|
| 1 | +; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5 |
| 2 | +; RUN: opt -S -passes=vector-combine < %s | FileCheck %s |
| 3 | + |
| 4 | +define <4 x i32> @add_v4i32(i32 %x, i32 %y) { |
| 5 | +; CHECK-LABEL: define <4 x i32> @add_v4i32( |
| 6 | +; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) { |
| 7 | +; CHECK-NEXT: [[X_HEAD:%.*]] = insertelement <4 x i32> poison, i32 [[X]], i32 0 |
| 8 | +; CHECK-NEXT: [[X_SPLAT:%.*]] = shufflevector <4 x i32> [[X_HEAD]], <4 x i32> poison, <4 x i32> zeroinitializer |
| 9 | +; CHECK-NEXT: [[RES:%.*]] = add <4 x i32> [[X_SPLAT]], splat (i32 42) |
| 10 | +; CHECK-NEXT: ret <4 x i32> [[RES]] |
| 11 | +; |
| 12 | + %x.head = insertelement <4 x i32> poison, i32 %x, i32 0 |
| 13 | + %x.splat = shufflevector <4 x i32> %x.head, <4 x i32> poison, <4 x i32> zeroinitializer |
| 14 | + %res = add <4 x i32> %x.splat, splat (i32 42) |
| 15 | + ret <4 x i32> %res |
| 16 | +} |
| 17 | + |
| 18 | +define <vscale x 4 x i32> @add_nxv4i32(i32 %x, i32 %y) { |
| 19 | +; CHECK-LABEL: define <vscale x 4 x i32> @add_nxv4i32( |
| 20 | +; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) { |
| 21 | +; CHECK-NEXT: [[Y_HEAD1:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[X]], i32 0 |
| 22 | +; CHECK-NEXT: [[Y_SPLAT1:%.*]] = shufflevector <vscale x 4 x i32> [[Y_HEAD1]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer |
| 23 | +; CHECK-NEXT: [[RES:%.*]] = add <vscale x 4 x i32> [[Y_SPLAT1]], splat (i32 42) |
| 24 | +; CHECK-NEXT: ret <vscale x 4 x i32> [[RES]] |
| 25 | +; |
| 26 | + %x.head = insertelement <vscale x 4 x i32> poison, i32 %x, i32 0 |
| 27 | + %x.splat = shufflevector <vscale x 4 x i32> %x.head, <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer |
| 28 | + %res = add <vscale x 4 x i32> %x.splat, splat (i32 42) |
| 29 | + ret <vscale x 4 x i32> %res |
| 30 | +} |
| 31 | + |
| 32 | +; Make sure that we can scalarize sequences of vector instructions. |
| 33 | +define <4 x i32> @add_mul_v4i32(i32 %x, i32 %y, i32 %z) { |
| 34 | +; CHECK-LABEL: define <4 x i32> @add_mul_v4i32( |
| 35 | +; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]], i32 [[Z:%.*]]) { |
| 36 | +; CHECK-NEXT: [[Z_HEAD1:%.*]] = insertelement <4 x i32> poison, i32 [[X]], i32 0 |
| 37 | +; CHECK-NEXT: [[Z_SPLAT1:%.*]] = shufflevector <4 x i32> [[Z_HEAD1]], <4 x i32> poison, <4 x i32> zeroinitializer |
| 38 | +; CHECK-NEXT: [[RES0:%.*]] = add <4 x i32> [[Z_SPLAT1]], splat (i32 42) |
| 39 | +; CHECK-NEXT: [[RES1:%.*]] = mul <4 x i32> [[RES0]], splat (i32 42) |
| 40 | +; CHECK-NEXT: ret <4 x i32> [[RES1]] |
| 41 | +; |
| 42 | + %x.head = insertelement <4 x i32> poison, i32 %x, i32 0 |
| 43 | + %x.splat = shufflevector <4 x i32> %x.head, <4 x i32> poison, <4 x i32> zeroinitializer |
| 44 | + %res0 = add <4 x i32> %x.splat, splat (i32 42) |
| 45 | + %res1 = mul <4 x i32> %res0, splat (i32 42) |
| 46 | + ret <4 x i32> %res1 |
| 47 | +} |
| 48 | + |
| 49 | +; Shouldn't be scalarized since %x.splat and %y.splat have other users. |
| 50 | +define <4 x i32> @other_users_v4i32(i32 %x, i32 %y, ptr %p, ptr %q) { |
| 51 | +; CHECK-LABEL: define <4 x i32> @other_users_v4i32( |
| 52 | +; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]], ptr [[P:%.*]], ptr [[Q:%.*]]) { |
| 53 | +; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[X]], i32 0 |
| 54 | +; CHECK-NEXT: [[RES:%.*]] = shufflevector <4 x i32> [[DOTSPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer |
| 55 | +; CHECK-NEXT: [[RES1:%.*]] = add <4 x i32> [[RES]], splat (i32 42) |
| 56 | +; CHECK-NEXT: store <4 x i32> [[RES]], ptr [[P]], align 16 |
| 57 | +; CHECK-NEXT: store <4 x i32> [[RES]], ptr [[Q]], align 16 |
| 58 | +; CHECK-NEXT: ret <4 x i32> [[RES1]] |
| 59 | +; |
| 60 | + %x.head = insertelement <4 x i32> poison, i32 %x, i32 0 |
| 61 | + %x.splat = shufflevector <4 x i32> %x.head, <4 x i32> poison, <4 x i32> zeroinitializer |
| 62 | + %res = add <4 x i32> %x.splat, splat (i32 42) |
| 63 | + store <4 x i32> %x.splat, ptr %p |
| 64 | + store <4 x i32> %x.splat, ptr %q |
| 65 | + ret <4 x i32> %res |
| 66 | +} |
| 67 | + |
| 68 | +define <4 x i1> @icmp_v4i32(i32 %x, i32 %y) { |
| 69 | +; CHECK-LABEL: define <4 x i1> @icmp_v4i32( |
| 70 | +; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) { |
| 71 | +; CHECK-NEXT: [[X_HEAD:%.*]] = insertelement <4 x i32> poison, i32 [[X]], i32 0 |
| 72 | +; CHECK-NEXT: [[X_SPLAT:%.*]] = shufflevector <4 x i32> [[X_HEAD]], <4 x i32> poison, <4 x i32> zeroinitializer |
| 73 | +; CHECK-NEXT: [[RES:%.*]] = icmp eq <4 x i32> [[X_SPLAT]], splat (i32 42) |
| 74 | +; CHECK-NEXT: ret <4 x i1> [[RES]] |
| 75 | +; |
| 76 | + %x.head = insertelement <4 x i32> poison, i32 %x, i32 0 |
| 77 | + %x.splat = shufflevector <4 x i32> %x.head, <4 x i32> poison, <4 x i32> zeroinitializer |
| 78 | + %res = icmp eq <4 x i32> %x.splat, splat (i32 42) |
| 79 | + ret <4 x i1> %res |
| 80 | +} |
| 81 | + |
| 82 | +define <vscale x 4 x i1> @icmp_nxv4i32(i32 %x, i32 %y) { |
| 83 | +; CHECK-LABEL: define <vscale x 4 x i1> @icmp_nxv4i32( |
| 84 | +; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) { |
| 85 | +; CHECK-NEXT: [[X_HEAD:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[X]], i32 0 |
| 86 | +; CHECK-NEXT: [[X_SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[X_HEAD]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer |
| 87 | +; CHECK-NEXT: [[RES:%.*]] = icmp eq <vscale x 4 x i32> [[X_SPLAT]], splat (i32 42) |
| 88 | +; CHECK-NEXT: ret <vscale x 4 x i1> [[RES]] |
| 89 | +; |
| 90 | + %x.head = insertelement <vscale x 4 x i32> poison, i32 %x, i32 0 |
| 91 | + %x.splat = shufflevector <vscale x 4 x i32> %x.head, <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer |
| 92 | + %res = icmp eq <vscale x 4 x i32> %x.splat, splat (i32 42) |
| 93 | + ret <vscale x 4 x i1> %res |
| 94 | +} |
0 commit comments