143 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --version 52; RUN: opt -S -passes=loop-vectorize -debug-only=loop-vectorize -mattr=avx512fp16 %s 2>&1 | FileCheck %s3; REQUIRES: asserts4target datalayout = "e-m:e-p:32:32-f64:32:64-f80:32-n8:16:32-S128"5target triple = "i386-unknown-linux-gnu"6 7@src = common local_unnamed_addr global [120 x half] zeroinitializer, align 48@dst = common local_unnamed_addr global [120 x half] zeroinitializer, align 49 10; Function Attrs: norecurse nounwind11define void @stride8(half %k, i32 %width_) {12entry:13 14; CHECK: Cost of 148 for VF 32: INTERLEAVE-GROUP with factor 8 at %0, ir<%arrayidx>15 16 %cmp72 = icmp sgt i32 %width_, 017 br i1 %cmp72, label %for.body.lr.ph, label %for.cond.cleanup18 19for.body.lr.ph: ; preds = %entry20 br label %for.body21 22for.cond.cleanup.loopexit: ; preds = %for.body23 br label %for.cond.cleanup24 25for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %entry26 ret void27 28for.body: ; preds = %for.body.lr.ph, %for.body29 %i.073 = phi i32 [ 0, %for.body.lr.ph ], [ %add46, %for.body ]30 %arrayidx = getelementptr inbounds [120 x half], ptr @src, i32 0, i32 %i.07331 %0 = load half, ptr %arrayidx, align 432 %mul = fmul fast half %0, %k33 %arrayidx2 = getelementptr inbounds [120 x half], ptr @dst, i32 0, i32 %i.07334 %1 = load half, ptr %arrayidx2, align 435 %add3 = fadd fast half %1, %mul36 store half %add3, ptr %arrayidx2, align 437 %add4 = or disjoint i32 %i.073, 138 %arrayidx5 = getelementptr inbounds [120 x half], ptr @src, i32 0, i32 %add439 %2 = load half, ptr %arrayidx5, align 440 %mul6 = fmul fast half %2, %k41 %arrayidx8 = getelementptr inbounds [120 x half], ptr @dst, i32 0, i32 %add442 %3 = load half, ptr %arrayidx8, align 443 %add9 = fadd fast half %3, %mul644 store half %add9, ptr %arrayidx8, align 445 %add10 = or disjoint i32 %i.073, 246 %arrayidx11 = getelementptr inbounds [120 x half], ptr @src, i32 0, i32 %add1047 %4 = load half, ptr %arrayidx11, align 448 %mul12 = fmul fast half %4, %k49 %arrayidx14 = getelementptr inbounds [120 x half], ptr @dst, i32 0, i32 %add1050 %5 = load half, ptr %arrayidx14, align 451 %add15 = fadd fast half %5, %mul1252 store half %add15, ptr %arrayidx14, align 453 %add16 = or disjoint i32 %i.073, 354 %arrayidx17 = getelementptr inbounds [120 x half], ptr @src, i32 0, i32 %add1655 %6 = load half, ptr %arrayidx17, align 456 %mul18 = fmul fast half %6, %k57 %arrayidx20 = getelementptr inbounds [120 x half], ptr @dst, i32 0, i32 %add1658 %7 = load half, ptr %arrayidx20, align 459 %add21 = fadd fast half %7, %mul1860 store half %add21, ptr %arrayidx20, align 461 %add22 = or disjoint i32 %i.073, 462 %arrayidx23 = getelementptr inbounds [120 x half], ptr @src, i32 0, i32 %add2263 %8 = load half, ptr %arrayidx23, align 464 %mul24 = fmul fast half %8, %k65 %arrayidx26 = getelementptr inbounds [120 x half], ptr @dst, i32 0, i32 %add2266 %9 = load half, ptr %arrayidx26, align 467 %add27 = fadd fast half %9, %mul2468 store half %add27, ptr %arrayidx26, align 469 %add28 = or disjoint i32 %i.073, 570 %arrayidx29 = getelementptr inbounds [120 x half], ptr @src, i32 0, i32 %add2871 %10 = load half, ptr %arrayidx29, align 472 %mul30 = fmul fast half %10, %k73 %arrayidx32 = getelementptr inbounds [120 x half], ptr @dst, i32 0, i32 %add2874 %11 = load half, ptr %arrayidx32, align 475 %add33 = fadd fast half %11, %mul3076 store half %add33, ptr %arrayidx32, align 477 %add34 = or disjoint i32 %i.073, 678 %arrayidx35 = getelementptr inbounds [120 x half], ptr @src, i32 0, i32 %add3479 %12 = load half, ptr %arrayidx35, align 480 %mul36 = fmul fast half %12, %k81 %arrayidx38 = getelementptr inbounds [120 x half], ptr @dst, i32 0, i32 %add3482 %13 = load half, ptr %arrayidx38, align 483 %add39 = fadd fast half %13, %mul3684 store half %add39, ptr %arrayidx38, align 485 %add40 = or disjoint i32 %i.073, 786 %arrayidx41 = getelementptr inbounds [120 x half], ptr @src, i32 0, i32 %add4087 %14 = load half, ptr %arrayidx41, align 488 %mul42 = fmul fast half %14, %k89 %arrayidx44 = getelementptr inbounds [120 x half], ptr @dst, i32 0, i32 %add4090 %15 = load half, ptr %arrayidx44, align 491 %add45 = fadd fast half %15, %mul4292 store half %add45, ptr %arrayidx44, align 493 %add46 = add nuw nsw i32 %i.073, 894 %cmp = icmp slt i32 %add46, %width_95 br i1 %cmp, label %for.body, label %for.cond.cleanup.loopexit96}97 98; Function Attrs: norecurse nounwind99define void @stride3(half %k, i32 %width_) {100entry:101 102; CHECK: LV: Found an estimated cost of 18 for VF 32 For instruction: %0 = load half103 104 %cmp27 = icmp sgt i32 %width_, 0105 br i1 %cmp27, label %for.body.lr.ph, label %for.cond.cleanup106 107for.body.lr.ph: ; preds = %entry108 br label %for.body109 110for.cond.cleanup: ; preds = %for.body, %entry111 ret void112 113for.body: ; preds = %for.body.lr.ph, %for.body114 %i.028 = phi i32 [ 0, %for.body.lr.ph ], [ %add16, %for.body ]115 %arrayidx = getelementptr inbounds [120 x half], ptr @src, i32 0, i32 %i.028116 %0 = load half, ptr %arrayidx, align 4117 %mul = fmul fast half %0, %k118 %arrayidx2 = getelementptr inbounds [120 x half], ptr @dst, i32 0, i32 %i.028119 %1 = load half, ptr %arrayidx2, align 4120 %add3 = fadd fast half %1, %mul121 store half %add3, ptr %arrayidx2, align 4122 %add4 = add nuw nsw i32 %i.028, 1123 %arrayidx5 = getelementptr inbounds [120 x half], ptr @src, i32 0, i32 %add4124 %2 = load half, ptr %arrayidx5, align 4125 %mul6 = fmul fast half %2, %k126 %arrayidx8 = getelementptr inbounds [120 x half], ptr @dst, i32 0, i32 %add4127 %3 = load half, ptr %arrayidx8, align 4128 %add9 = fadd fast half %3, %mul6129 store half %add9, ptr %arrayidx8, align 4130 %add10 = add nuw nsw i32 %i.028, 2131 %arrayidx11 = getelementptr inbounds [120 x half], ptr @src, i32 0, i32 %add10132 %4 = load half, ptr %arrayidx11, align 4133 %mul12 = fmul fast half %4, %k134 %arrayidx14 = getelementptr inbounds [120 x half], ptr @dst, i32 0, i32 %add10135 %5 = load half, ptr %arrayidx14, align 4136 %add15 = fadd fast half %5, %mul12137 store half %add15, ptr %arrayidx14, align 4138 %add16 = add nuw nsw i32 %i.028, 3139 %cmp = icmp slt i32 %add16, %width_140 br i1 %cmp, label %for.body, label %for.cond.cleanup141}142 143