1; Test the handling of base + index + 12-bit displacement addresses for 2; large frames, in cases where no 20-bit form exists. The tests here 3; assume z10 register pressure, without the high words being available. 4; 5; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 | \ 6; RUN: FileCheck -check-prefix=CHECK-NOFP %s 7; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 -frame-pointer=all | \ 8; RUN: FileCheck -check-prefix=CHECK-FP %s 9 10declare void @foo(ptr %ptr1, ptr %ptr2) 11 12; This file tests what happens when a displacement is converted from 13; being relative to the start of a frame object to being relative to 14; the frame itself. In some cases the test is only possible if two 15; objects are allocated. 16; 17; Rather than rely on a particular order for those objects, the tests 18; instead allocate two objects of the same size and apply the test to 19; both of them. For consistency, all tests follow this model, even if 20; one object would actually be enough. 21 22; First check the highest in-range offset after conversion, which is 4092 23; for word-addressing instructions like LDEB. 24; 25; The last in-range doubleword offset is 4088. Since the frame has two 26; emergency spill slots at 160(%r15), the amount that we need to allocate 27; in order to put another object at offset 4088 is (4088 - 176) / 4 = 978 28; words. 29define void @f1(ptr %dst) { 30; CHECK-NOFP-LABEL: f1: 31; CHECK-NOFP: ldeb {{%f[0-7]}}, 4092(%r15) 32; CHECK-NOFP: br %r14 33; 34; CHECK-FP-LABEL: f1: 35; CHECK-FP: ldeb {{%f[0-7]}}, 4092(%r11) 36; CHECK-FP: br %r14 37 %region1 = alloca [978 x float], align 8 38 %region2 = alloca [978 x float], align 8 39 call void @foo(ptr %region1, ptr %region2) 40 %ptr1 = getelementptr inbounds [978 x float], ptr %region1, i64 0, i64 1 41 %ptr2 = getelementptr inbounds [978 x float], ptr %region2, i64 0, i64 1 42 %float1 = load float, ptr %ptr1 43 %float2 = load float, ptr %ptr2 44 %double1 = fpext float %float1 to double 45 %double2 = fpext float %float2 to double 46 store volatile double %double1, ptr %dst 47 store volatile double %double2, ptr %dst 48 ret void 49} 50 51; Test the first out-of-range offset. 52define void @f2(ptr %dst) { 53; CHECK-NOFP-LABEL: f2: 54; CHECK-NOFP: lghi %r1, 4096 55; CHECK-NOFP: ldeb {{%f[0-7]}}, 0(%r1,%r15) 56; CHECK-NOFP: br %r14 57; 58; CHECK-FP-LABEL: f2: 59; CHECK-FP: lghi %r1, 4096 60; CHECK-FP: ldeb {{%f[0-7]}}, 0(%r1,%r11) 61; CHECK-FP: br %r14 62 %region1 = alloca [978 x float], align 8 63 %region2 = alloca [978 x float], align 8 64 call void @foo(ptr %region1, ptr %region2) 65 %ptr1 = getelementptr inbounds [978 x float], ptr %region1, i64 0, i64 2 66 %ptr2 = getelementptr inbounds [978 x float], ptr %region2, i64 0, i64 2 67 %float1 = load float, ptr %ptr1 68 %float2 = load float, ptr %ptr2 69 %double1 = fpext float %float1 to double 70 %double2 = fpext float %float2 to double 71 store volatile double %double1, ptr %dst 72 store volatile double %double2, ptr %dst 73 ret void 74} 75 76; Test the next offset after that. 77define void @f3(ptr %dst) { 78; CHECK-NOFP-LABEL: f3: 79; CHECK-NOFP: lghi %r1, 4096 80; CHECK-NOFP: ldeb {{%f[0-7]}}, 4(%r1,%r15) 81; CHECK-NOFP: br %r14 82; 83; CHECK-FP-LABEL: f3: 84; CHECK-FP: lghi %r1, 4096 85; CHECK-FP: ldeb {{%f[0-7]}}, 4(%r1,%r11) 86; CHECK-FP: br %r14 87 %region1 = alloca [978 x float], align 8 88 %region2 = alloca [978 x float], align 8 89 call void @foo(ptr %region1, ptr %region2) 90 %ptr1 = getelementptr inbounds [978 x float], ptr %region1, i64 0, i64 3 91 %ptr2 = getelementptr inbounds [978 x float], ptr %region2, i64 0, i64 3 92 %float1 = load float, ptr %ptr1 93 %float2 = load float, ptr %ptr2 94 %double1 = fpext float %float1 to double 95 %double2 = fpext float %float2 to double 96 store volatile double %double1, ptr %dst 97 store volatile double %double2, ptr %dst 98 ret void 99} 100 101; Add 4096 bytes (1024 words) to the size of each object and repeat. 102define void @f4(ptr %dst) { 103; CHECK-NOFP-LABEL: f4: 104; CHECK-NOFP: lghi %r1, 4096 105; CHECK-NOFP: ldeb {{%f[0-7]}}, 4092(%r1,%r15) 106; CHECK-NOFP: br %r14 107; 108; CHECK-FP-LABEL: f4: 109; CHECK-FP: lghi %r1, 4096 110; CHECK-FP: ldeb {{%f[0-7]}}, 4092(%r1,%r11) 111; CHECK-FP: br %r14 112 %region1 = alloca [2002 x float], align 8 113 %region2 = alloca [2002 x float], align 8 114 call void @foo(ptr %region1, ptr %region2) 115 %ptr1 = getelementptr inbounds [2002 x float], ptr %region1, i64 0, i64 1 116 %ptr2 = getelementptr inbounds [2002 x float], ptr %region2, i64 0, i64 1 117 %float1 = load float, ptr %ptr1 118 %float2 = load float, ptr %ptr2 119 %double1 = fpext float %float1 to double 120 %double2 = fpext float %float2 to double 121 store volatile double %double1, ptr %dst 122 store volatile double %double2, ptr %dst 123 ret void 124} 125 126; ...as above. 127define void @f5(ptr %dst) { 128; CHECK-NOFP-LABEL: f5: 129; CHECK-NOFP: lghi %r1, 8192 130; CHECK-NOFP: ldeb {{%f[0-7]}}, 0(%r1,%r15) 131; CHECK-NOFP: br %r14 132; 133; CHECK-FP-LABEL: f5: 134; CHECK-FP: lghi %r1, 8192 135; CHECK-FP: ldeb {{%f[0-7]}}, 0(%r1,%r11) 136; CHECK-FP: br %r14 137 %region1 = alloca [2002 x float], align 8 138 %region2 = alloca [2002 x float], align 8 139 call void @foo(ptr %region1, ptr %region2) 140 %ptr1 = getelementptr inbounds [2002 x float], ptr %region1, i64 0, i64 2 141 %ptr2 = getelementptr inbounds [2002 x float], ptr %region2, i64 0, i64 2 142 %float1 = load float, ptr %ptr1 143 %float2 = load float, ptr %ptr2 144 %double1 = fpext float %float1 to double 145 %double2 = fpext float %float2 to double 146 store volatile double %double1, ptr %dst 147 store volatile double %double2, ptr %dst 148 ret void 149} 150 151; ...as above. 152define void @f6(ptr %dst) { 153; CHECK-NOFP-LABEL: f6: 154; CHECK-NOFP: lghi %r1, 8192 155; CHECK-NOFP: ldeb {{%f[0-7]}}, 4(%r1,%r15) 156; CHECK-NOFP: br %r14 157; 158; CHECK-FP-LABEL: f6: 159; CHECK-FP: lghi %r1, 8192 160; CHECK-FP: ldeb {{%f[0-7]}}, 4(%r1,%r11) 161; CHECK-FP: br %r14 162 %region1 = alloca [2002 x float], align 8 163 %region2 = alloca [2002 x float], align 8 164 call void @foo(ptr %region1, ptr %region2) 165 %ptr1 = getelementptr inbounds [2002 x float], ptr %region1, i64 0, i64 3 166 %ptr2 = getelementptr inbounds [2002 x float], ptr %region2, i64 0, i64 3 167 %float1 = load float, ptr %ptr1 168 %float2 = load float, ptr %ptr2 169 %double1 = fpext float %float1 to double 170 %double2 = fpext float %float2 to double 171 store volatile double %double1, ptr %dst 172 store volatile double %double2, ptr %dst 173 ret void 174} 175 176; Now try an offset of 4092 from the start of the object, with the object 177; being at offset 8192. This time we need objects of (8192 - 168) / 4 = 2004 178; words. 179define void @f7(ptr %dst) { 180; CHECK-NOFP-LABEL: f7: 181; CHECK-NOFP: lghi %r1, 8192 182; CHECK-NOFP: ldeb {{%f[0-7]}}, 4092(%r1,%r15) 183; CHECK-NOFP: br %r14 184; 185; CHECK-FP-LABEL: f7: 186; CHECK-FP: lghi %r1, 8192 187; CHECK-FP: ldeb {{%f[0-7]}}, 4092(%r1,%r11) 188; CHECK-FP: br %r14 189 %region1 = alloca [2004 x float], align 8 190 %region2 = alloca [2004 x float], align 8 191 call void @foo(ptr %region1, ptr %region2) 192 %ptr1 = getelementptr inbounds [2004 x float], ptr %region1, i64 0, i64 1023 193 %ptr2 = getelementptr inbounds [2004 x float], ptr %region2, i64 0, i64 1023 194 %float1 = load float, ptr %ptr1 195 %float2 = load float, ptr %ptr2 196 %double1 = fpext float %float1 to double 197 %double2 = fpext float %float2 to double 198 store volatile double %double1, ptr %dst 199 store volatile double %double2, ptr %dst 200 ret void 201} 202 203; Keep the object-relative offset the same but bump the size of the 204; objects by one doubleword. 205define void @f8(ptr %dst) { 206; CHECK-NOFP-LABEL: f8: 207; CHECK-NOFP: lghi %r1, 12288 208; CHECK-NOFP: ldeb {{%f[0-7]}}, 4(%r1,%r15) 209; CHECK-NOFP: br %r14 210; 211; CHECK-FP-LABEL: f8: 212; CHECK-FP: lghi %r1, 12288 213; CHECK-FP: ldeb {{%f[0-7]}}, 4(%r1,%r11) 214; CHECK-FP: br %r14 215 %region1 = alloca [2006 x float], align 8 216 %region2 = alloca [2006 x float], align 8 217 call void @foo(ptr %region1, ptr %region2) 218 %ptr1 = getelementptr inbounds [2006 x float], ptr %region1, i64 0, i64 1023 219 %ptr2 = getelementptr inbounds [2006 x float], ptr %region2, i64 0, i64 1023 220 %float1 = load float, ptr %ptr1 221 %float2 = load float, ptr %ptr2 222 %double1 = fpext float %float1 to double 223 %double2 = fpext float %float2 to double 224 store volatile double %double1, ptr %dst 225 store volatile double %double2, ptr %dst 226 ret void 227} 228 229; Check a case where the original displacement is out of range. The backend 230; should force an LAY from the outset. We don't yet do any kind of anchor 231; optimization, so there should be no offset on the LDEB itself. 232define void @f9(ptr %dst) { 233; CHECK-NOFP-LABEL: f9: 234; CHECK-NOFP: lay %r1, 12296(%r15) 235; CHECK-NOFP: ldeb {{%f[0-7]}}, 0(%r1) 236; CHECK-NOFP: br %r14 237; 238; CHECK-FP-LABEL: f9: 239; CHECK-FP: lay %r1, 12296(%r11) 240; CHECK-FP: ldeb {{%f[0-7]}}, 0(%r1) 241; CHECK-FP: br %r14 242 %region1 = alloca [2006 x float], align 8 243 %region2 = alloca [2006 x float], align 8 244 call void @foo(ptr %region1, ptr %region2) 245 %ptr1 = getelementptr inbounds [2006 x float], ptr %region1, i64 0, i64 1024 246 %ptr2 = getelementptr inbounds [2006 x float], ptr %region2, i64 0, i64 1024 247 %float1 = load float, ptr %ptr1 248 %float2 = load float, ptr %ptr2 249 %double1 = fpext float %float1 to double 250 %double2 = fpext float %float2 to double 251 store volatile double %double1, ptr %dst 252 store volatile double %double2, ptr %dst 253 ret void 254} 255 256; Repeat f2 in a case that needs the emergency spill slots, because all 257; call-clobbered and allocated call-saved registers are live. Note that 258; %vptr and %dst are copied to call-saved registers, freeing up %r2 and 259; %r3 during the main test. 260define void @f10(ptr %vptr, ptr %dst) { 261; CHECK-NOFP-LABEL: f10: 262; CHECK-NOFP: stg [[REGISTER:%r[1-9][0-4]?]], [[OFFSET:160|168]](%r15) 263; CHECK-NOFP: lghi [[REGISTER]], 4096 264; CHECK-NOFP: ldeb {{%f[0-7]}}, 0([[REGISTER]],%r15) 265; CHECK-NOFP: lg [[REGISTER]], [[OFFSET]](%r15) 266; CHECK-NOFP: br %r14 267; 268; CHECK-FP-LABEL: f10: 269; CHECK-FP: stg [[REGISTER:%r[1-9][0-4]?]], [[OFFSET:160|168]](%r11) 270; CHECK-FP: lghi [[REGISTER]], 4096 271; CHECK-FP: ldeb {{%f[0-7]}}, 0([[REGISTER]],%r11) 272; CHECK-FP: lg [[REGISTER]], [[OFFSET]](%r11) 273; CHECK-FP: br %r14 274 %region1 = alloca [978 x float], align 8 275 %region2 = alloca [978 x float], align 8 276 call void @foo(ptr %region1, ptr %region2) 277 %ptr1 = getelementptr inbounds [978 x float], ptr %region1, i64 0, i64 2 278 %ptr2 = getelementptr inbounds [978 x float], ptr %region2, i64 0, i64 2 279 %i0 = load volatile i32, ptr %vptr 280 %i1 = load volatile i32, ptr %vptr 281 %i2 = load volatile i32, ptr %vptr 282 %i3 = load volatile i32, ptr %vptr 283 %i4 = load volatile i32, ptr %vptr 284 %i5 = load volatile i32, ptr %vptr 285 %i14 = load volatile i32, ptr %vptr 286 %float1 = load float, ptr %ptr1 287 %float2 = load float, ptr %ptr2 288 %double1 = fpext float %float1 to double 289 %double2 = fpext float %float2 to double 290 store volatile double %double1, ptr %dst 291 store volatile double %double2, ptr %dst 292 store volatile i32 %i0, ptr %vptr 293 store volatile i32 %i1, ptr %vptr 294 store volatile i32 %i2, ptr %vptr 295 store volatile i32 %i3, ptr %vptr 296 store volatile i32 %i4, ptr %vptr 297 store volatile i32 %i5, ptr %vptr 298 store volatile i32 %i14, ptr %vptr 299 ret void 300} 301 302; Repeat f2 in a case where the index register is already occupied. 303define void @f11(ptr %dst, i64 %index) { 304; CHECK-NOFP-LABEL: f11: 305; CHECK-NOFP: lgr [[REGISTER:%r[1-9][0-5]?]], %r3 306; CHECK-NOFP: lay %r1, 4096(%r15) 307; CHECK-NOFP: ldeb {{%f[0-7]}}, 0([[REGISTER]],%r1) 308; CHECK-NOFP: br %r14 309; 310; CHECK-FP-LABEL: f11: 311; CHECK-FP: lgr [[REGISTER:%r[1-9][0-5]?]], %r3 312; CHECK-FP: lay %r1, 4096(%r11) 313; CHECK-FP: ldeb {{%f[0-7]}}, 0([[REGISTER]],%r1) 314; CHECK-FP: br %r14 315 %region1 = alloca [978 x float], align 8 316 %region2 = alloca [978 x float], align 8 317 call void @foo(ptr %region1, ptr %region2) 318 %elem1 = getelementptr inbounds [978 x float], ptr %region1, i64 0, i64 2 319 %elem2 = getelementptr inbounds [978 x float], ptr %region2, i64 0, i64 2 320 %base1 = ptrtoint ptr %elem1 to i64 321 %base2 = ptrtoint ptr %elem2 to i64 322 %addr1 = add i64 %base1, %index 323 %addr2 = add i64 %base2, %index 324 %ptr1 = inttoptr i64 %addr1 to ptr 325 %ptr2 = inttoptr i64 %addr2 to ptr 326 %float1 = load float, ptr %ptr1 327 %float2 = load float, ptr %ptr2 328 %double1 = fpext float %float1 to double 329 %double2 = fpext float %float2 to double 330 store volatile double %double1, ptr %dst 331 store volatile double %double2, ptr %dst 332 ret void 333} 334