424 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 32; RUN: opt < %s -mtriple=nvptx64-nvidia-cuda -passes=separate-const-offset-from-gep \3; RUN: -reassociate-geps-verify-no-dead-code -S | FileCheck %s4 5; Several unit tests for separate-const-offset-from-gep. The transformation6; heavily relies on TargetTransformInfo, so we put these tests under7; target-specific folders.8 9%struct.S = type { float, double }10 11@struct_array = global [1024 x %struct.S] zeroinitializer, align 1612@float_2d_array = global [32 x [32 x float]] zeroinitializer, align 413 14; We should not extract any struct field indices, because fields in a struct15; may have different types.16define ptr @struct(i32 %i) {17; CHECK-LABEL: define ptr @struct(18; CHECK-SAME: i32 [[I:%.*]]) {19; CHECK-NEXT: entry:20; CHECK-NEXT: [[TMP0:%.*]] = sext i32 [[I]] to i6421; CHECK-NEXT: [[TMP1:%.*]] = getelementptr [1024 x %struct.S], ptr @struct_array, i64 0, i64 [[TMP0]], i32 122; CHECK-NEXT: [[P2:%.*]] = getelementptr i8, ptr [[TMP1]], i64 8023; CHECK-NEXT: ret ptr [[P2]]24;25entry:26 %add = add nsw i32 %i, 527 %idxprom = sext i32 %add to i6428 %p = getelementptr inbounds [1024 x %struct.S], ptr @struct_array, i64 0, i64 %idxprom, i32 129 ret ptr %p30}31 32; We should be able to trace into sext(a + b) if a + b is non-negative33; (e.g., used as an index of an inbounds GEP) and one of a and b is34; non-negative.35define ptr @sext_add(i32 %i, i32 %j) {36; CHECK-LABEL: define ptr @sext_add(37; CHECK-SAME: i32 [[I:%.*]], i32 [[J:%.*]]) {38; CHECK-NEXT: entry:39; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[J]], -240; CHECK-NEXT: [[TMP1:%.*]] = sext i32 [[TMP0]] to i6441; CHECK-NEXT: [[TMP2:%.*]] = sext i32 [[I]] to i6442; CHECK-NEXT: [[TMP3:%.*]] = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 [[TMP2]], i64 [[TMP1]]43; CHECK-NEXT: [[P1:%.*]] = getelementptr i8, ptr [[TMP3]], i64 12844; CHECK-NEXT: ret ptr [[P1]]45;46entry:47 %0 = add i32 %i, 148 %1 = sext i32 %0 to i64 ; inbound sext(i + 1) = sext(i) + 149 %2 = add i32 %j, -250 ; However, inbound sext(j + -2) != sext(j) + -2, e.g., j = INT_MIN51 %3 = sext i32 %2 to i6452 %p = getelementptr inbounds [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 %1, i64 %353 ret ptr %p54}55 56; We should be able to trace into sext/zext if it can be distributed to both57; operands, e.g., sext (add nsw a, b) == add nsw (sext a), (sext b)58;59; This test verifies we can transform60; gep base, a + sext(b +nsw 1), c + zext(d +nuw 1)61; to62; gep base, a + sext(b), c + zext(d); gep ..., 1 * 32 + 163define ptr @ext_add_no_overflow(i64 %a, i32 %b, i64 %c, i32 %d) {64; CHECK-LABEL: define ptr @ext_add_no_overflow(65; CHECK-SAME: i64 [[A:%.*]], i32 [[B:%.*]], i64 [[C:%.*]], i32 [[D:%.*]]) {66; CHECK-NEXT: [[TMP1:%.*]] = sext i32 [[B]] to i6467; CHECK-NEXT: [[I2:%.*]] = add i64 [[A]], [[TMP1]]68; CHECK-NEXT: [[TMP2:%.*]] = zext i32 [[D]] to i6469; CHECK-NEXT: [[J4:%.*]] = add i64 [[C]], [[TMP2]]70; CHECK-NEXT: [[TMP3:%.*]] = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 [[I2]], i64 [[J4]]71; CHECK-NEXT: [[P5:%.*]] = getelementptr i8, ptr [[TMP3]], i64 13272; CHECK-NEXT: ret ptr [[P5]]73;74 %b1 = add nsw i32 %b, 175 %b2 = sext i32 %b1 to i6476 %i = add i64 %a, %b2 ; i = a + sext(b +nsw 1)77 %d1 = add nuw i32 %d, 178 %d2 = zext i32 %d1 to i6479 %j = add i64 %c, %d2 ; j = c + zext(d +nuw 1)80 %p = getelementptr inbounds [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 %i, i64 %j81 ret ptr %p82}83 84; Verifies we handle nested sext/zext correctly.85define void @sext_zext(i32 %a, i32 %b, ptr %out1, ptr %out2) {86; CHECK-LABEL: define void @sext_zext(87; CHECK-SAME: i32 [[A:%.*]], i32 [[B:%.*]], ptr [[OUT1:%.*]], ptr [[OUT2:%.*]]) {88; CHECK-NEXT: entry:89; CHECK-NEXT: [[TMP0:%.*]] = add nsw i32 [[B]], 290; CHECK-NEXT: [[TMP1:%.*]] = sext i32 [[TMP0]] to i4891; CHECK-NEXT: [[TMP2:%.*]] = zext i48 [[TMP1]] to i6492; CHECK-NEXT: [[TMP3:%.*]] = sext i32 [[A]] to i4893; CHECK-NEXT: [[TMP4:%.*]] = zext i48 [[TMP3]] to i6494; CHECK-NEXT: [[TMP5:%.*]] = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 [[TMP4]], i64 [[TMP2]]95; CHECK-NEXT: [[P11:%.*]] = getelementptr i8, ptr [[TMP5]], i64 12896; CHECK-NEXT: store ptr [[P11]], ptr [[OUT1]], align 897; CHECK-NEXT: [[TMP6:%.*]] = add nsw i32 [[B]], 498; CHECK-NEXT: [[TMP7:%.*]] = zext i32 [[TMP6]] to i4899; CHECK-NEXT: [[TMP8:%.*]] = sext i48 [[TMP7]] to i64100; CHECK-NEXT: [[TMP9:%.*]] = zext i32 [[A]] to i48101; CHECK-NEXT: [[TMP10:%.*]] = sext i48 [[TMP9]] to i64102; CHECK-NEXT: [[TMP11:%.*]] = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 [[TMP10]], i64 [[TMP8]]103; CHECK-NEXT: [[P22:%.*]] = getelementptr i8, ptr [[TMP11]], i64 384104; CHECK-NEXT: store ptr [[P22]], ptr [[OUT2]], align 8105; CHECK-NEXT: ret void106;107entry:108 %0 = add nsw nuw i32 %a, 1109 %1 = sext i32 %0 to i48110 %2 = zext i48 %1 to i64 ; zext(sext(a +nsw nuw 1)) = zext(sext(a)) + 1111 %3 = add nsw i32 %b, 2112 %4 = sext i32 %3 to i48113 %5 = zext i48 %4 to i64 ; zext(sext(b +nsw 2)) != zext(sext(b)) + 2114 %p1 = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 %2, i64 %5115 store ptr %p1, ptr %out1116 %6 = add nuw i32 %a, 3117 %7 = zext i32 %6 to i48118 %8 = sext i48 %7 to i64 ; sext(zext(a +nuw 3)) = zext(a +nuw 3) = zext(a) + 3119 %9 = add nsw i32 %b, 4120 %10 = zext i32 %9 to i48121 %11 = sext i48 %10 to i64 ; sext(zext(b +nsw 4)) != zext(b) + 4122 %p2 = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 %8, i64 %11123 store ptr %p2, ptr %out2124 ret void125}126 127; Similar to @ext_add_no_overflow, we should be able to trace into s/zext if128; its operand is an OR and the two operands of the OR have no common bits.129define ptr @sext_or(i64 %a, i32 %b) {130; CHECK-LABEL: define ptr @sext_or(131; CHECK-SAME: i64 [[A:%.*]], i32 [[B:%.*]]) {132; CHECK-NEXT: entry:133; CHECK-NEXT: [[B1:%.*]] = shl i32 [[B]], 2134; CHECK-NEXT: [[B3:%.*]] = or i32 [[B1]], 4135; CHECK-NEXT: [[B3_EXT:%.*]] = sext i32 [[B3]] to i64136; CHECK-NEXT: [[J:%.*]] = add i64 [[A]], [[B3_EXT]]137; CHECK-NEXT: [[TMP0:%.*]] = zext i32 [[B1]] to i64138; CHECK-NEXT: [[I2:%.*]] = add i64 [[A]], [[TMP0]]139; CHECK-NEXT: [[TMP1:%.*]] = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 [[I2]], i64 [[J]]140; CHECK-NEXT: [[P3:%.*]] = getelementptr i8, ptr [[TMP1]], i64 128141; CHECK-NEXT: ret ptr [[P3]]142;143entry:144 %b1 = shl i32 %b, 2145 %b2 = or disjoint i32 %b1, 1 ; (b << 2) and 1 have no common bits146 %b3 = or i32 %b1, 4 ; (b << 2) and 4 may have common bits147 %b2.ext = zext i32 %b2 to i64148 %b3.ext = sext i32 %b3 to i64149 %i = add i64 %a, %b2.ext150 %j = add i64 %a, %b3.ext151 %p = getelementptr inbounds [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 %i, i64 %j152 ret ptr %p153}154 155; The subexpression (b + 5) is used in both "i = a + (b + 5)" and "*out = b +156; 5". When extracting the constant offset 5, make sure "*out = b + 5" isn't157; affected.158define ptr @expr(i64 %a, i64 %b, ptr %out) {159; CHECK-LABEL: define ptr @expr(160; CHECK-SAME: i64 [[A:%.*]], i64 [[B:%.*]], ptr [[OUT:%.*]]) {161; CHECK-NEXT: entry:162; CHECK-NEXT: [[B5:%.*]] = add i64 [[B]], 5163; CHECK-NEXT: [[I2:%.*]] = add i64 [[B]], [[A]]164; CHECK-NEXT: [[TMP0:%.*]] = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 [[I2]], i64 0165; CHECK-NEXT: [[P3:%.*]] = getelementptr i8, ptr [[TMP0]], i64 640166; CHECK-NEXT: store i64 [[B5]], ptr [[OUT]], align 8167; CHECK-NEXT: ret ptr [[P3]]168;169entry:170 %b5 = add i64 %b, 5171 %i = add i64 %b5, %a172 %p = getelementptr inbounds [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 %i, i64 0173 store i64 %b5, ptr %out174 ret ptr %p175}176 177; d + sext(a +nsw (b +nsw (c +nsw 8))) => (d + sext(a) + sext(b) + sext(c)) + 8178define ptr @sext_expr(i32 %a, i32 %b, i32 %c, i64 %d) {179; CHECK-LABEL: define ptr @sext_expr(180; CHECK-SAME: i32 [[A:%.*]], i32 [[B:%.*]], i32 [[C:%.*]], i64 [[D:%.*]]) {181; CHECK-NEXT: entry:182; CHECK-NEXT: [[TMP0:%.*]] = sext i32 [[A]] to i64183; CHECK-NEXT: [[TMP1:%.*]] = sext i32 [[B]] to i64184; CHECK-NEXT: [[TMP2:%.*]] = sext i32 [[C]] to i64185; CHECK-NEXT: [[TMP3:%.*]] = add i64 [[TMP1]], [[TMP2]]186; CHECK-NEXT: [[TMP4:%.*]] = add i64 [[TMP0]], [[TMP3]]187; CHECK-NEXT: [[I1:%.*]] = add i64 [[D]], [[TMP4]]188; CHECK-NEXT: [[TMP5:%.*]] = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 0, i64 [[I1]]189; CHECK-NEXT: [[P2:%.*]] = getelementptr i8, ptr [[TMP5]], i64 32190; CHECK-NEXT: ret ptr [[P2]]191;192entry:193 %0 = add nsw i32 %c, 8194 %1 = add nsw i32 %b, %0195 %2 = add nsw i32 %a, %1196 %3 = sext i32 %2 to i64197 %i = add i64 %d, %3198 %p = getelementptr inbounds [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 0, i64 %i199 ret ptr %p200}201 202; Verifies we handle "sub" correctly.203define ptr @sub(i64 %i, i64 %j) {204; CHECK-LABEL: define ptr @sub(205; CHECK-SAME: i64 [[I:%.*]], i64 [[J:%.*]]) {206; CHECK-NEXT: [[J22:%.*]] = sub i64 0, [[J]]207; CHECK-NEXT: [[TMP1:%.*]] = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 [[I]], i64 [[J22]]208; CHECK-NEXT: [[P3:%.*]] = getelementptr i8, ptr [[TMP1]], i64 -620209; CHECK-NEXT: ret ptr [[P3]]210;211 %i2 = sub i64 %i, 5 ; i - 5212 %j2 = sub i64 5, %j ; 5 - i213 %p = getelementptr inbounds [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 %i2, i64 %j2214 ret ptr %p215}216 217%struct.Packed = type <{ [3 x i32], [8 x i64] }> ; <> means packed218 219; Verifies we can emit correct uglygep if the address is not natually aligned.220define ptr @packed_struct(i32 %i, i32 %j) {221; CHECK-LABEL: define ptr @packed_struct(222; CHECK-SAME: i32 [[I:%.*]], i32 [[J:%.*]]) {223; CHECK-NEXT: entry:224; CHECK-NEXT: [[S:%.*]] = alloca [1024 x %struct.Packed], align 16225; CHECK-NEXT: [[TMP0:%.*]] = sext i32 [[I]] to i64226; CHECK-NEXT: [[TMP1:%.*]] = sext i32 [[J]] to i64227; CHECK-NEXT: [[TMP2:%.*]] = getelementptr [1024 x %struct.Packed], ptr [[S]], i64 0, i64 [[TMP0]], i32 1, i64 [[TMP1]]228; CHECK-NEXT: [[ARRAYIDX33:%.*]] = getelementptr i8, ptr [[TMP2]], i64 100229; CHECK-NEXT: ret ptr [[ARRAYIDX33]]230;231entry:232 %s = alloca [1024 x %struct.Packed], align 16233 %add = add nsw i32 %j, 3234 %idxprom = sext i32 %add to i64235 %add1 = add nsw i32 %i, 1236 %idxprom2 = sext i32 %add1 to i64237 %arrayidx3 = getelementptr inbounds [1024 x %struct.Packed], ptr %s, i64 0, i64 %idxprom2, i32 1, i64 %idxprom238 ret ptr %arrayidx3239}240 241; We shouldn't be able to extract the 8 from "zext(a +nuw (b + 8))",242; because "zext(b + 8) != zext(b) + 8"243define ptr @zext_expr(i32 %a, i32 %b) {244; CHECK-LABEL: define ptr @zext_expr(245; CHECK-SAME: i32 [[A:%.*]], i32 [[B:%.*]]) {246; CHECK-NEXT: entry:247; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[B]], 8248; CHECK-NEXT: [[TMP1:%.*]] = add nuw i32 [[A]], [[TMP0]]249; CHECK-NEXT: [[I:%.*]] = zext i32 [[TMP1]] to i64250; CHECK-NEXT: [[P:%.*]] = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 0, i64 [[I]]251; CHECK-NEXT: ret ptr [[P]]252;253entry:254 %0 = add i32 %b, 8255 %1 = add nuw i32 %a, %0256 %i = zext i32 %1 to i64257 %p = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 0, i64 %i258 ret ptr %p259}260 261; Per http://llvm.org/docs/LangRef.html#id181, the indices of a off-bound gep262; should be considered sign-extended to the pointer size. Therefore,263; gep base, (add i32 a, b) != gep (gep base, i32 a), i32 b264; because265; sext(a + b) != sext(a) + sext(b)266;267; This test verifies we do not illegitimately extract the 8 from268; gep base, (i32 a + 8)269define ptr @i32_add(i32 %a) {270; CHECK-LABEL: define ptr @i32_add(271; CHECK-SAME: i32 [[A:%.*]]) {272; CHECK-NEXT: entry:273; CHECK-NEXT: [[I:%.*]] = add i32 [[A]], 8274; CHECK-NEXT: [[IDXPROM:%.*]] = sext i32 [[I]] to i64275; CHECK-NEXT: [[P:%.*]] = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 0, i64 [[IDXPROM]]276; CHECK-NEXT: ret ptr [[P]]277;278entry:279 %i = add i32 %a, 8280 %p = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 0, i32 %i281 ret ptr %p282}283 284; Verifies that we compute the correct constant offset when the index is285; sign-extended and then zero-extended. The old version of our code failed to286; handle this case because it simply computed the constant offset as the287; sign-extended value of the constant part of the GEP index.288define ptr @apint(i1 %a) {289; CHECK-LABEL: define ptr @apint(290; CHECK-SAME: i1 [[A:%.*]]) {291; CHECK-NEXT: entry:292; CHECK-NEXT: [[TMP0:%.*]] = sext i1 [[A]] to i4293; CHECK-NEXT: [[TMP1:%.*]] = zext i4 [[TMP0]] to i64294; CHECK-NEXT: [[TMP2:%.*]] = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 0, i64 [[TMP1]]295; CHECK-NEXT: [[P1:%.*]] = getelementptr i8, ptr [[TMP2]], i64 60296; CHECK-NEXT: ret ptr [[P1]]297;298entry:299 %0 = add nsw nuw i1 %a, 1300 %1 = sext i1 %0 to i4301 %2 = zext i4 %1 to i64 ; zext (sext i1 1 to i4) to i64 = 15302 %p = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 0, i64 %2303 ret ptr %p304}305 306; Do not trace into binary operators other than ADD, SUB, and OR.307define ptr @and(i64 %a) {308; CHECK-LABEL: define ptr @and(309; CHECK-SAME: i64 [[A:%.*]]) {310; CHECK-NEXT: entry:311; CHECK-NEXT: [[TMP0:%.*]] = shl i64 [[A]], 2312; CHECK-NEXT: [[TMP1:%.*]] = and i64 [[TMP0]], 1313; CHECK-NEXT: [[P:%.*]] = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 0, i64 [[TMP1]]314; CHECK-NEXT: ret ptr [[P]]315;316entry:317 %0 = shl i64 %a, 2318 %1 = and i64 %0, 1319 %p = getelementptr [32 x [32 x float]], ptr @float_2d_array, i64 0, i64 0, i64 %1320 ret ptr %p321}322 323; The code that rebuilds an OR expression used to be buggy, and failed on this324; test.325define ptr @shl_add_or(i64 %a, ptr %ptr) {326; CHECK-LABEL: define ptr @shl_add_or(327; CHECK-SAME: i64 [[A:%.*]], ptr [[PTR:%.*]]) {328; CHECK-NEXT: entry:329; CHECK-NEXT: [[SHL:%.*]] = shl i64 [[A]], 2330; CHECK-NEXT: [[OR2:%.*]] = add i64 [[SHL]], 1331; CHECK-NEXT: [[TMP0:%.*]] = getelementptr float, ptr [[PTR]], i64 [[OR2]]332; CHECK-NEXT: [[P3:%.*]] = getelementptr i8, ptr [[TMP0]], i64 48333; CHECK-NEXT: ret ptr [[P3]]334;335entry:336 %shl = shl i64 %a, 2337 %add = add i64 %shl, 12338 %or = or disjoint i64 %add, 1339 ; ((a << 2) + 12) and 1 have no common bits. Therefore,340 ; SeparateConstOffsetFromGEP is able to extract the 12.341 ; TODO(jingyue): We could reassociate the expression to combine 12 and 1.342 %p = getelementptr float, ptr %ptr, i64 %or343 ret ptr %p344}345 346; The source code used to be buggy in checking347; (AccumulativeByteOffset % ElementTypeSizeOfGEP == 0)348; where AccumulativeByteOffset is signed but ElementTypeSizeOfGEP is unsigned.349; The compiler would promote AccumulativeByteOffset to unsigned, causing350; unexpected results. For example, while -64 % (int64_t)24 != 0,351; -64 % (uint64_t)24 == 0.352%struct3 = type { i64, i32 }353%struct2 = type { %struct3, i32 }354%struct1 = type { i64, %struct2 }355%struct0 = type { i32, i32, ptr, [100 x %struct1] }356define ptr @sign_mod_unsign(ptr %ptr, i64 %idx) {357; CHECK-LABEL: define ptr @sign_mod_unsign(358; CHECK-SAME: ptr [[PTR:%.*]], i64 [[IDX:%.*]]) {359; CHECK-NEXT: entry:360; CHECK-NEXT: [[TMP0:%.*]] = getelementptr [[STRUCT0:%.*]], ptr [[PTR]], i64 0, i32 3, i64 [[IDX]], i32 1361; CHECK-NEXT: [[PTR22:%.*]] = getelementptr i8, ptr [[TMP0]], i64 -64362; CHECK-NEXT: ret ptr [[PTR22]]363;364entry:365 %arrayidx = add nsw i64 %idx, -2366 %ptr2 = getelementptr inbounds %struct0, ptr %ptr, i64 0, i32 3, i64 %arrayidx, i32 1367 ret ptr %ptr2368}369 370; Check that we can see through explicit trunc() instruction.371define ptr @trunk_explicit(ptr %ptr, i64 %idx) {372; CHECK-LABEL: define ptr @trunk_explicit(373; CHECK-SAME: ptr [[PTR:%.*]], i64 [[IDX:%.*]]) {374; CHECK-NEXT: entry:375; CHECK-NEXT: [[TMP0:%.*]] = getelementptr [[STRUCT0:%.*]], ptr [[PTR]], i64 0, i32 3, i64 [[IDX]], i32 1376; CHECK-NEXT: [[PTR21:%.*]] = getelementptr i8, ptr [[TMP0]], i64 3216377; CHECK-NEXT: ret ptr [[PTR21]]378;379entry:380 %idx0 = trunc i64 1 to i32381 %ptr2 = getelementptr inbounds %struct0, ptr %ptr, i32 %idx0, i32 3, i64 %idx, i32 1382 ret ptr %ptr2383}384 385; Check that we can deal with trunc inserted by386; canonicalizeArrayIndicesToPointerSize() if size of an index is larger than387; that of the pointer.388define ptr @trunk_long_idx(ptr %ptr, i64 %idx) {389; CHECK-LABEL: define ptr @trunk_long_idx(390; CHECK-SAME: ptr [[PTR:%.*]], i64 [[IDX:%.*]]) {391; CHECK-NEXT: entry:392; CHECK-NEXT: [[TMP0:%.*]] = getelementptr [[STRUCT0:%.*]], ptr [[PTR]], i64 0, i32 3, i64 [[IDX]], i32 1393; CHECK-NEXT: [[PTR21:%.*]] = getelementptr i8, ptr [[TMP0]], i64 3216394; CHECK-NEXT: ret ptr [[PTR21]]395;396entry:397 %ptr2 = getelementptr inbounds %struct0, ptr %ptr, i65 1, i32 3, i64 %idx, i32 1398 ret ptr %ptr2399}400 401; Do not extract large constant offset that cannot be folded in to PTX402; addressing mode403define void @large_offset(ptr %out, i32 %in) {404; CHECK-LABEL: define void @large_offset(405; CHECK-SAME: ptr [[OUT:%.*]], i32 [[IN:%.*]]) {406; CHECK-NEXT: entry:407; CHECK-NEXT: [[TMP0:%.*]] = tail call i32 @llvm.nvvm.read.ptx.sreg.tid.x()408; CHECK-NEXT: [[ADD:%.*]] = add nuw nsw i32 [[TMP0]], 536870912409; CHECK-NEXT: [[IDX:%.*]] = zext nneg i32 [[ADD]] to i64410; CHECK-NEXT: [[GETELEM:%.*]] = getelementptr inbounds i32, ptr [[OUT]], i64 [[IDX]]411; CHECK-NEXT: store i32 [[IN]], ptr [[GETELEM]], align 4412; CHECK-NEXT: ret void413;414entry:415 %0 = tail call i32 @llvm.nvvm.read.ptx.sreg.tid.x()416 %add = add nuw nsw i32 %0, 536870912417 %idx = zext nneg i32 %add to i64418 %getElem = getelementptr inbounds i32, ptr %out, i64 %idx419 store i32 %in, ptr %getElem, align 4420 ret void421}422 423declare i32 @llvm.nvvm.read.ptx.sreg.tid.x()424