xref: /llvm-project/llvm/test/CodeGen/ARM/ssat-unroll-loops.ll (revision c4a60c9d34375e73fc2da5e02215eabe4bc90e8f) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=armv6t2-eabi %s -o - | FileCheck %s --check-prefix=CHECK
3
4; Checks SSAT is still generated when loop unrolling is on
5
6define void @ssat_unroll(ptr %pSrcA, ptr %pSrcB, ptr %pDst, i32 %blockSize) {
7; CHECK-LABEL: ssat_unroll:
8; CHECK:       @ %bb.0: @ %entry
9; CHECK-NEXT:    cmp r3, #0
10; CHECK-NEXT:    bxeq lr
11; CHECK-NEXT:  .LBB0_1: @ %while.body.preheader
12; CHECK-NEXT:    .save {r11, lr}
13; CHECK-NEXT:    push {r11, lr}
14; CHECK-NEXT:    sub r12, r3, #1
15; CHECK-NEXT:    tst r3, #1
16; CHECK-NEXT:    beq .LBB0_3
17; CHECK-NEXT:  @ %bb.2: @ %while.body.prol.preheader
18; CHECK-NEXT:    ldrsh lr, [r0], #2
19; CHECK-NEXT:    ldrsh r3, [r1], #2
20; CHECK-NEXT:    smulbb r3, r3, lr
21; CHECK-NEXT:    ssat r3, #16, r3, asr #14
22; CHECK-NEXT:    strh r3, [r2], #2
23; CHECK-NEXT:    mov r3, r12
24; CHECK-NEXT:  .LBB0_3: @ %while.body.prol.loopexit
25; CHECK-NEXT:    cmp r12, #0
26; CHECK-NEXT:    beq .LBB0_5
27; CHECK-NEXT:  .LBB0_4: @ %while.body
28; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
29; CHECK-NEXT:    ldrsh r12, [r0]
30; CHECK-NEXT:    subs r3, r3, #2
31; CHECK-NEXT:    ldrsh lr, [r1]
32; CHECK-NEXT:    smulbb r12, lr, r12
33; CHECK-NEXT:    ssat r12, #16, r12, asr #14
34; CHECK-NEXT:    strh r12, [r2]
35; CHECK-NEXT:    ldrsh r12, [r0, #2]
36; CHECK-NEXT:    add r0, r0, #4
37; CHECK-NEXT:    ldrsh lr, [r1, #2]
38; CHECK-NEXT:    add r1, r1, #4
39; CHECK-NEXT:    smulbb r12, lr, r12
40; CHECK-NEXT:    ssat r12, #16, r12, asr #14
41; CHECK-NEXT:    strh r12, [r2, #2]
42; CHECK-NEXT:    add r2, r2, #4
43; CHECK-NEXT:    bne .LBB0_4
44; CHECK-NEXT:  .LBB0_5:
45; CHECK-NEXT:    pop {r11, lr}
46; CHECK-NEXT:    bx lr
47entry:
48  %cmp.not7 = icmp eq i32 %blockSize, 0
49  br i1 %cmp.not7, label %while.end, label %while.body.preheader
50
51while.body.preheader:                             ; preds = %entry
52  %0 = add i32 %blockSize, -1
53  %xtraiter = and i32 %blockSize, 1
54  %lcmp.mod.not = icmp eq i32 %xtraiter, 0
55  br i1 %lcmp.mod.not, label %while.body.prol.loopexit, label %while.body.prol.preheader
56
57while.body.prol.preheader:                        ; preds = %while.body.preheader
58  %incdec.ptr.prol = getelementptr inbounds i16, ptr %pSrcA, i32 1
59  %1 = load i16, ptr %pSrcA
60  %conv.prol = sext i16 %1 to i32
61  %incdec.ptr1.prol = getelementptr inbounds i16, ptr %pSrcB, i32 1
62  %2 = load i16, ptr %pSrcB
63  %conv2.prol = sext i16 %2 to i32
64  %mul.prol = mul nsw i32 %conv2.prol, %conv.prol
65  %shr.prol = ashr i32 %mul.prol, 14
66  %3 = icmp sgt i32 %shr.prol, -32768
67  %4 = select i1 %3, i32 %shr.prol, i32 -32768
68  %5 = icmp slt i32 %4, 32767
69  %spec.select.i.prol = select i1 %5, i32 %4, i32 32767
70  %conv3.prol = trunc i32 %spec.select.i.prol to i16
71  %incdec.ptr4.prol = getelementptr inbounds i16, ptr %pDst, i32 1
72  store i16 %conv3.prol, ptr %pDst
73  br label %while.body.prol.loopexit
74
75while.body.prol.loopexit:                         ; preds = %while.body.prol.preheader, %while.body.preheader
76  %blkCnt.011.unr = phi i32 [ %blockSize, %while.body.preheader ], [ %0, %while.body.prol.preheader ]
77  %pSrcA.addr.010.unr = phi ptr [ %pSrcA, %while.body.preheader ], [ %incdec.ptr.prol, %while.body.prol.preheader ]
78  %pDst.addr.09.unr = phi ptr [ %pDst, %while.body.preheader ], [ %incdec.ptr4.prol, %while.body.prol.preheader ]
79  %pSrcB.addr.08.unr = phi ptr [ %pSrcB, %while.body.preheader ], [ %incdec.ptr1.prol, %while.body.prol.preheader ]
80  %6 = icmp eq i32 %0, 0
81  br i1 %6, label %while.end, label %while.body
82
83while.body:                                       ; preds = %while.body.prol.loopexit, %while.body
84  %blkCnt.011 = phi i32 [ %dec.1, %while.body ], [ %blkCnt.011.unr, %while.body.prol.loopexit ]
85  %pSrcA.addr.010 = phi ptr [ %incdec.ptr.1, %while.body ], [ %pSrcA.addr.010.unr, %while.body.prol.loopexit ]
86  %pDst.addr.09 = phi ptr [ %incdec.ptr4.1, %while.body ], [ %pDst.addr.09.unr, %while.body.prol.loopexit ]
87  %pSrcB.addr.08 = phi ptr [ %incdec.ptr1.1, %while.body ], [ %pSrcB.addr.08.unr, %while.body.prol.loopexit ]
88  %incdec.ptr = getelementptr inbounds i16, ptr %pSrcA.addr.010, i32 1
89  %7 = load i16, ptr %pSrcA.addr.010
90  %conv = sext i16 %7 to i32
91  %incdec.ptr1 = getelementptr inbounds i16, ptr %pSrcB.addr.08, i32 1
92  %8 = load i16, ptr %pSrcB.addr.08
93  %conv2 = sext i16 %8 to i32
94  %mul = mul nsw i32 %conv2, %conv
95  %shr = ashr i32 %mul, 14
96  %9 = icmp sgt i32 %shr, -32768
97  %10 = select i1 %9, i32 %shr, i32 -32768
98  %11 = icmp slt i32 %10, 32767
99  %spec.select.i = select i1 %11, i32 %10, i32 32767
100  %conv3 = trunc i32 %spec.select.i to i16
101  %incdec.ptr4 = getelementptr inbounds i16, ptr %pDst.addr.09, i32 1
102  store i16 %conv3, ptr %pDst.addr.09
103  %incdec.ptr.1 = getelementptr inbounds i16, ptr %pSrcA.addr.010, i32 2
104  %12 = load i16, ptr %incdec.ptr
105  %conv.1 = sext i16 %12 to i32
106  %incdec.ptr1.1 = getelementptr inbounds i16, ptr %pSrcB.addr.08, i32 2
107  %13 = load i16, ptr %incdec.ptr1
108  %conv2.1 = sext i16 %13 to i32
109  %mul.1 = mul nsw i32 %conv2.1, %conv.1
110  %shr.1 = ashr i32 %mul.1, 14
111  %14 = icmp sgt i32 %shr.1, -32768
112  %15 = select i1 %14, i32 %shr.1, i32 -32768
113  %16 = icmp slt i32 %15, 32767
114  %spec.select.i.1 = select i1 %16, i32 %15, i32 32767
115  %conv3.1 = trunc i32 %spec.select.i.1 to i16
116  %incdec.ptr4.1 = getelementptr inbounds i16, ptr %pDst.addr.09, i32 2
117  store i16 %conv3.1, ptr %incdec.ptr4
118  %dec.1 = add i32 %blkCnt.011, -2
119  %cmp.not.1 = icmp eq i32 %dec.1, 0
120  br i1 %cmp.not.1, label %while.end, label %while.body
121
122while.end:                                        ; preds = %while.body, %while.body.prol.loopexit, %entry
123  ret void
124}
125
126define void @ssat_unroll_minmax(ptr nocapture readonly %pSrcA, ptr nocapture readonly %pSrcB, ptr nocapture writeonly %pDst, i32 %blockSize) {
127; CHECK-LABEL: ssat_unroll_minmax:
128; CHECK:       @ %bb.0: @ %entry
129; CHECK-NEXT:    cmp r3, #0
130; CHECK-NEXT:    bxeq lr
131; CHECK-NEXT:  .LBB1_1: @ %while.body.preheader
132; CHECK-NEXT:    .save {r11, lr}
133; CHECK-NEXT:    push {r11, lr}
134; CHECK-NEXT:    sub r12, r3, #1
135; CHECK-NEXT:    tst r3, #1
136; CHECK-NEXT:    beq .LBB1_3
137; CHECK-NEXT:  @ %bb.2: @ %while.body.prol.preheader
138; CHECK-NEXT:    ldrsh lr, [r0], #2
139; CHECK-NEXT:    ldrsh r3, [r1], #2
140; CHECK-NEXT:    smulbb r3, r3, lr
141; CHECK-NEXT:    ssat r3, #16, r3, asr #14
142; CHECK-NEXT:    strh r3, [r2], #2
143; CHECK-NEXT:    mov r3, r12
144; CHECK-NEXT:  .LBB1_3: @ %while.body.prol.loopexit
145; CHECK-NEXT:    cmp r12, #0
146; CHECK-NEXT:    beq .LBB1_5
147; CHECK-NEXT:  .LBB1_4: @ %while.body
148; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
149; CHECK-NEXT:    ldrsh r12, [r0]
150; CHECK-NEXT:    subs r3, r3, #2
151; CHECK-NEXT:    ldrsh lr, [r1]
152; CHECK-NEXT:    smulbb r12, lr, r12
153; CHECK-NEXT:    ssat r12, #16, r12, asr #14
154; CHECK-NEXT:    strh r12, [r2]
155; CHECK-NEXT:    ldrsh r12, [r0, #2]
156; CHECK-NEXT:    add r0, r0, #4
157; CHECK-NEXT:    ldrsh lr, [r1, #2]
158; CHECK-NEXT:    add r1, r1, #4
159; CHECK-NEXT:    smulbb r12, lr, r12
160; CHECK-NEXT:    ssat r12, #16, r12, asr #14
161; CHECK-NEXT:    strh r12, [r2, #2]
162; CHECK-NEXT:    add r2, r2, #4
163; CHECK-NEXT:    bne .LBB1_4
164; CHECK-NEXT:  .LBB1_5:
165; CHECK-NEXT:    pop {r11, lr}
166; CHECK-NEXT:    bx lr
167entry:
168  %cmp.not7 = icmp eq i32 %blockSize, 0
169  br i1 %cmp.not7, label %while.end, label %while.body.preheader
170
171while.body.preheader:                             ; preds = %entry
172  %0 = add i32 %blockSize, -1
173  %xtraiter = and i32 %blockSize, 1
174  %lcmp.mod.not = icmp eq i32 %xtraiter, 0
175  br i1 %lcmp.mod.not, label %while.body.prol.loopexit, label %while.body.prol.preheader
176
177while.body.prol.preheader:                        ; preds = %while.body.preheader
178  %incdec.ptr.prol = getelementptr inbounds i16, ptr %pSrcA, i64 1
179  %1 = load i16, ptr %pSrcA, align 2
180  %conv.prol = sext i16 %1 to i32
181  %incdec.ptr1.prol = getelementptr inbounds i16, ptr %pSrcB, i64 1
182  %2 = load i16, ptr %pSrcB, align 2
183  %conv2.prol = sext i16 %2 to i32
184  %mul.prol = mul nsw i32 %conv2.prol, %conv.prol
185  %shr.prol = ashr i32 %mul.prol, 14
186  %3 = call i32 @llvm.smax.i32(i32 %shr.prol, i32 -32768)
187  %4 = call i32 @llvm.smin.i32(i32 %3, i32 32767)
188  %conv3.prol = trunc i32 %4 to i16
189  %incdec.ptr4.prol = getelementptr inbounds i16, ptr %pDst, i64 1
190  store i16 %conv3.prol, ptr %pDst, align 2
191  br label %while.body.prol.loopexit
192
193while.body.prol.loopexit:                         ; preds = %while.body.prol.preheader, %while.body.preheader
194  %blkCnt.011.unr = phi i32 [ %blockSize, %while.body.preheader ], [ %0, %while.body.prol.preheader ]
195  %pSrcA.addr.010.unr = phi ptr [ %pSrcA, %while.body.preheader ], [ %incdec.ptr.prol, %while.body.prol.preheader ]
196  %pDst.addr.09.unr = phi ptr [ %pDst, %while.body.preheader ], [ %incdec.ptr4.prol, %while.body.prol.preheader ]
197  %pSrcB.addr.08.unr = phi ptr [ %pSrcB, %while.body.preheader ], [ %incdec.ptr1.prol, %while.body.prol.preheader ]
198  %5 = icmp eq i32 %0, 0
199  br i1 %5, label %while.end, label %while.body
200
201while.body:                                       ; preds = %while.body.prol.loopexit, %while.body
202  %blkCnt.011 = phi i32 [ %dec.1, %while.body ], [ %blkCnt.011.unr, %while.body.prol.loopexit ]
203  %pSrcA.addr.010 = phi ptr [ %incdec.ptr.1, %while.body ], [ %pSrcA.addr.010.unr, %while.body.prol.loopexit ]
204  %pDst.addr.09 = phi ptr [ %incdec.ptr4.1, %while.body ], [ %pDst.addr.09.unr, %while.body.prol.loopexit ]
205  %pSrcB.addr.08 = phi ptr [ %incdec.ptr1.1, %while.body ], [ %pSrcB.addr.08.unr, %while.body.prol.loopexit ]
206  %incdec.ptr = getelementptr inbounds i16, ptr %pSrcA.addr.010, i64 1
207  %6 = load i16, ptr %pSrcA.addr.010, align 2
208  %conv = sext i16 %6 to i32
209  %incdec.ptr1 = getelementptr inbounds i16, ptr %pSrcB.addr.08, i64 1
210  %7 = load i16, ptr %pSrcB.addr.08, align 2
211  %conv2 = sext i16 %7 to i32
212  %mul = mul nsw i32 %conv2, %conv
213  %shr = ashr i32 %mul, 14
214  %8 = call i32 @llvm.smax.i32(i32 %shr, i32 -32768)
215  %9 = call i32 @llvm.smin.i32(i32 %8, i32 32767)
216  %conv3 = trunc i32 %9 to i16
217  %incdec.ptr4 = getelementptr inbounds i16, ptr %pDst.addr.09, i64 1
218  store i16 %conv3, ptr %pDst.addr.09, align 2
219  %incdec.ptr.1 = getelementptr inbounds i16, ptr %pSrcA.addr.010, i64 2
220  %10 = load i16, ptr %incdec.ptr, align 2
221  %conv.1 = sext i16 %10 to i32
222  %incdec.ptr1.1 = getelementptr inbounds i16, ptr %pSrcB.addr.08, i64 2
223  %11 = load i16, ptr %incdec.ptr1, align 2
224  %conv2.1 = sext i16 %11 to i32
225  %mul.1 = mul nsw i32 %conv2.1, %conv.1
226  %shr.1 = ashr i32 %mul.1, 14
227  %12 = call i32 @llvm.smax.i32(i32 %shr.1, i32 -32768)
228  %13 = call i32 @llvm.smin.i32(i32 %12, i32 32767)
229  %conv3.1 = trunc i32 %13 to i16
230  %incdec.ptr4.1 = getelementptr inbounds i16, ptr %pDst.addr.09, i64 2
231  store i16 %conv3.1, ptr %incdec.ptr4, align 2
232  %dec.1 = add i32 %blkCnt.011, -2
233  %cmp.not.1 = icmp eq i32 %dec.1, 0
234  br i1 %cmp.not.1, label %while.end, label %while.body
235
236while.end:                                        ; preds = %while.body, %while.body.prol.loopexit, %entry
237  ret void
238}
239
240declare i32 @llvm.smax.i32(i32, i32) #1
241declare i32 @llvm.smin.i32(i32, i32) #1
242