECK: .LBB0_[[LOOP]]:
+; CHECK: = mem{{u?}}b([[REG:(r[0-9])+]]+#-1)
+; CHECK-NOT: memb([[REG]]{{\+?}}#0) =
+; CHECK: }
+; CHECK: }{{[ \t]*}}:endloop0
+
+@g0 = external global [1000000 x i8], align 8
+
+; Function Attrs: nounwind
+define void @f0(i32 %a0, [1000 x i8]* %a1, [1000 x i8]* %a2) #0 {
+b0:
+ br i1 undef, label %b1, label %b7
+
+b1: ; preds = %b0
+ br i1 undef, label %b2, label %b6
+
+b2: ; preds = %b5, %b1
+ br i1 undef, label %b3, label %b5
+
+b3: ; preds = %b3, %b2
+ %v0 = phi i32 [ %v17, %b3 ], [ 1, %b2 ]
+ %v1 = phi i32 [ %v16, %b3 ], [ 0, %b2 ]
+ %v2 = add nsw i32 %v0, -1
+ %v3 = getelementptr inbounds [1000 x i8], [1000 x i8]* %a1, i32 undef, i32 %v2
+ %v4 = load i8, i8* %v3, align 1, !tbaa !0
+ %v5 = zext i8 %v4 to i32
+ %v6 = getelementptr inbounds [1000000 x i8], [1000000 x i8]* @g0, i32 0, i32 %v1
+ %v7 = load i8, i8* %v6, align 1, !tbaa !0
+ %v8 = sext i8 %v7 to i32
+ %v9 = getelementptr inbounds [1000 x i8], [1000 x i8]* %a2, i32 undef, i32 %v0
+ %v10 = load i8, i8* %v9, align 1, !tbaa !0
+ %v11 = sext i8 %v10 to i32
+ %v12 = mul nsw i32 %v11, %v8
+ %v13 = add nsw i32 %v12, %v5
+ %v14 = trunc i32 %v13 to i8
+ %v15 = getelementptr inbounds [1000 x i8], [1000 x i8]* %a1, i32 undef, i32 %v0
+ store i8 %v14, i8* %v15, align 1, !tbaa !0
+ %v16 = add nsw i32 %v1, 1
+ %v17 = add nsw i32 %v0, 1
+ %v18 = icmp eq i32 %v17, %a0
+ br i1 %v18, label %b4, label %b3
+
+b4: ; preds = %b3
+ br label %b5
+
+b5: ; preds = %b4, %b2
+ br i1 undef, label %b6, label %b2
+
+b6: ; preds = %b5, %b1
+ unreachable
+
+b7: ; preds = %b0
+ ret void
+}
+
+attributes #0 = { nounwind "target-cpu"="hexagonv55" }
+
+!0 = !{!1, !1, i64 0}
+!1 = !{!"omnipotent char", !2, i64 0}
+!2 = !{!"Simple C/C++ TBAA"}
--- /dev/null
+; RUN: llc -march=hexagon -enable-pipeliner -pipeliner-max-stages=3 < %s | FileCheck %s
+
+%s.0 = type { i16, i8, i8, i16, i8, i8, i16, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i32, i16, i8, i8, %s.1, [2 x [16 x %s.2]], i32 (i8*, i8*, i8*, i8*, i8*)*, %s.3*, %s.3*, [120 x i8], i8, i8, %s.4*, [2 x [120 x [8 x i8]]], [56 x i8], [2 x [121 x %s.5]], [2 x %s.5], %s.5*, %s.5*, i32, i32, i16, i8, i8, %s.7, %s.9, %s.11, %s.8*, %s.8* }
+%s.1 = type { i8, i8, i8, i8, i8, i8, i8, i8, i32, i8, [16 x i8], i8, [4 x i8], [32 x i16], [32 x i16], [2 x i8], [4 x i8], [2 x [4 x i8]], [2 x [4 x i8]], i32, i32, i16, i8 }
+%s.2 = type { [2 x i16] }
+%s.3 = type { i16*, i16*, i32, i32 }
+%s.4 = type { i8*, i8*, i8*, i32, i32, i32, i32 }
+%s.5 = type { %s.6, [2 x [4 x %s.2]], [2 x [2 x i8]], [2 x i8] }
+%s.6 = type { i8, i8, i8, i8, i8, i8, i8, i8, i32 }
+%s.7 = type { [12 x %s.8], [4 x %s.8], [2 x %s.8], [4 x %s.8], [6 x %s.8], [2 x [7 x %s.8]], [4 x %s.8], [3 x [4 x %s.8]], [3 x %s.8], [3 x %s.8] }
+%s.8 = type { i8, i8 }
+%s.9 = type { [371 x %s.8], [6 x %s.10] }
+%s.10 = type { %s.8*, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8 }
+%s.11 = type { i32, i32, i8* }
+
+; Function Attrs: nounwind
+define void @f0(%s.0* %a0) #0 {
+b0:
+ %v0 = load i8, i8* undef, align 1, !tbaa !0
+ %v1 = icmp eq i8 %v0, 1
+ br i1 %v1, label %b1, label %b2
+
+; CHECK: loop0(.LBB0_[[LOOP:.]],
+; CHECK: .LBB0_[[LOOP]]:
+; CHECK: memh([[REG0:(r[0-9]+)]]+#0) = #0
+; CHECK: }{{[ \t]*}}:endloop0
+
+b1: ; preds = %b1, %b0
+ %v2 = phi i16* [ %v17, %b1 ], [ undef, %b0 ]
+ %v3 = phi i32 [ %v18, %b1 ], [ 0, %b0 ]
+ %v4 = getelementptr inbounds %s.0, %s.0* %a0, i32 0, i32 25, i32 10, i32 %v3
+ %v5 = load i8, i8* %v4, align 1, !tbaa !0
+ %v6 = zext i8 %v5 to i16
+ %v7 = add nsw i32 %v3, 1
+ %v8 = getelementptr inbounds %s.0, %s.0* %a0, i32 0, i32 25, i32 10, i32 %v7
+ %v9 = load i8, i8* %v8, align 1, !tbaa !0
+ %v10 = or i16 0, %v6
+ %v11 = load i8, i8* undef, align 1, !tbaa !0
+ %v12 = zext i8 %v11 to i16
+ %v13 = shl nuw i16 %v12, 8
+ %v14 = or i16 %v10, %v13
+ %v15 = or i16 %v14, 0
+ %v16 = getelementptr inbounds i16, i16* %v2, i32 1
+ store i16* %v16, i16** null, align 4, !tbaa !3
+ store i16 %v15, i16* %v2, align 2, !tbaa !5
+ %v17 = getelementptr inbounds i16, i16* %v2, i32 2
+ store i16* %v17, i16** null, align 4, !tbaa !3
+ store i16 0, i16* %v16, align 2, !tbaa !5
+ %v18 = add nsw i32 %v3, 8
+ %v19 = icmp slt i32 %v18, undef
+ br i1 %v19, label %b1, label %b2
+
+b2: ; preds = %b1, %b0
+ ret void
+}
+
+attributes #0 = { nounwind "target-cpu"="hexagonv55" }
+
+!0 = !{!1, !1, i64 0}
+!1 = !{!"omnipotent char", !2}
+!2 = !{!"Simple C/C++ TBAA"}
+!3 = !{!4, !4, i64 0}
+!4 = !{!"any pointer", !1}
+!5 = !{!6, !6, i64 0}
+!6 = !{!"short", !1}
--- /dev/null
+; RUN: llc -march=hexagon -pipeliner-ignore-recmii -pipeliner-max-stages=2 -enable-pipeliner < %s | FileCheck %s
+
+; This is a loop we pipeline to three packets, though we could do bettter.
+
+; CHECK: loop0(.LBB0_[[LOOP:.]],
+; CHECK: .LBB0_[[LOOP]]:
+; CHECK: {
+; CHECK: }
+; CHECK: {
+; CHECK: }
+; CHECK: {
+; CHECK: }{{[ \t]*}}:endloop0
+
+; Function Attrs: nounwind
+define void @f0(i32* nocapture %a0, i16 signext %a1) #0 {
+b0:
+ %v0 = sext i16 %a1 to i32
+ %v1 = add i32 %v0, -1
+ %v2 = icmp sgt i32 %v1, 0
+ br i1 %v2, label %b1, label %b4
+
+b1: ; preds = %b0
+ %v3 = getelementptr i32, i32* %a0, i32 %v1
+ %v4 = load i32, i32* %v3, align 4
+ br label %b2
+
+b2: ; preds = %b2, %b1
+ %v5 = phi i32 [ %v16, %b2 ], [ %v1, %b1 ]
+ %v6 = phi i32 [ %v5, %b2 ], [ %v0, %b1 ]
+ %v7 = phi i32 [ %v10, %b2 ], [ %v4, %b1 ]
+ %v8 = add nsw i32 %v6, -2
+ %v9 = getelementptr inbounds i32, i32* %a0, i32 %v8
+ %v10 = load i32, i32* %v9, align 4, !tbaa !0
+ %v11 = tail call i64 @llvm.hexagon.M2.dpmpyss.s0(i32 %v10, i32 7946)
+ %v12 = tail call i64 @llvm.hexagon.S2.asl.r.p(i64 %v11, i32 -13)
+ %v13 = getelementptr inbounds i32, i32* %a0, i32 %v5
+ %v14 = tail call i32 @llvm.hexagon.A2.sat(i64 %v12)
+ %v15 = tail call i32 @llvm.hexagon.A2.subsat(i32 %v7, i32 %v14)
+ store i32 %v15, i32* %v13, align 4, !tbaa !0
+ %v16 = add nsw i32 %v5, -1
+ %v17 = icmp sgt i32 %v16, 0
+ br i1 %v17, label %b2, label %b3
+
+b3: ; preds = %b2
+ br label %b4
+
+b4: ; preds = %b3, %b0
+ ret void
+}
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hexagon.S2.asl.r.p(i64, i32) #1
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hexagon.M2.dpmpyss.s0(i32, i32) #1
+
+; Function Attrs: nounwind readnone
+declare i32 @llvm.hexagon.A2.subsat(i32, i32) #1
+
+; Function Attrs: nounwind readnone
+declare i32 @llvm.hexagon.A2.sat(i64) #1
+
+attributes #0 = { nounwind "target-cpu"="hexagonv55" }
+attributes #1 = { nounwind readnone }
+
+!0 = !{!1, !1, i64 0}
+!1 = !{!"int", !2}
+!2 = !{!"omnipotent char", !3}
+!3 = !{!"Simple C/C++ TBAA"}
--- /dev/null
+; RUN: llc -march=hexagon -O3 -fp-contract=fast -pipeliner-max-stages=3 < %s
+; REQUIRES: asserts
+
+; Check Phis are generated correctly in epilogs after setting -swp-max-stages=3
+
+@g0 = private unnamed_addr constant [6 x i8] c"s4116\00", align 1
+
+; Function Attrs: noinline nounwind
+define void @f0(i32 %a0, i32 %a1, float* nocapture readonly %a2, [1000 x float]* nocapture readonly %a3, i32* nocapture readonly %a4, i32 %a5) #0 {
+b0:
+ %v0 = tail call i32 bitcast (i32 (...)* @f1 to i32 ()*)() #2
+ %v1 = sitofp i32 %v0 to double
+ %v2 = add nsw i32 %a1, -1
+ %v3 = icmp sgt i32 %a1, 1
+ br i1 %v3, label %b1, label %b3
+
+b1: ; preds = %b1, %b0
+ %v4 = phi float [ %v13, %b1 ], [ 0.000000e+00, %b0 ]
+ %v5 = phi float* [ %v16, %b1 ], [ %a2, %b0 ]
+ %v6 = phi i32* [ %v17, %b1 ], [ %a4, %b0 ]
+ %v7 = phi i32 [ %v14, %b1 ], [ 0, %b0 ]
+ %v8 = load float, float* %v5, align 4
+ %v9 = load i32, i32* %v6, align 4
+ %v10 = getelementptr inbounds [1000 x float], [1000 x float]* %a3, i32 %v9, i32 %a5
+ %v11 = load float, float* %v10, align 4
+ %v12 = fmul float %v8, %v11
+ %v13 = fadd float %v4, %v12
+ %v14 = add nuw nsw i32 %v7, 1
+ %v15 = icmp slt i32 %v14, %v2
+ %v16 = getelementptr float, float* %v5, i32 1
+ %v17 = getelementptr i32, i32* %v6, i32 1
+ br i1 %v15, label %b1, label %b2
+
+b2: ; preds = %b1
+ %v18 = fpext float %v13 to double
+ br label %b3
+
+b3: ; preds = %b2, %b0
+ %v19 = phi double [ %v18, %b2 ], [ 0.000000e+00, %b0 ]
+ tail call void @f2(double %v19, i32 %a1, i32 %a1, double %v1, i8* getelementptr inbounds ([6 x i8], [6 x i8]* @g0, i32 0, i32 0)) #2
+ ret void
+}
+
+declare i32 @f1(...) #1
+
+declare void @f2(double, i32, i32, double, i8*) #1
+
+attributes #0 = { noinline nounwind "target-cpu"="hexagonv60" }
+attributes #1 = { "target-cpu"="hexagonv60" }
+attributes #2 = { nounwind }
--- /dev/null
+; RUN: llc -march=hexagon < %s | FileCheck %s
+
+; Test that the pipeliner generates correct code when attempting to reuse
+; an existing phi. This test case contains a phi that references another
+; phi (the value from the previous iteration), and when there is a use that
+; is schedule in a later iteration. When this occurs, the pipeliner was
+; using a value from the wrong iteration.
+
+; CHECK: loop0(.LBB0_[[LOOP:.]],
+; CHECK: .LBB0_[[LOOP]]:
+; CHECK: vlalign([[VREG1:v([0-9]+)]],[[VREG2:v([0-9]+)]],#2)
+; CHECK: [[VREG2]]:{{[0-9]+}} = vcombine([[VREG1]],v{{[0-9]+}})
+; CHECK: }{{[ \t]*}}:endloop0
+
+; Function Attrs: nounwind
+define void @f0(i32 %a0, i32 %a1) local_unnamed_addr #0 {
+b0:
+ %v0 = shl nsw i32 %a0, 1
+ %v1 = sub i32 0, %v0
+ %v2 = sub i32 0, %a0
+ %v3 = getelementptr inbounds i8, i8* undef, i32 %v1
+ %v4 = getelementptr inbounds i8, i8* undef, i32 %v2
+ %v5 = getelementptr inbounds i8, i8* undef, i32 %a0
+ %v6 = getelementptr inbounds i8, i8* undef, i32 %v0
+ %v7 = getelementptr inbounds i8, i8* %v6, i32 64
+ %v8 = bitcast i8* %v7 to <16 x i32>*
+ %v9 = getelementptr inbounds i8, i8* %v5, i32 64
+ %v10 = bitcast i8* %v9 to <16 x i32>*
+ %v11 = getelementptr inbounds i8, i8* undef, i32 64
+ %v12 = bitcast i8* %v11 to <16 x i32>*
+ %v13 = getelementptr inbounds i8, i8* %v4, i32 64
+ %v14 = bitcast i8* %v13 to <16 x i32>*
+ %v15 = getelementptr inbounds i8, i8* %v3, i32 64
+ %v16 = bitcast i8* %v15 to <16 x i32>*
+ br label %b1
+
+b1: ; preds = %b1, %b0
+ %v17 = phi <16 x i32>* [ %v59, %b1 ], [ undef, %b0 ]
+ %v18 = phi <16 x i32>* [ %v34, %b1 ], [ %v8, %b0 ]
+ %v19 = phi <16 x i32>* [ %v32, %b1 ], [ %v10, %b0 ]
+ %v20 = phi <16 x i32>* [ %v30, %b1 ], [ %v12, %b0 ]
+ %v21 = phi <16 x i32>* [ %v28, %b1 ], [ %v14, %b0 ]
+ %v22 = phi <16 x i32>* [ %v26, %b1 ], [ %v16, %b0 ]
+ %v23 = phi <32 x i32> [ %v39, %b1 ], [ undef, %b0 ]
+ %v24 = phi <32 x i32> [ %v23, %b1 ], [ undef, %b0 ]
+ %v25 = phi i32 [ %v60, %b1 ], [ %a1, %b0 ]
+ %v26 = getelementptr inbounds <16 x i32>, <16 x i32>* %v22, i32 1
+ %v27 = load <16 x i32>, <16 x i32>* %v22, align 64
+ %v28 = getelementptr inbounds <16 x i32>, <16 x i32>* %v21, i32 1
+ %v29 = load <16 x i32>, <16 x i32>* %v21, align 64
+ %v30 = getelementptr inbounds <16 x i32>, <16 x i32>* %v20, i32 1
+ %v31 = load <16 x i32>, <16 x i32>* %v20, align 64
+ %v32 = getelementptr inbounds <16 x i32>, <16 x i32>* %v19, i32 1
+ %v33 = load <16 x i32>, <16 x i32>* %v19, align 64
+ %v34 = getelementptr inbounds <16 x i32>, <16 x i32>* %v18, i32 1
+ %v35 = load <16 x i32>, <16 x i32>* %v18, align 64
+ %v36 = tail call <32 x i32> @llvm.hexagon.V6.vaddubh(<16 x i32> %v27, <16 x i32> %v35) #2
+ %v37 = tail call <32 x i32> @llvm.hexagon.V6.vmpybus.acc(<32 x i32> %v36, <16 x i32> %v31, i32 101058054) #2
+ %v38 = tail call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %v33, <16 x i32> %v29) #2
+ %v39 = tail call <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32> %v37, <32 x i32> %v38, i32 67372036) #2
+ %v40 = tail call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %v23) #2
+ %v41 = tail call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %v24) #2
+ %v42 = tail call <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32> %v40, <16 x i32> %v41, i32 2) #2
+ %v43 = tail call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %v23) #2
+ %v44 = tail call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %v24) #2
+ %v45 = tail call <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32> %v43, <16 x i32> %v44, i32 2) #2
+ %v46 = tail call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %v39) #2
+ %v47 = tail call <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32> %v46, <16 x i32> %v40, i32 2) #2
+ %v48 = tail call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %v39) #2
+ %v49 = tail call <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32> %v48, <16 x i32> %v43, i32 2) #2
+ %v50 = tail call <16 x i32> @llvm.hexagon.V6.vaddh(<16 x i32> %v45, <16 x i32> %v43) #2
+ %v51 = tail call <16 x i32> @llvm.hexagon.V6.vaddh(<16 x i32> %v40, <16 x i32> %v47) #2
+ %v52 = tail call <16 x i32> @llvm.hexagon.V6.vaddh(<16 x i32> %v42, <16 x i32> %v47) #2
+ %v53 = tail call <16 x i32> @llvm.hexagon.V6.vmpyiwb.acc(<16 x i32> %v52, <16 x i32> %v40, i32 101058054) #2
+ %v54 = tail call <16 x i32> @llvm.hexagon.V6.vmpyiwb.acc(<16 x i32> %v53, <16 x i32> %v50, i32 67372036) #2
+ %v55 = tail call <16 x i32> @llvm.hexagon.V6.vaddh(<16 x i32> %v45, <16 x i32> %v49) #2
+ %v56 = tail call <16 x i32> @llvm.hexagon.V6.vmpyiwb.acc(<16 x i32> %v55, <16 x i32> %v43, i32 101058054) #2
+ %v57 = tail call <16 x i32> @llvm.hexagon.V6.vmpyiwb.acc(<16 x i32> %v56, <16 x i32> %v51, i32 67372036) #2
+ %v58 = tail call <16 x i32> @llvm.hexagon.V6.vshuffob(<16 x i32> %v57, <16 x i32> %v54) #2
+ %v59 = getelementptr inbounds <16 x i32>, <16 x i32>* %v17, i32 1
+ store <16 x i32> %v58, <16 x i32>* %v17, align 64
+ %v60 = add nsw i32 %v25, -64
+ %v61 = icmp sgt i32 %v25, 128
+ br i1 %v61, label %b1, label %b2
+
+b2: ; preds = %b1
+ ret void
+}
+
+; Function Attrs: nounwind readnone
+declare <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32>, <16 x i32>) #1
+
+declare <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32>, <32 x i32>, i32) #1
+declare <16 x i32> @llvm.hexagon.V6.hi(<32 x i32>) #1
+declare <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32>, <16 x i32>, i32) #1
+declare <16 x i32> @llvm.hexagon.V6.lo(<32 x i32>) #1
+declare <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32>, <16 x i32>, i32) #1
+declare <16 x i32> @llvm.hexagon.V6.vaddh(<16 x i32>, <16 x i32>) #1
+declare <32 x i32> @llvm.hexagon.V6.vaddubh(<16 x i32>, <16 x i32>) #1
+declare <32 x i32> @llvm.hexagon.V6.vmpybus.acc(<32 x i32>, <16 x i32>, i32) #1
+declare <16 x i32> @llvm.hexagon.V6.vmpyiwb.acc(<16 x i32>, <16 x i32>, i32) #1
+declare <16 x i32> @llvm.hexagon.V6.vshuffob(<16 x i32>, <16 x i32>) #1
+
+attributes #0 = { nounwind "target-cpu"="hexagonv65" "target-features"="+hvxv65,+hvx-length64b" }
+attributes #1 = { nounwind readnone }
+attributes #2 = { nounwind }
--- /dev/null
+; RUN: llc -march=hexagon -O2 < %s | FileCheck %s
+
+; We do not pipeline sigma yet, but the non-pipelined version
+; with good scheduling is pretty fast. The compiler generates
+; 19 packets, and the assembly version is 16.
+
+; CHECK: loop0(.LBB0_[[LOOP:.]],
+; CHECK: .LBB0_[[LOOP]]:
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK: }
+; CHECK-NOT: }
+; CHECK: }{{[ \t]*}}:endloop
+
+@g0 = external constant [10 x i16], align 128
+
+declare i32 @llvm.hexagon.S2.vsplatrb(i32) #0
+declare <16 x i32> @llvm.hexagon.V6.vshuffh(<16 x i32>) #0
+declare <16 x i32> @llvm.hexagon.V6.lvsplatw(i32) #0
+declare <16 x i32> @llvm.hexagon.V6.vd0() #0
+declare <32 x i32> @llvm.hexagon.V6.vsububh(<16 x i32>, <16 x i32>) #0
+declare <16 x i32> @llvm.hexagon.V6.vabsdiffub(<16 x i32>, <16 x i32>) #0
+declare <512 x i1> @llvm.hexagon.V6.vgtub(<16 x i32>, <16 x i32>) #0
+declare <16 x i32> @llvm.hexagon.V6.vmux(<512 x i1>, <16 x i32>, <16 x i32>) #0
+declare <16 x i32> @llvm.hexagon.V6.vaddbnq(<512 x i1>, <16 x i32>, <16 x i32>) #0
+declare <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32>, <16 x i32>) #0
+declare <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32>, <32 x i32>, i32) #0
+declare <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32>, <16 x i32>, i32) #0
+declare <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32>, <16 x i32>, i32) #0
+declare <32 x i32> @llvm.hexagon.V6.vlutvwh(<16 x i32>, <16 x i32>, i32) #0
+declare <16 x i32> @llvm.hexagon.V6.vmpyhvsrs(<16 x i32>, <16 x i32>) #0
+declare <16 x i32> @llvm.hexagon.V6.lo(<32 x i32>) #0
+declare <16 x i32> @llvm.hexagon.V6.hi(<32 x i32>) #0
+declare <16 x i32> @llvm.hexagon.V6.vsathub(<16 x i32>, <16 x i32>) #0
+
+define void @f0(i8* nocapture readonly %a0, i32 %a1, i32 %a2, i32 %a3, i8 zeroext %a4, i8* nocapture %a5) #1 {
+b0:
+ %v0 = add nsw i32 %a3, -1
+ %v1 = icmp sgt i32 %v0, 1
+ br i1 %v1, label %b1, label %b8
+
+b1: ; preds = %b0
+ %v2 = mul i32 %a1, 2
+ %v3 = load <16 x i32>, <16 x i32>* bitcast ([10 x i16]* @g0 to <16 x i32>*), align 128
+ %v4 = tail call <16 x i32> @llvm.hexagon.V6.vshuffh(<16 x i32> %v3) #2
+ %v5 = zext i8 %a4 to i32
+ %v6 = tail call i32 @llvm.hexagon.S2.vsplatrb(i32 %v5) #2
+ %v7 = tail call <16 x i32> @llvm.hexagon.V6.lvsplatw(i32 %v6) #2
+ %v8 = tail call <16 x i32> @llvm.hexagon.V6.vd0() #2
+ %v9 = tail call <16 x i32> @llvm.hexagon.V6.lvsplatw(i32 16843009) #2
+ %v10 = tail call <16 x i32> @llvm.hexagon.V6.lvsplatw(i32 33686018) #2
+ %v11 = icmp sgt i32 %a2, 64
+ %v12 = add i32 %a1, 64
+ %v13 = add i32 %v12, %a1
+ %v14 = icmp sgt i32 %a2, 0
+ %v15 = add i32 %a3, -2
+ %v16 = bitcast i8* %a0 to <16 x i32>*
+ %v17 = load <16 x i32>, <16 x i32>* %v16, align 64
+ br label %b2
+
+b2: ; preds = %b7, %b1
+ %v18 = phi <16 x i32> [ %v17, %b1 ], [ %v28, %b7 ]
+ %v19 = phi i8* [ %a0, %b1 ], [ %v23, %b7 ]
+ %v20 = phi i8* [ %a5, %b1 ], [ %v22, %b7 ]
+ %v21 = phi i32 [ 1, %b1 ], [ %v118, %b7 ]
+ %v22 = getelementptr inbounds i8, i8* %v20, i32 %a1
+ %v23 = getelementptr inbounds i8, i8* %v19, i32 %a1
+ %v24 = bitcast i8* %v23 to <16 x i32>*
+ %v25 = getelementptr inbounds i8, i8* %v19, i32 %v2
+ %v26 = bitcast i8* %v25 to <16 x i32>*
+ %v27 = bitcast i8* %v22 to <16 x i32>*
+ %v28 = load <16 x i32>, <16 x i32>* %v24, align 64
+ %v29 = load <16 x i32>, <16 x i32>* %v26, align 64
+ br i1 %v11, label %b3, label %b4
+
+b3: ; preds = %b2
+ %v30 = getelementptr inbounds i8, i8* %v19, i32 64
+ %v31 = getelementptr inbounds i8, i8* %v19, i32 %v12
+ %v32 = bitcast i8* %v31 to <16 x i32>*
+ %v33 = getelementptr inbounds i8, i8* %v19, i32 %v13
+ %v34 = bitcast i8* %v33 to <16 x i32>*
+ br label %b5
+
+b4: ; preds = %b2
+ br i1 %v14, label %b5, label %b7
+
+b5: ; preds = %b4, %b3
+ %v35 = phi <16 x i32>* [ %v26, %b4 ], [ %v34, %b3 ]
+ %v36 = phi <16 x i32>* [ %v24, %b4 ], [ %v32, %b3 ]
+ %v37 = phi i8* [ %v19, %b4 ], [ %v30, %b3 ]
+ %v38 = bitcast i8* %v37 to <16 x i32>*
+ br label %b6
+
+b6: ; preds = %b6, %b5
+ %v39 = phi <16 x i32>* [ %v108, %b6 ], [ %v27, %b5 ]
+ %v40 = phi <16 x i32>* [ %v115, %b6 ], [ %v35, %b5 ]
+ %v41 = phi <16 x i32>* [ %v114, %b6 ], [ %v36, %b5 ]
+ %v42 = phi <16 x i32>* [ %v113, %b6 ], [ %v38, %b5 ]
+ %v43 = phi i32 [ %v116, %b6 ], [ %a2, %b5 ]
+ %v44 = phi <16 x i32> [ %v45, %b6 ], [ %v8, %b5 ]
+ %v45 = phi <16 x i32> [ %v50, %b6 ], [ %v18, %b5 ]
+ %v46 = phi <16 x i32> [ %v47, %b6 ], [ %v8, %b5 ]
+ %v47 = phi <16 x i32> [ %v51, %b6 ], [ %v28, %b5 ]
+ %v48 = phi <16 x i32> [ %v49, %b6 ], [ %v8, %b5 ]
+ %v49 = phi <16 x i32> [ %v52, %b6 ], [ %v29, %b5 ]
+ %v50 = load <16 x i32>, <16 x i32>* %v42, align 64
+ %v51 = load <16 x i32>, <16 x i32>* %v41, align 64
+ %v52 = load <16 x i32>, <16 x i32>* %v40, align 64
+ %v53 = tail call <32 x i32> @llvm.hexagon.V6.vsububh(<16 x i32> %v8, <16 x i32> %v47) #2
+ %v54 = tail call <16 x i32> @llvm.hexagon.V6.vabsdiffub(<16 x i32> %v45, <16 x i32> %v47) #2
+ %v55 = tail call <16 x i32> @llvm.hexagon.V6.vabsdiffub(<16 x i32> %v49, <16 x i32> %v47) #2
+ %v56 = tail call <512 x i1> @llvm.hexagon.V6.vgtub(<16 x i32> %v54, <16 x i32> %v7) #2
+ %v57 = tail call <512 x i1> @llvm.hexagon.V6.vgtub(<16 x i32> %v55, <16 x i32> %v7) #2
+ %v58 = tail call <16 x i32> @llvm.hexagon.V6.vmux(<512 x i1> %v56, <16 x i32> %v9, <16 x i32> %v10) #2
+ %v59 = tail call <16 x i32> @llvm.hexagon.V6.vaddbnq(<512 x i1> %v57, <16 x i32> %v58, <16 x i32> %v9) #2
+ %v60 = tail call <16 x i32> @llvm.hexagon.V6.vmux(<512 x i1> %v56, <16 x i32> %v8, <16 x i32> %v45) #2
+ %v61 = tail call <16 x i32> @llvm.hexagon.V6.vmux(<512 x i1> %v57, <16 x i32> %v8, <16 x i32> %v49) #2
+ %v62 = tail call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %v61, <16 x i32> %v60) #2
+ %v63 = tail call <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32> %v53, <32 x i32> %v62, i32 -1) #2
+ %v64 = tail call <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32> %v45, <16 x i32> %v44, i32 1) #2
+ %v65 = tail call <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32> %v49, <16 x i32> %v48, i32 1) #2
+ %v66 = tail call <16 x i32> @llvm.hexagon.V6.vabsdiffub(<16 x i32> %v64, <16 x i32> %v47) #2
+ %v67 = tail call <16 x i32> @llvm.hexagon.V6.vabsdiffub(<16 x i32> %v65, <16 x i32> %v47) #2
+ %v68 = tail call <512 x i1> @llvm.hexagon.V6.vgtub(<16 x i32> %v66, <16 x i32> %v7) #2
+ %v69 = tail call <512 x i1> @llvm.hexagon.V6.vgtub(<16 x i32> %v67, <16 x i32> %v7) #2
+ %v70 = tail call <16 x i32> @llvm.hexagon.V6.vaddbnq(<512 x i1> %v68, <16 x i32> %v59, <16 x i32> %v9) #2
+ %v71 = tail call <16 x i32> @llvm.hexagon.V6.vaddbnq(<512 x i1> %v69, <16 x i32> %v70, <16 x i32> %v9) #2
+ %v72 = tail call <16 x i32> @llvm.hexagon.V6.vmux(<512 x i1> %v68, <16 x i32> %v8, <16 x i32> %v64) #2
+ %v73 = tail call <16 x i32> @llvm.hexagon.V6.vmux(<512 x i1> %v69, <16 x i32> %v8, <16 x i32> %v65) #2
+ %v74 = tail call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %v73, <16 x i32> %v72) #2
+ %v75 = tail call <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32> %v63, <32 x i32> %v74, i32 -1) #2
+ %v76 = tail call <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32> %v50, <16 x i32> %v45, i32 1) #2
+ %v77 = tail call <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32> %v52, <16 x i32> %v49, i32 1) #2
+ %v78 = tail call <16 x i32> @llvm.hexagon.V6.vabsdiffub(<16 x i32> %v76, <16 x i32> %v47) #2
+ %v79 = tail call <16 x i32> @llvm.hexagon.V6.vabsdiffub(<16 x i32> %v77, <16 x i32> %v47) #2
+ %v80 = tail call <512 x i1> @llvm.hexagon.V6.vgtub(<16 x i32> %v78, <16 x i32> %v7) #2
+ %v81 = tail call <512 x i1> @llvm.hexagon.V6.vgtub(<16 x i32> %v79, <16 x i32> %v7) #2
+ %v82 = tail call <16 x i32> @llvm.hexagon.V6.vaddbnq(<512 x i1> %v80, <16 x i32> %v71, <16 x i32> %v9) #2
+ %v83 = tail call <16 x i32> @llvm.hexagon.V6.vaddbnq(<512 x i1> %v81, <16 x i32> %v82, <16 x i32> %v9) #2
+ %v84 = tail call <16 x i32> @llvm.hexagon.V6.vmux(<512 x i1> %v80, <16 x i32> %v8, <16 x i32> %v76) #2
+ %v85 = tail call <16 x i32> @llvm.hexagon.V6.vmux(<512 x i1> %v81, <16 x i32> %v8, <16 x i32> %v77) #2
+ %v86 = tail call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %v85, <16 x i32> %v84) #2
+ %v87 = tail call <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32> %v75, <32 x i32> %v86, i32 -1) #2
+ %v88 = tail call <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32> %v47, <16 x i32> %v46, i32 1) #2
+ %v89 = tail call <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32> %v51, <16 x i32> %v47, i32 1) #2
+ %v90 = tail call <16 x i32> @llvm.hexagon.V6.vabsdiffub(<16 x i32> %v88, <16 x i32> %v47) #2
+ %v91 = tail call <16 x i32> @llvm.hexagon.V6.vabsdiffub(<16 x i32> %v89, <16 x i32> %v47) #2
+ %v92 = tail call <512 x i1> @llvm.hexagon.V6.vgtub(<16 x i32> %v90, <16 x i32> %v7) #2
+ %v93 = tail call <512 x i1> @llvm.hexagon.V6.vgtub(<16 x i32> %v91, <16 x i32> %v7) #2
+ %v94 = tail call <16 x i32> @llvm.hexagon.V6.vaddbnq(<512 x i1> %v92, <16 x i32> %v83, <16 x i32> %v9) #2
+ %v95 = tail call <16 x i32> @llvm.hexagon.V6.vaddbnq(<512 x i1> %v93, <16 x i32> %v94, <16 x i32> %v9) #2
+ %v96 = tail call <16 x i32> @llvm.hexagon.V6.vmux(<512 x i1> %v92, <16 x i32> %v8, <16 x i32> %v88) #2
+ %v97 = tail call <16 x i32> @llvm.hexagon.V6.vmux(<512 x i1> %v93, <16 x i32> %v8, <16 x i32> %v89) #2
+ %v98 = tail call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %v97, <16 x i32> %v96) #2
+ %v99 = tail call <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32> %v87, <32 x i32> %v98, i32 -1) #2
+ %v100 = tail call <32 x i32> @llvm.hexagon.V6.vlutvwh(<16 x i32> %v95, <16 x i32> %v4, i32 0) #2
+ %v101 = tail call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %v99) #2
+ %v102 = tail call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %v100) #2
+ %v103 = tail call <16 x i32> @llvm.hexagon.V6.vmpyhvsrs(<16 x i32> %v101, <16 x i32> %v102) #2
+ %v104 = tail call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %v99) #2
+ %v105 = tail call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %v100) #2
+ %v106 = tail call <16 x i32> @llvm.hexagon.V6.vmpyhvsrs(<16 x i32> %v104, <16 x i32> %v105) #2
+ %v107 = tail call <16 x i32> @llvm.hexagon.V6.vsathub(<16 x i32> %v106, <16 x i32> %v103) #2
+ %v108 = getelementptr inbounds <16 x i32>, <16 x i32>* %v39, i32 1
+ store <16 x i32> %v107, <16 x i32>* %v39, align 64
+ %v109 = icmp sgt i32 %v43, 128
+ %v110 = getelementptr inbounds <16 x i32>, <16 x i32>* %v42, i32 1
+ %v111 = getelementptr inbounds <16 x i32>, <16 x i32>* %v41, i32 1
+ %v112 = getelementptr inbounds <16 x i32>, <16 x i32>* %v40, i32 1
+ %v113 = select i1 %v109, <16 x i32>* %v110, <16 x i32>* %v42
+ %v114 = select i1 %v109, <16 x i32>* %v111, <16 x i32>* %v41
+ %v115 = select i1 %v109, <16 x i32>* %v112, <16 x i32>* %v40
+ %v116 = add nsw i32 %v43, -64
+ %v117 = icmp sgt i32 %v43, 64
+ br i1 %v117, label %b6, label %b7
+
+b7: ; preds = %b6, %b4
+ %v118 = add nuw nsw i32 %v21, 1
+ %v119 = icmp eq i32 %v21, %v15
+ br i1 %v119, label %b8, label %b2
+
+b8: ; preds = %b7, %b0
+ ret void
+}
+
+attributes #0 = { nounwind readnone }
+attributes #1 = { nounwind "target-cpu"="hexagonv60" "target-features"="+hvx,+hvx-length64b" }
+attributes #2 = { nounwind }
--- /dev/null
+; RUN: llc -march=hexagon -O3 -hexagon-eif=0 < %s | FileCheck %s
+; Without TFR cleanup, the last block contained
+; {
+; r3 = xor(r1, r2)
+; r1 = #0
+; }
+; {
+; r7 = r1
+; r0 = zxtb(r3)
+; }
+; After TFR cleanup, the copy "r7 = r1" should be simplified to "r7 = #0".
+; There shouldn't be any register copies in that block anymore.
+;
+; CHECK: LBB0_5:
+; CHECK-NOT: r{{[0-9]+}} = r{{[0-9]+}}
+
+target triple = "hexagon"
+
+; Function Attrs: nounwind readnone
+define i64 @f0(i64 %a0, i32 %a1, i32 %a2) #0 {
+b0:
+ %v0 = trunc i64 %a0 to i32
+ %v1 = lshr i64 %a0, 32
+ %v2 = trunc i64 %v1 to i32
+ %v3 = lshr i64 %a0, 40
+ %v4 = lshr i64 %a0, 48
+ %v5 = trunc i64 %v4 to i16
+ %v6 = icmp sgt i32 %a2, %a1
+ %v7 = lshr i32 %v0, 10
+ br i1 %v6, label %b1, label %b2
+
+b1: ; preds = %b0
+ %v8 = add nsw i32 %v7, 4190971
+ %v9 = and i32 %v8, 4194303
+ %v10 = shl nuw nsw i64 %v3, 24
+ %v11 = trunc i64 %v10 to i32
+ %v12 = ashr exact i32 %v11, 24
+ %v13 = or i32 %v12, 102
+ br label %b3
+
+b2: ; preds = %b0
+ %v14 = add nsw i32 %v7, 4189760
+ %v15 = trunc i64 %v3 to i32
+ %v16 = or i32 %v15, 119
+ br label %b3
+
+b3: ; preds = %b2, %b1
+ %v17 = phi i32 [ %v13, %b1 ], [ %v16, %b2 ]
+ %v18 = phi i32 [ %v9, %b1 ], [ %v14, %b2 ]
+ %v19 = shl i32 %v18, 10
+ %v20 = icmp sgt i32 %a1, %a2
+ br i1 %v20, label %b4, label %b5
+
+b4: ; preds = %b3
+ %v21 = and i32 %v2, 140
+ %v22 = or i32 %v21, 115
+ %v23 = and i16 %v5, 12345
+ br label %b6
+
+b5: ; preds = %b3
+ %v24 = xor i32 %v2, 23
+ %v25 = or i16 %v5, 12345
+ br label %b6
+
+b6: ; preds = %b5, %b4
+ %v26 = phi i16 [ %v23, %b4 ], [ %v25, %b5 ]
+ %v27 = phi i32 [ %v22, %b4 ], [ %v24, %b5 ]
+ %v28 = zext i16 %v26 to i64
+ %v29 = shl nuw i64 %v28, 48
+ %v30 = and i32 %v27, 255
+ %v31 = zext i32 %v30 to i64
+ %v32 = shl nuw nsw i64 %v31, 40
+ %v33 = and i32 %v17, 255
+ %v34 = zext i32 %v33 to i64
+ %v35 = shl nuw nsw i64 %v34, 32
+ %v36 = zext i32 %v19 to i64
+ %v37 = or i64 %v36, %v35
+ %v38 = or i64 %v37, %v29
+ %v39 = or i64 %v38, %v32
+ ret i64 %v39
+}
+
+attributes #0 = { nounwind readnone "target-cpu"="hexagonv55" }
--- /dev/null
+; RUN: llc -march=hexagon -disable-hexagon-shuffle=1 -O2 -enable-pipeliner=false < %s | FileCheck %s
+
+; Generate vmemu (unaligned).
+; CHECK: vmem
+; CHECK: vmem
+; CHECK: vmemu
+; CHECK-NOT: vmem
+
+target triple = "hexagon"
+
+; Function Attrs: nounwind
+define void @f0(i16* nocapture readonly %a0, i32 %a1, i32 %a2, i16* nocapture %a3) #0 {
+b0:
+ %v0 = mul i32 %a2, -2
+ %v1 = add i32 %v0, 64
+ %v2 = tail call <16 x i32> @llvm.hexagon.V6.vsubw(<16 x i32> undef, <16 x i32> undef)
+ %v3 = bitcast i16* %a3 to <16 x i32>*
+ %v4 = sdiv i32 %a1, 32
+ %v5 = icmp sgt i32 %a1, 31
+ br i1 %v5, label %b1, label %b4
+
+b1: ; preds = %b0
+ %v6 = bitcast i16* %a0 to <16 x i32>*
+ %v7 = icmp sgt i32 %a1, 63
+ %v8 = mul i32 %v4, 32
+ %v9 = select i1 %v7, i32 %v8, i32 32
+ %v10 = getelementptr i16, i16* %a3, i32 %v9
+ br label %b2
+
+b2: ; preds = %b2, %b1
+ %v11 = phi i32 [ 0, %b1 ], [ %v19, %b2 ]
+ %v12 = phi <16 x i32> [ %v2, %b1 ], [ %v16, %b2 ]
+ %v13 = phi <16 x i32>* [ %v3, %b1 ], [ %v18, %b2 ]
+ %v14 = phi <16 x i32>* [ %v6, %b1 ], [ %v15, %b2 ]
+ %v15 = getelementptr inbounds <16 x i32>, <16 x i32>* %v14, i32 1
+ %v16 = load <16 x i32>, <16 x i32>* %v14, align 4, !tbaa !0
+ %v17 = tail call <16 x i32> @llvm.hexagon.V6.valignb(<16 x i32> %v16, <16 x i32> %v12, i32 %v1)
+ %v18 = getelementptr inbounds <16 x i32>, <16 x i32>* %v13, i32 1
+ store <16 x i32> %v17, <16 x i32>* %v13, align 4, !tbaa !0
+ %v19 = add nsw i32 %v11, 1
+ %v20 = icmp slt i32 %v19, %v4
+ br i1 %v20, label %b2, label %b3
+
+b3: ; preds = %b2
+ %v21 = bitcast i16* %v10 to <16 x i32>*
+ br label %b4
+
+b4: ; preds = %b3, %b0
+ %v22 = phi <16 x i32> [ %v16, %b3 ], [ %v2, %b0 ]
+ %v23 = phi <16 x i32>* [ %v21, %b3 ], [ %v3, %b0 ]
+ %v24 = tail call <16 x i32> @llvm.hexagon.V6.valignb(<16 x i32> %v2, <16 x i32> %v22, i32 %v1)
+ store <16 x i32> %v24, <16 x i32>* %v23, align 4, !tbaa !0
+ ret void
+}
+
+; Function Attrs: nounwind readnone
+declare <16 x i32> @llvm.hexagon.V6.vsubw(<16 x i32>, <16 x i32>) #1
+
+; Function Attrs: nounwind readnone
+declare <16 x i32> @llvm.hexagon.V6.valignb(<16 x i32>, <16 x i32>, i32) #1
+
+attributes #0 = { nounwind "target-cpu"="hexagonv60" "target-features"="+hvxv60,+hvx-length64b" }
+attributes #1 = { nounwind readnone }
+
+!0 = !{!1, !1, i64 0}
+!1 = !{!"omnipotent char", !2, i64 0}
+!2 = !{!"Simple C/C++ TBAA"}
--- /dev/null
+; RUN: llc -march=hexagon -disable-hexagon-shuffle=0 -O2 < %s | FileCheck %s
+
+; Generate vmemu (unaligned).
+; CHECK: vmemu
+; CHECK: vmemu
+; CHECK: vmemu
+; CHECK-NOT: vmem
+
+target triple = "hexagon"
+
+; Function Attrs: nounwind
+define void @f0(i8* nocapture readonly %a0, i8* nocapture readonly %a1, i8* nocapture %a2) #0 {
+b0:
+ %v0 = bitcast i8* %a0 to <16 x i32>*
+ %v1 = load <16 x i32>, <16 x i32>* %v0, align 4, !tbaa !0
+ %v2 = bitcast i8* %a1 to <16 x i32>*
+ %v3 = load <16 x i32>, <16 x i32>* %v2, align 4, !tbaa !0
+ %v4 = tail call <16 x i32> @llvm.hexagon.V6.vaddw(<16 x i32> %v1, <16 x i32> %v3)
+ %v5 = bitcast i8* %a2 to <16 x i32>*
+ store <16 x i32> %v4, <16 x i32>* %v5, align 4, !tbaa !0
+ ret void
+}
+
+; Function Attrs: nounwind readnone
+declare <16 x i32> @llvm.hexagon.V6.vaddw(<16 x i32>, <16 x i32>) #1
+
+attributes #0 = { nounwind "target-cpu"="hexagonv60" "target-features"="+hvxv60,+hvx-length64b" }
+attributes #1 = { nounwind readnone }
+
+!0 = !{!1, !1, i64 0}
+!1 = !{!"omnipotent char", !2, i64 0}
+!2 = !{!"Simple C/C++ TBAA"}
--- /dev/null
+; RUN: llc -march=hexagon -O2 -enable-pipeliner=false < %s | FileCheck %s
+; RUN: llc -march=hexagon -O2 -debug-only=pipeliner < %s -o - 2>&1 | FileCheck %s --check-prefix=CHECK-SWP
+
+; CHECK: {
+; CHECK-DAG: v{{[0-9]*}} = vmem(r{{[0-9]*}}++#1)
+; CHECK-DAG: vmem(r{{[0-9]*}}++#1) = v{{[0-9]*}}.new
+; CHECK: }{{[ \t]*}}:endloop0
+
+; CHECK-SWP: Schedule Found? 1
+; CHECK-SWP: {
+; CHECK-DAG-SWP: v{{[0-9]*}}.cur = vmem(r{{[0-9]*}}++#1)
+; CHECK-DAG-SWP: vmem(r{{[0-9]*}}++#1) = v{{[0-9]*}}.new
+; CHECK-SWP: }{{[ \t]*}}:endloop0
+
+target triple = "hexagon"
+
+; Function Attrs: nounwind
+define void @f0(i16* nocapture readonly %a0, i32 %a1, i32 %a2, i16* nocapture %a3) #0 {
+b0:
+ %v0 = mul i32 %a2, -2
+ %v1 = add i32 %v0, 64
+ %v2 = tail call <16 x i32> @llvm.hexagon.V6.vsubw(<16 x i32> undef, <16 x i32> undef)
+ %v3 = bitcast i16* %a3 to <16 x i32>*
+ %v4 = sdiv i32 %a1, 32
+ %v5 = icmp sgt i32 %a1, 31
+ br i1 %v5, label %b1, label %b4
+
+b1: ; preds = %b0
+ %v6 = bitcast i16* %a0 to <16 x i32>*
+ %v7 = icmp sgt i32 %a1, 63
+ %v8 = mul i32 %v4, 32
+ %v9 = select i1 %v7, i32 %v8, i32 32
+ %v10 = getelementptr i16, i16* %a3, i32 %v9
+ br label %b2
+
+b2: ; preds = %b2, %b1
+ %v11 = phi i32 [ 0, %b1 ], [ %v19, %b2 ]
+ %v12 = phi <16 x i32> [ %v2, %b1 ], [ %v16, %b2 ]
+ %v13 = phi <16 x i32>* [ %v3, %b1 ], [ %v18, %b2 ]
+ %v14 = phi <16 x i32>* [ %v6, %b1 ], [ %v15, %b2 ]
+ %v15 = getelementptr inbounds <16 x i32>, <16 x i32>* %v14, i32 1
+ %v16 = load <16 x i32>, <16 x i32>* %v14, align 64, !tbaa !0
+ %v17 = tail call <16 x i32> @llvm.hexagon.V6.valignb(<16 x i32> %v16, <16 x i32> %v12, i32 %v1)
+ %v18 = getelementptr inbounds <16 x i32>, <16 x i32>* %v13, i32 1
+ store <16 x i32> %v17, <16 x i32>* %v13, align 64, !tbaa !0
+ %v19 = add nsw i32 %v11, 1
+ %v20 = icmp slt i32 %v19, %v4
+ br i1 %v20, label %b2, label %b3
+
+b3: ; preds = %b2
+ %v21 = bitcast i16* %v10 to <16 x i32>*
+ br label %b4
+
+b4: ; preds = %b3, %b0
+ %v22 = phi <16 x i32> [ %v16, %b3 ], [ %v2, %b0 ]
+ %v23 = phi <16 x i32>* [ %v21, %b3 ], [ %v3, %b0 ]
+ %v24 = tail call <16 x i32> @llvm.hexagon.V6.valignb(<16 x i32> %v2, <16 x i32> %v22, i32 %v1)
+ store <16 x i32> %v24, <16 x i32>* %v23, align 64, !tbaa !0
+ ret void
+}
+
+; Function Attrs: nounwind readnone
+declare <16 x i32> @llvm.hexagon.V6.vsubw(<16 x i32>, <16 x i32>) #1
+
+; Function Attrs: nounwind readnone
+declare <16 x i32> @llvm.hexagon.V6.valignb(<16 x i32>, <16 x i32>, i32) #1
+
+attributes #0 = { nounwind "target-cpu"="hexagonv60" "target-features"="+hvxv60,+hvx-length64b" }
+attributes #1 = { nounwind readnone }
+
+!0 = !{!1, !1, i64 0}
+!1 = !{!"omnipotent char", !2, i64 0}
+!2 = !{!"Simple C/C++ TBAA"}