1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
2 ; RUN: opt < %s -passes=loop-vectorize -mtriple=x86_64-unknown-linux -S -mcpu=slm | FileCheck %s
5 ; This test should not be vectorized in X86\SLM arch
6 ; Vectorizing the 64bit multiply in this case is wrong since
7 ; it can be done with a lower bit mode (notice that the sources is 16bit)
8 ; Also addq\subq (quad word) has a high cost on SLM arch.
9 ; this test has a bad performance (regression of -70%) if vectorized on SLM arch
11 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
12 target triple = "x86_64-unknown-linux-gnu"
14 define i32 @no_vec(i32 %LastIndex, ptr nocapture readonly %InputData, i16 signext %lag, i16 signext %Scale) {
15 ; CHECK-LABEL: define i32 @no_vec
16 ; CHECK-SAME: (i32 [[LASTINDEX:%.*]], ptr nocapture readonly [[INPUTDATA:%.*]], i16 signext [[LAG:%.*]], i16 signext [[SCALE:%.*]]) #[[ATTR0:[0-9]+]] {
18 ; CHECK-NEXT: [[CMP17:%.*]] = icmp sgt i32 [[LASTINDEX]], 0
19 ; CHECK-NEXT: br i1 [[CMP17]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_COND_CLEANUP:%.*]]
20 ; CHECK: for.body.lr.ph:
21 ; CHECK-NEXT: [[CONV5:%.*]] = sext i16 [[SCALE]] to i64
22 ; CHECK-NEXT: [[SH_PROM:%.*]] = and i64 [[CONV5]], 4294967295
23 ; CHECK-NEXT: [[TMP0:%.*]] = sext i16 [[LAG]] to i64
24 ; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[LASTINDEX]] to i64
25 ; CHECK-NEXT: br label [[FOR_BODY:%.*]]
26 ; CHECK: for.cond.cleanup.loopexit:
27 ; CHECK-NEXT: [[ADD7_LCSSA:%.*]] = phi i64 [ [[ADD7:%.*]], [[FOR_BODY]] ]
28 ; CHECK-NEXT: [[CONV8:%.*]] = trunc i64 [[ADD7_LCSSA]] to i32
29 ; CHECK-NEXT: br label [[FOR_COND_CLEANUP]]
30 ; CHECK: for.cond.cleanup:
31 ; CHECK-NEXT: [[ACCUMULATOR_0_LCSSA:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[CONV8]], [[FOR_COND_CLEANUP_LOOPEXIT:%.*]] ]
32 ; CHECK-NEXT: ret i32 [[ACCUMULATOR_0_LCSSA]]
34 ; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[FOR_BODY_LR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
35 ; CHECK-NEXT: [[ACCUMULATOR_018:%.*]] = phi i64 [ 0, [[FOR_BODY_LR_PH]] ], [ [[ADD7]], [[FOR_BODY]] ]
36 ; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i16, ptr [[INPUTDATA]], i64 [[INDVARS_IV]]
37 ; CHECK-NEXT: [[TMP1:%.*]] = load i16, ptr [[ARRAYIDX]], align 2
38 ; CHECK-NEXT: [[CONV:%.*]] = sext i16 [[TMP1]] to i64
39 ; CHECK-NEXT: [[TMP2:%.*]] = add nsw i64 [[INDVARS_IV]], [[TMP0]]
40 ; CHECK-NEXT: [[ARRAYIDX3:%.*]] = getelementptr inbounds i16, ptr [[INPUTDATA]], i64 [[TMP2]]
41 ; CHECK-NEXT: [[TMP3:%.*]] = load i16, ptr [[ARRAYIDX3]], align 2
42 ; CHECK-NEXT: [[CONV4:%.*]] = sext i16 [[TMP3]] to i64
43 ; CHECK-NEXT: [[MUL:%.*]] = mul nsw i64 [[CONV4]], [[CONV]]
44 ; CHECK-NEXT: [[SHR:%.*]] = ashr i64 [[MUL]], [[SH_PROM]]
45 ; CHECK-NEXT: [[ADD7]] = add i64 [[SHR]], [[ACCUMULATOR_018]]
46 ; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
47 ; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]
48 ; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP_LOOPEXIT]], label [[FOR_BODY]]
51 %cmp17 = icmp sgt i32 %LastIndex, 0
52 br i1 %cmp17, label %for.body.lr.ph, label %for.cond.cleanup
54 for.body.lr.ph: ; preds = %entry
55 %conv5 = sext i16 %Scale to i64
56 %sh_prom = and i64 %conv5, 4294967295
57 %0 = sext i16 %lag to i64
58 %wide.trip.count = zext i32 %LastIndex to i64
61 for.cond.cleanup.loopexit: ; preds = %for.body
62 %conv8 = trunc i64 %add7 to i32
63 br label %for.cond.cleanup
65 for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %entry
66 %Accumulator.0.lcssa = phi i32 [ 0, %entry ], [ %conv8, %for.cond.cleanup.loopexit ]
67 ret i32 %Accumulator.0.lcssa
69 for.body: ; preds = %for.body, %for.body.lr.ph
70 %indvars.iv = phi i64 [ 0, %for.body.lr.ph ], [ %indvars.iv.next, %for.body ]
71 %Accumulator.018 = phi i64 [ 0, %for.body.lr.ph ], [ %add7, %for.body ]
72 %arrayidx = getelementptr inbounds i16, ptr %InputData, i64 %indvars.iv
73 %1 = load i16, ptr %arrayidx, align 2
74 %conv = sext i16 %1 to i64
75 %2 = add nsw i64 %indvars.iv, %0
76 %arrayidx3 = getelementptr inbounds i16, ptr %InputData, i64 %2
77 %3 = load i16, ptr %arrayidx3, align 2
78 %conv4 = sext i16 %3 to i64
79 %mul = mul nsw i64 %conv4, %conv
80 %shr = ashr i64 %mul, %sh_prom
81 %add7 = add i64 %shr, %Accumulator.018
82 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
83 %exitcond = icmp eq i64 %indvars.iv.next, %wide.trip.count
84 br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body