108 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=aarch64-unknown-unknown | FileCheck %s3 4; Test LSR for giving small constants, which get re-associated as unfolded5; offset, a chance to get combined with loop-invariant registers (same as6; large constants which do not fit as add immediate operands). LSR7; favors here to bump the base pointer outside the loop.8 9; float test(float *arr, long long start, float threshold) {10; for (long long i = start; i != 0; ++i) {11; float x = arr[i + 7];12; if (x > threshold)13; return x;14; }15; return -7;16; }17define float @test1(ptr nocapture readonly %arr, i64 %start, float %threshold) {18; CHECK-LABEL: test1:19; CHECK: // %bb.0: // %entry20; CHECK-NEXT: cbz x1, .LBB0_421; CHECK-NEXT: // %bb.1: // %for.body.preheader22; CHECK-NEXT: add x8, x0, #2823; CHECK-NEXT: .LBB0_2: // %for.body24; CHECK-NEXT: // =>This Inner Loop Header: Depth=125; CHECK-NEXT: ldr s1, [x8, x1, lsl #2]26; CHECK-NEXT: fcmp s1, s027; CHECK-NEXT: b.gt .LBB0_528; CHECK-NEXT: // %bb.3: // %for.cond29; CHECK-NEXT: // in Loop: Header=BB0_2 Depth=130; CHECK-NEXT: add x1, x1, #131; CHECK-NEXT: cbnz x1, .LBB0_232; CHECK-NEXT: .LBB0_4:33; CHECK-NEXT: fmov s0, #-7.0000000034; CHECK-NEXT: ret35; CHECK-NEXT: .LBB0_5: // %cleanup236; CHECK-NEXT: fmov s0, s137; CHECK-NEXT: ret38entry:39 %cmp11 = icmp eq i64 %start, 040 br i1 %cmp11, label %cleanup2, label %for.body41 42for.cond: ; preds = %for.body43 %cmp = icmp eq i64 %inc, 044 br i1 %cmp, label %cleanup2, label %for.body45 46for.body: ; preds = %entry, %for.cond47 %i.012 = phi i64 [ %inc, %for.cond ], [ %start, %entry ]48 %add = add nsw i64 %i.012, 749 %arrayidx = getelementptr inbounds float, ptr %arr, i64 %add50 %0 = load float, ptr %arrayidx, align 451 %cmp1 = fcmp ogt float %0, %threshold52 %inc = add nsw i64 %i.012, 153 br i1 %cmp1, label %cleanup2, label %for.cond54 55cleanup2: ; preds = %for.cond, %for.body, %entry56 %1 = phi float [ -7.000000e+00, %entry ], [ %0, %for.body ], [ -7.000000e+00, %for.cond ]57 ret float %158}59 60; Same as test1, except i has another use:61; if (x > threshold) ---> if (x > threshold + i)62define float @test2(ptr nocapture readonly %arr, i64 %start, float %threshold) {63; CHECK-LABEL: test2:64; CHECK: // %bb.0: // %entry65; CHECK-NEXT: cbz x1, .LBB1_466; CHECK-NEXT: // %bb.1: // %for.body.preheader67; CHECK-NEXT: add x8, x0, #2868; CHECK-NEXT: .LBB1_2: // %for.body69; CHECK-NEXT: // =>This Inner Loop Header: Depth=170; CHECK-NEXT: scvtf s1, x171; CHECK-NEXT: fadd s2, s1, s072; CHECK-NEXT: ldr s1, [x8, x1, lsl #2]73; CHECK-NEXT: fcmp s1, s274; CHECK-NEXT: b.gt .LBB1_575; CHECK-NEXT: // %bb.3: // %for.cond76; CHECK-NEXT: // in Loop: Header=BB1_2 Depth=177; CHECK-NEXT: add x1, x1, #178; CHECK-NEXT: cbnz x1, .LBB1_279; CHECK-NEXT: .LBB1_4:80; CHECK-NEXT: fmov s0, #-7.0000000081; CHECK-NEXT: ret82; CHECK-NEXT: .LBB1_5: // %cleanup483; CHECK-NEXT: fmov s0, s184; CHECK-NEXT: ret85entry:86 %cmp14 = icmp eq i64 %start, 087 br i1 %cmp14, label %cleanup4, label %for.body88 89for.cond: ; preds = %for.body90 %cmp = icmp eq i64 %inc, 091 br i1 %cmp, label %cleanup4, label %for.body92 93for.body: ; preds = %entry, %for.cond94 %i.015 = phi i64 [ %inc, %for.cond ], [ %start, %entry ]95 %add = add nsw i64 %i.015, 796 %arrayidx = getelementptr inbounds float, ptr %arr, i64 %add97 %0 = load float, ptr %arrayidx, align 498 %conv = sitofp i64 %i.015 to float99 %add1 = fadd float %conv, %threshold100 %cmp2 = fcmp ogt float %0, %add1101 %inc = add nsw i64 %i.015, 1102 br i1 %cmp2, label %cleanup4, label %for.cond103 104cleanup4: ; preds = %for.cond, %for.body, %entry105 %1 = phi float [ -7.000000e+00, %entry ], [ %0, %for.body ], [ -7.000000e+00, %for.cond ]106 ret float %1107}108