228 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=riscv64 -mattr=+v | FileCheck %s3 4 5; Make sure we don't emit a pair of shift for the zext in the preheader. We6; can tell that bit 31 is 0 in the preheader and rely on %n already being7; sign extended without adding zeros explicitly.8define void @test1(ptr nocapture noundef %a, i32 noundef signext %n) {9; CHECK-LABEL: test1:10; CHECK: # %bb.0: # %entry11; CHECK-NEXT: blez a1, .LBB0_312; CHECK-NEXT: # %bb.1: # %for.body.preheader13; CHECK-NEXT: slli a1, a1, 214; CHECK-NEXT: add a1, a0, a115; CHECK-NEXT: .LBB0_2: # %for.body16; CHECK-NEXT: # =>This Inner Loop Header: Depth=117; CHECK-NEXT: lw a2, 0(a0)18; CHECK-NEXT: addi a2, a2, 419; CHECK-NEXT: sw a2, 0(a0)20; CHECK-NEXT: addi a0, a0, 421; CHECK-NEXT: bne a0, a1, .LBB0_222; CHECK-NEXT: .LBB0_3: # %for.cond.cleanup23; CHECK-NEXT: ret24entry:25 %cmp3 = icmp sgt i32 %n, 026 br i1 %cmp3, label %for.body.preheader, label %for.cond.cleanup27 28for.body.preheader: ; preds = %entry29 %wide.trip.count = zext nneg i32 %n to i6430 br label %for.body31 32for.cond.cleanup: ; preds = %for.body, %entry33 ret void34 35for.body: ; preds = %for.body.preheader, %for.body36 %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next, %for.body ]37 %arrayidx = getelementptr inbounds i32, ptr %a, i64 %indvars.iv38 %0 = load i32, ptr %arrayidx, align 439 %add = add nsw i32 %0, 440 store i32 %add, ptr %arrayidx, align 441 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 142 %exitcond.not = icmp eq i64 %indvars.iv.next, %wide.trip.count43 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body44}45 46; Make sure we convert the 4294967294 in for.body.preheader.new to -2 based on47; the upper 33 bits being zero by the dominating condition %cmp3.48define void @test2(ptr nocapture noundef %a, i32 noundef signext %n) {49; CHECK-LABEL: test2:50; CHECK: # %bb.0: # %entry51; CHECK-NEXT: blez a1, .LBB1_752; CHECK-NEXT: # %bb.1: # %for.body.preheader53; CHECK-NEXT: li a3, 154; CHECK-NEXT: andi a2, a1, 155; CHECK-NEXT: bne a1, a3, .LBB1_356; CHECK-NEXT: # %bb.2:57; CHECK-NEXT: li a3, 058; CHECK-NEXT: j .LBB1_559; CHECK-NEXT: .LBB1_3: # %for.body.preheader.new60; CHECK-NEXT: li a3, 061; CHECK-NEXT: andi a1, a1, -262; CHECK-NEXT: addi a4, a0, 463; CHECK-NEXT: .LBB1_4: # %for.body64; CHECK-NEXT: # =>This Inner Loop Header: Depth=165; CHECK-NEXT: lw a5, -4(a4)66; CHECK-NEXT: lw a6, 0(a4)67; CHECK-NEXT: addi a3, a3, 268; CHECK-NEXT: addi a5, a5, 469; CHECK-NEXT: addi a6, a6, 470; CHECK-NEXT: sw a5, -4(a4)71; CHECK-NEXT: sw a6, 0(a4)72; CHECK-NEXT: addi a4, a4, 873; CHECK-NEXT: bne a1, a3, .LBB1_474; CHECK-NEXT: .LBB1_5: # %for.cond.cleanup.loopexit.unr-lcssa75; CHECK-NEXT: beqz a2, .LBB1_776; CHECK-NEXT: # %bb.6: # %for.body.epil77; CHECK-NEXT: slli a3, a3, 278; CHECK-NEXT: add a0, a0, a379; CHECK-NEXT: lw a1, 0(a0)80; CHECK-NEXT: addi a1, a1, 481; CHECK-NEXT: sw a1, 0(a0)82; CHECK-NEXT: .LBB1_7: # %for.cond.cleanup83; CHECK-NEXT: ret84entry:85 %cmp3 = icmp sgt i32 %n, 086 br i1 %cmp3, label %for.body.preheader, label %for.cond.cleanup87 88for.body.preheader: ; preds = %entry89 %wide.trip.count = zext nneg i32 %n to i6490 %xtraiter = and i64 %wide.trip.count, 191 %0 = icmp eq i32 %n, 192 br i1 %0, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body.preheader.new93 94for.body.preheader.new: ; preds = %for.body.preheader95 %unroll_iter = and i64 %wide.trip.count, 429496729496 br label %for.body97 98for.cond.cleanup.loopexit.unr-lcssa: ; preds = %for.body, %for.body.preheader99 %indvars.iv.unr = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next.1, %for.body ]100 %lcmp.mod.not = icmp eq i64 %xtraiter, 0101 br i1 %lcmp.mod.not, label %for.cond.cleanup, label %for.body.epil102 103for.body.epil: ; preds = %for.cond.cleanup.loopexit.unr-lcssa104 %arrayidx.epil = getelementptr inbounds i32, ptr %a, i64 %indvars.iv.unr105 %1 = load i32, ptr %arrayidx.epil, align 4106 %add.epil = add nsw i32 %1, 4107 store i32 %add.epil, ptr %arrayidx.epil, align 4108 br label %for.cond.cleanup109 110for.cond.cleanup: ; preds = %for.body.epil, %for.cond.cleanup.loopexit.unr-lcssa, %entry111 ret void112 113for.body: ; preds = %for.body, %for.body.preheader.new114 %indvars.iv = phi i64 [ 0, %for.body.preheader.new ], [ %indvars.iv.next.1, %for.body ]115 %niter = phi i64 [ 0, %for.body.preheader.new ], [ %niter.next.1, %for.body ]116 %arrayidx = getelementptr inbounds i32, ptr %a, i64 %indvars.iv117 %2 = load i32, ptr %arrayidx, align 4118 %add = add nsw i32 %2, 4119 store i32 %add, ptr %arrayidx, align 4120 %indvars.iv.next = or disjoint i64 %indvars.iv, 1121 %arrayidx.1 = getelementptr inbounds i32, ptr %a, i64 %indvars.iv.next122 %3 = load i32, ptr %arrayidx.1, align 4123 %add.1 = add nsw i32 %3, 4124 store i32 %add.1, ptr %arrayidx.1, align 4125 %indvars.iv.next.1 = add nuw nsw i64 %indvars.iv, 2126 %niter.next.1 = add i64 %niter, 2127 %niter.ncmp.1 = icmp eq i64 %niter.next.1, %unroll_iter128 br i1 %niter.ncmp.1, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body129}130 131define i1 @widen_anyof_rdx(ptr %p, i64 %n) {132; CHECK-LABEL: widen_anyof_rdx:133; CHECK: # %bb.0: # %entry134; CHECK-NEXT: li a2, 0135; CHECK-NEXT: vsetvli a3, zero, e8, mf2, ta, ma136; CHECK-NEXT: vmv.v.i v8, 0137; CHECK-NEXT: .LBB2_1: # %loop138; CHECK-NEXT: # =>This Inner Loop Header: Depth=1139; CHECK-NEXT: sub a3, a1, a2140; CHECK-NEXT: slli a4, a2, 2141; CHECK-NEXT: vsetvli a3, a3, e32, m2, ta, ma142; CHECK-NEXT: add a4, a0, a4143; CHECK-NEXT: vle32.v v10, (a4)144; CHECK-NEXT: vmsne.vi v0, v10, 0145; CHECK-NEXT: add a2, a2, a3146; CHECK-NEXT: vsetvli zero, zero, e8, mf2, tu, ma147; CHECK-NEXT: vmerge.vim v8, v8, 1, v0148; CHECK-NEXT: blt a2, a1, .LBB2_1149; CHECK-NEXT: # %bb.2: # %exit150; CHECK-NEXT: vsetvli a0, zero, e8, mf2, ta, ma151; CHECK-NEXT: vand.vi v8, v8, 1152; CHECK-NEXT: vmsne.vi v8, v8, 0153; CHECK-NEXT: vcpop.m a0, v8154; CHECK-NEXT: snez a0, a0155; CHECK-NEXT: ret156entry:157 br label %loop158loop:159 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]160 %phi = phi <vscale x 4 x i1> [ zeroinitializer, %entry ], [ %rec, %loop ]161 %avl = sub i64 %n, %iv162 %evl = call i32 @llvm.experimental.get.vector.length(i64 %avl, i32 4, i1 true)163 164 %gep = getelementptr i32, ptr %p, i64 %iv165 %x = call <vscale x 4 x i32> @llvm.vp.load(ptr %gep, <vscale x 4 x i1> splat (i1 true), i32 %evl)166 %cmp = icmp ne <vscale x 4 x i32> %x, zeroinitializer167 %rec = call <vscale x 4 x i1> @llvm.vp.merge(<vscale x 4 x i1> %cmp, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i1> %phi, i32 %evl)168 169 %evl.zext = zext i32 %evl to i64170 %iv.next = add i64 %iv, %evl.zext171 %done = icmp sge i64 %iv.next, %n172 br i1 %done, label %exit, label %loop173exit:174 %res = call i1 @llvm.vector.reduce.or(<vscale x 4 x i1> %rec)175 ret i1 %res176}177 178 179define i1 @widen_anyof_rdx_use_in_loop(ptr %p, i64 %n) {180; CHECK-LABEL: widen_anyof_rdx_use_in_loop:181; CHECK: # %bb.0: # %entry182; CHECK-NEXT: li a2, 0183; CHECK-NEXT: vsetvli a3, zero, e8, mf2, ta, ma184; CHECK-NEXT: vmv.v.i v8, 0185; CHECK-NEXT: .LBB3_1: # %loop186; CHECK-NEXT: # =>This Inner Loop Header: Depth=1187; CHECK-NEXT: sub a3, a1, a2188; CHECK-NEXT: slli a4, a2, 2189; CHECK-NEXT: vsetvli a3, a3, e32, m2, ta, ma190; CHECK-NEXT: add a4, a0, a4191; CHECK-NEXT: vle32.v v10, (a4)192; CHECK-NEXT: vmsne.vi v0, v10, 0193; CHECK-NEXT: vsetvli zero, zero, e8, mf2, tu, ma194; CHECK-NEXT: vmerge.vim v8, v8, 1, v0195; CHECK-NEXT: vsetvli a5, zero, e8, mf2, ta, ma196; CHECK-NEXT: vand.vi v9, v8, 1197; CHECK-NEXT: vmsne.vi v9, v9, 0198; CHECK-NEXT: add a2, a2, a3199; CHECK-NEXT: vsm.v v9, (a4)200; CHECK-NEXT: blt a2, a1, .LBB3_1201; CHECK-NEXT: # %bb.2: # %exit202; CHECK-NEXT: vcpop.m a0, v9203; CHECK-NEXT: snez a0, a0204; CHECK-NEXT: ret205entry:206 br label %loop207loop:208 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]209 %phi = phi <vscale x 4 x i1> [ zeroinitializer, %entry ], [ %rec, %loop ]210 %avl = sub i64 %n, %iv211 %evl = call i32 @llvm.experimental.get.vector.length(i64 %avl, i32 4, i1 true)212 213 %gep = getelementptr i32, ptr %p, i64 %iv214 %x = call <vscale x 4 x i32> @llvm.vp.load(ptr %gep, <vscale x 4 x i1> splat (i1 true), i32 %evl)215 %cmp = icmp ne <vscale x 4 x i32> %x, zeroinitializer216 %rec = call <vscale x 4 x i1> @llvm.vp.merge(<vscale x 4 x i1> %cmp, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i1> %phi, i32 %evl)217 218 store <vscale x 4 x i1> %rec, ptr %gep219 220 %evl.zext = zext i32 %evl to i64221 %iv.next = add i64 %iv, %evl.zext222 %done = icmp sge i64 %iv.next, %n223 br i1 %done, label %exit, label %loop224exit:225 %res = call i1 @llvm.vector.reduce.or(<vscale x 4 x i1> %rec)226 ret i1 %res227}228