1 ; RUN: llc < %s -mcpu=cortex-a9 -verify-coalescing -verify-machineinstrs | FileCheck %s
2 target datalayout = "e-p:32:32:32-i1:8:32-i8:8:32-i16:16:32-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:32:64-v128:32:128-a0:0:32-n32-S32"
3 target triple = "thumbv7-apple-ios0.0.0"
6 ; The vld2 and vst2 are not aligned wrt each other, the second Q loaded is the
8 ; The coalescer must find a super-register larger than QQ to eliminate the copy
9 ; setting up the vst2 data.
14 define void @f(ptr %p, i32 %c) nounwind ssp {
16 %vld2 = tail call { <4 x float>, <4 x float> } @llvm.arm.neon.vld2.v4f32.p0(ptr %p, i32 4)
17 %vld221 = extractvalue { <4 x float>, <4 x float> } %vld2, 1
18 %add.ptr = getelementptr inbounds float, ptr %p, i32 8
19 tail call void @llvm.arm.neon.vst2.p0.v4f32(ptr %add.ptr, <4 x float> %vld221, <4 x float> undef, i32 4)
24 ; FIXME: This function still has copies.
25 define void @f1(ptr %p, i32 %c) nounwind ssp {
27 %vld2 = tail call { <4 x float>, <4 x float> } @llvm.arm.neon.vld2.v4f32.p0(ptr %p, i32 4)
28 %vld221 = extractvalue { <4 x float>, <4 x float> } %vld2, 1
29 %add.ptr = getelementptr inbounds float, ptr %p, i32 8
30 %vld22 = tail call { <4 x float>, <4 x float> } @llvm.arm.neon.vld2.v4f32.p0(ptr %add.ptr, i32 4)
31 %vld2215 = extractvalue { <4 x float>, <4 x float> } %vld22, 0
32 tail call void @llvm.arm.neon.vst2.p0.v4f32(ptr %add.ptr, <4 x float> %vld221, <4 x float> %vld2215, i32 4)
37 ; FIXME: This function still has copies.
38 define void @f2(ptr %p, i32 %c) nounwind ssp {
40 %vld2 = tail call { <4 x float>, <4 x float> } @llvm.arm.neon.vld2.v4f32.p0(ptr %p, i32 4)
41 %vld224 = extractvalue { <4 x float>, <4 x float> } %vld2, 1
44 do.body: ; preds = %do.body, %entry
45 %qq0.0.1.0 = phi <4 x float> [ %vld224, %entry ], [ %vld2216, %do.body ]
46 %c.addr.0 = phi i32 [ %c, %entry ], [ %dec, %do.body ]
47 %p.addr.0 = phi ptr [ %p, %entry ], [ %add.ptr, %do.body ]
48 %add.ptr = getelementptr inbounds float, ptr %p.addr.0, i32 8
49 %vld22 = tail call { <4 x float>, <4 x float> } @llvm.arm.neon.vld2.v4f32.p0(ptr %add.ptr, i32 4)
50 %vld2215 = extractvalue { <4 x float>, <4 x float> } %vld22, 0
51 %vld2216 = extractvalue { <4 x float>, <4 x float> } %vld22, 1
52 tail call void @llvm.arm.neon.vst2.p0.v4f32(ptr %add.ptr, <4 x float> %qq0.0.1.0, <4 x float> %vld2215, i32 4)
53 %dec = add nsw i32 %c.addr.0, -1
54 %tobool = icmp eq i32 %dec, 0
55 br i1 %tobool, label %do.end, label %do.body
57 do.end: ; preds = %do.body
61 declare { <4 x float>, <4 x float> } @llvm.arm.neon.vld2.v4f32.p0(ptr, i32) nounwind readonly
62 declare void @llvm.arm.neon.vst2.p0.v4f32(ptr, <4 x float>, <4 x float>, i32) nounwind
65 ; This function has lane insertions that span basic blocks.
66 ; The trivial REG_SEQUENCE lowering can't handle that, but the coalescer can.
68 ; void f3(ptr p, ptr q) {
80 define void @f3(ptr %p, ptr %q) nounwind ssp {
82 %arrayidx = getelementptr inbounds float, ptr %p, i32 3
83 %0 = load float, ptr %arrayidx, align 4
84 %vecins = insertelement <2 x float> undef, float %0, i32 1
85 %tobool = icmp eq ptr %q, null
86 br i1 %tobool, label %if.else, label %if.then
88 if.then: ; preds = %entry
89 %1 = load float, ptr %q, align 4
90 %arrayidx2 = getelementptr inbounds float, ptr %q, i32 1
91 %2 = load float, ptr %arrayidx2, align 4
92 %add = fadd float %1, %2
93 %vecins3 = insertelement <2 x float> %vecins, float %add, i32 0
96 if.else: ; preds = %entry
97 %arrayidx4 = getelementptr inbounds float, ptr %p, i32 2
98 %3 = load float, ptr %arrayidx4, align 4
99 %vecins5 = insertelement <2 x float> %vecins, float %3, i32 0
102 if.end: ; preds = %if.else, %if.then
103 %x.0 = phi <2 x float> [ %vecins3, %if.then ], [ %vecins5, %if.else ]
104 %add.ptr = getelementptr inbounds float, ptr %p, i32 4
105 tail call void @llvm.arm.neon.vst1.p0.v2f32(ptr %add.ptr, <2 x float> %x.0, i32 4)
109 declare void @llvm.arm.neon.vst1.p0.v2f32(ptr, <2 x float>, i32) nounwind
110 declare <2 x float> @llvm.arm.neon.vld1.v2f32.p0(ptr, i32) nounwind readonly
113 ; This function inserts a lane into a fully defined vector.
114 ; The destination lane isn't read, so the subregs can coalesce.
117 define void @f4(ptr %p, ptr %q) nounwind ssp {
119 %vld1 = tail call <2 x float> @llvm.arm.neon.vld1.v2f32.p0(ptr %p, i32 4)
120 %tobool = icmp eq ptr %q, null
121 br i1 %tobool, label %if.end, label %if.then
123 if.then: ; preds = %entry
124 %0 = load float, ptr %q, align 4
125 %arrayidx1 = getelementptr inbounds float, ptr %q, i32 1
126 %1 = load float, ptr %arrayidx1, align 4
127 %add = fadd float %0, %1
128 %vecins = insertelement <2 x float> %vld1, float %add, i32 1
131 if.end: ; preds = %entry, %if.then
132 %x.0 = phi <2 x float> [ %vecins, %if.then ], [ %vld1, %entry ]
133 tail call void @llvm.arm.neon.vst1.p0.v2f32(ptr %p, <2 x float> %x.0, i32 4)
138 ; Coalesce vector lanes through phis.
139 ; CHECK: vmov.f32 {{.*}}, #1.0
143 ; We may leave the last insertelement in the if.end block.
144 ; It is inserting the %add value into a dead lane, but %add causes interference
145 ; in the entry block, and we don't do dead lane checks across basic blocks.
146 define void @f5(ptr %p, ptr %q) nounwind ssp {
148 %vld1 = tail call <4 x float> @llvm.arm.neon.vld1.v4f32.p0(ptr %p, i32 4)
149 %vecext = extractelement <4 x float> %vld1, i32 0
150 %vecext1 = extractelement <4 x float> %vld1, i32 1
151 %vecext2 = extractelement <4 x float> %vld1, i32 2
152 %vecext3 = extractelement <4 x float> %vld1, i32 3
153 %add = fadd float %vecext3, 1.000000e+00
154 %tobool = icmp eq ptr %q, null
155 br i1 %tobool, label %if.end, label %if.then
157 if.then: ; preds = %entry
158 %arrayidx = getelementptr inbounds float, ptr %q, i32 1
159 %0 = load float, ptr %arrayidx, align 4
160 %add4 = fadd float %vecext, %0
161 %1 = load float, ptr %q, align 4
162 %add6 = fadd float %vecext1, %1
163 %arrayidx7 = getelementptr inbounds float, ptr %q, i32 2
164 %2 = load float, ptr %arrayidx7, align 4
165 %add8 = fadd float %vecext2, %2
168 if.end: ; preds = %entry, %if.then
169 %a.0 = phi float [ %add4, %if.then ], [ %vecext, %entry ]
170 %b.0 = phi float [ %add6, %if.then ], [ %vecext1, %entry ]
171 %c.0 = phi float [ %add8, %if.then ], [ %vecext2, %entry ]
172 %vecinit = insertelement <4 x float> undef, float %a.0, i32 0
173 %vecinit9 = insertelement <4 x float> %vecinit, float %b.0, i32 1
174 %vecinit10 = insertelement <4 x float> %vecinit9, float %c.0, i32 2
175 %vecinit11 = insertelement <4 x float> %vecinit10, float %add, i32 3
176 tail call void @llvm.arm.neon.vst1.p0.v4f32(ptr %p, <4 x float> %vecinit11, i32 4)
180 declare <4 x float> @llvm.arm.neon.vld1.v4f32.p0(ptr, i32) nounwind readonly
182 declare void @llvm.arm.neon.vst1.p0.v4f32(ptr, <4 x float>, i32) nounwind
185 define void @pr13999() nounwind readonly {
187 br i1 true, label %outer_loop, label %loop.end
190 %d = phi double [ 0.0, %entry ], [ %add, %after_inner_loop ]
191 %0 = insertelement <2 x double> <double 0.0, double 0.0>, double %d, i32 0
192 br i1 undef, label %after_inner_loop, label %inner_loop
195 br i1 true, label %after_inner_loop, label %inner_loop
198 %1 = phi <2 x double> [ %0, %outer_loop ], [ <double 0.0, double 0.0>,
200 %2 = extractelement <2 x double> %1, i32 1
201 %add = fadd double 1.0, %2
202 br i1 false, label %loop.end, label %outer_loop
205 %d.end = phi double [ 0.0, %entry ], [ %add, %after_inner_loop ]
210 define arm_aapcs_vfpcc i32 @pr14078(ptr nocapture %arg, ptr nocapture %arg1, i32 %arg2) nounwind uwtable readonly {
212 br i1 undef, label %bb31, label %bb3
214 bb3: ; preds = %bb12, %bb
215 %tmp = shufflevector <2 x i64> undef, <2 x i64> undef, <1 x i32> zeroinitializer
216 %tmp4 = bitcast <1 x i64> %tmp to <2 x float>
217 %tmp5 = shufflevector <2 x float> %tmp4, <2 x float> undef, <4 x i32> zeroinitializer
218 %tmp6 = bitcast <4 x float> %tmp5 to <2 x i64>
219 %tmp7 = shufflevector <2 x i64> %tmp6, <2 x i64> undef, <1 x i32> zeroinitializer
220 %tmp8 = bitcast <1 x i64> %tmp7 to <2 x float>
221 %tmp9 = tail call <2 x float> @baz(<2 x float> <float 0xFFFFFFFFE0000000, float 0.000000e+00>, <2 x float> %tmp8, <2 x float> zeroinitializer) nounwind
222 br i1 undef, label %bb10, label %bb12
225 %tmp11 = load <4 x float>, ptr undef, align 8
228 bb12: ; preds = %bb10, %bb3
229 %tmp13 = shufflevector <2 x float> %tmp9, <2 x float> zeroinitializer, <2 x i32> <i32 0, i32 2>
230 %tmp14 = bitcast <2 x float> %tmp13 to <1 x i64>
231 %tmp15 = shufflevector <1 x i64> %tmp14, <1 x i64> zeroinitializer, <2 x i32> <i32 0, i32 1>
232 %tmp16 = bitcast <2 x i64> %tmp15 to <4 x float>
233 %tmp17 = fmul <4 x float> zeroinitializer, %tmp16
234 %tmp18 = bitcast <4 x float> %tmp17 to <2 x i64>
235 %tmp19 = shufflevector <2 x i64> %tmp18, <2 x i64> undef, <1 x i32> zeroinitializer
236 %tmp20 = bitcast <1 x i64> %tmp19 to <2 x float>
237 %tmp21 = tail call <2 x float> @baz67(<2 x float> %tmp20, <2 x float> undef) nounwind
238 %tmp22 = tail call <2 x float> @baz67(<2 x float> %tmp21, <2 x float> %tmp21) nounwind
239 %tmp23 = shufflevector <2 x float> %tmp22, <2 x float> undef, <4 x i32> zeroinitializer
240 %tmp24 = bitcast <4 x float> %tmp23 to <2 x i64>
241 %tmp25 = shufflevector <2 x i64> %tmp24, <2 x i64> undef, <1 x i32> zeroinitializer
242 %tmp26 = bitcast <1 x i64> %tmp25 to <2 x float>
243 %tmp27 = extractelement <2 x float> %tmp26, i32 0
244 %tmp28 = fcmp olt float %tmp27, 0.000000e+00
245 %tmp29 = select i1 %tmp28, i32 0, i32 undef
246 %tmp30 = icmp ult i32 undef, %arg2
247 br i1 %tmp30, label %bb3, label %bb31
249 bb31: ; preds = %bb12, %bb
250 %tmp32 = phi i32 [ 1, %bb ], [ %tmp29, %bb12 ]
254 declare <2 x float> @baz(<2 x float>, <2 x float>, <2 x float>) nounwind readnone
256 declare <2 x float> @baz67(<2 x float>, <2 x float>) nounwind readnone
258 %struct.wombat.5 = type { %struct.quux, %struct.quux, %struct.quux, %struct.quux }
259 %struct.quux = type { <4 x float> }
262 define linkonce_odr arm_aapcs_vfpcc %struct.wombat.5 @pr14079(ptr nocapture %arg, ptr nocapture %arg1, ptr nocapture %arg2) nounwind uwtable inlinehint {
264 %tmp = shufflevector <2 x i64> zeroinitializer, <2 x i64> undef, <1 x i32> zeroinitializer
265 %tmp3 = bitcast <1 x i64> %tmp to <2 x float>
266 %tmp4 = shufflevector <2 x float> %tmp3, <2 x float> zeroinitializer, <2 x i32> <i32 1, i32 3>
267 %tmp5 = shufflevector <2 x float> %tmp4, <2 x float> undef, <2 x i32> <i32 1, i32 3>
268 %tmp6 = bitcast <2 x float> %tmp5 to <1 x i64>
269 %tmp7 = shufflevector <1 x i64> undef, <1 x i64> %tmp6, <2 x i32> <i32 0, i32 1>
270 %tmp8 = bitcast <2 x i64> %tmp7 to <4 x float>
271 %tmp9 = shufflevector <2 x i64> zeroinitializer, <2 x i64> undef, <1 x i32> <i32 1>
272 %tmp10 = bitcast <1 x i64> %tmp9 to <2 x float>
273 %tmp11 = shufflevector <2 x float> %tmp10, <2 x float> undef, <2 x i32> <i32 0, i32 2>
274 %tmp12 = shufflevector <2 x float> %tmp11, <2 x float> undef, <2 x i32> <i32 0, i32 2>
275 %tmp13 = bitcast <2 x float> %tmp12 to <1 x i64>
276 %tmp14 = shufflevector <1 x i64> %tmp13, <1 x i64> undef, <2 x i32> <i32 0, i32 1>
277 %tmp15 = bitcast <2 x i64> %tmp14 to <4 x float>
278 %tmp16 = insertvalue %struct.wombat.5 undef, <4 x float> %tmp8, 1, 0
279 %tmp17 = insertvalue %struct.wombat.5 %tmp16, <4 x float> %tmp15, 2, 0
280 %tmp18 = insertvalue %struct.wombat.5 %tmp17, <4 x float> undef, 3, 0
281 ret %struct.wombat.5 %tmp18
284 ; CHECK: adjustCopiesBackFrom
285 ; The shuffle in if.else3 must be preserved even though adjustCopiesBackFrom
286 ; is tempted to remove it.
288 define internal void @adjustCopiesBackFrom(ptr noalias nocapture sret(<2 x i64>) %agg.result, <2 x i64> %in) {
290 %0 = extractelement <2 x i64> %in, i32 0
291 %cmp = icmp slt i64 %0, 1
292 %.in = select i1 %cmp, <2 x i64> <i64 0, i64 undef>, <2 x i64> %in
293 %1 = extractelement <2 x i64> %in, i32 1
294 %cmp1 = icmp slt i64 %1, 1
295 br i1 %cmp1, label %if.then2, label %if.else3
297 if.then2: ; preds = %entry
298 %2 = insertelement <2 x i64> %.in, i64 0, i32 1
301 if.else3: ; preds = %entry
302 %3 = shufflevector <2 x i64> %.in, <2 x i64> %in, <2 x i32> <i32 0, i32 3>
305 if.end4: ; preds = %if.else3, %if.then2
306 %result.2 = phi <2 x i64> [ %2, %if.then2 ], [ %3, %if.else3 ]
307 store <2 x i64> %result.2, ptr %agg.result, align 128
311 ; <rdar://problem/12758887>
312 ; RegisterCoalescer::updateRegDefsUses() could visit an instruction more than
313 ; once under rare circumstances. When widening a register from QPR to DTriple
314 ; with the original virtual register in dsub_1_dsub_2, the double rewrite would
315 ; produce an invalid sub-register.
317 ; This is because dsub_1_dsub_2 is not an idempotent sub-register index.
318 ; It will translate %vr:dsub_0 -> %vr:dsub_1.
319 define hidden fastcc void @radar12758887() nounwind optsize ssp {
321 br i1 undef, label %for.body, label %for.end70
323 for.body: ; preds = %for.end, %entry
324 br i1 undef, label %for.body29, label %for.end
326 for.body29: ; preds = %for.body29, %for.body
327 %0 = load <2 x double>, ptr null, align 1
328 %splat40 = shufflevector <2 x double> %0, <2 x double> undef, <2 x i32> zeroinitializer
329 %mul41 = fmul <2 x double> undef, %splat40
330 %add42 = fadd <2 x double> undef, %mul41
331 %splat44 = shufflevector <2 x double> %0, <2 x double> undef, <2 x i32> <i32 1, i32 1>
332 %mul45 = fmul <2 x double> undef, %splat44
333 %add46 = fadd <2 x double> undef, %mul45
334 br i1 undef, label %for.end, label %for.body29
336 for.end: ; preds = %for.body29, %for.body
337 %accumR2.0.lcssa = phi <2 x double> [ zeroinitializer, %for.body ], [ %add42, %for.body29 ]
338 %accumI2.0.lcssa = phi <2 x double> [ zeroinitializer, %for.body ], [ %add46, %for.body29 ]
339 %1 = shufflevector <2 x double> %accumI2.0.lcssa, <2 x double> undef, <2 x i32> <i32 1, i32 0>
340 %add58 = fadd <2 x double> undef, %1
341 %mul61 = fmul <2 x double> %add58, undef
342 %add63 = fadd <2 x double> undef, %mul61
343 %add64 = fadd <2 x double> undef, %add63
344 %add67 = fadd <2 x double> undef, %add64
345 store <2 x double> %add67, ptr undef, align 1
346 br i1 undef, label %for.end70, label %for.body
348 for.end70: ; preds = %for.end, %entry