113 lines · plain
1; RUN: opt %loadNPMPolly -polly-invariant-load-hoisting=true '-passes=polly-custom<scops>' -polly-print-scops -disable-output < %s 2>&1 | FileCheck %s2; RUN: opt %loadNPMPolly -S '-passes=polly<no-default-opts>' -polly-invariant-load-hoisting=true < %s | FileCheck %s --check-prefix=IR3; RUN: opt %loadNPMPolly -S '-passes=polly<no-default-opts>' -polly-invariant-load-hoisting=true --polly-overflow-tracking=always < %s | FileCheck %s --check-prefix=IRA4;5; As (p + q) can overflow we have to check that we load from6; I[p + q] only if it does not.7;8; CHECK: Invariant Accesses: {9; CHECK-NEXT: ReadAccess := [Reduction Type: NONE] [Scalar: 0]10; CHECK-NEXT: [N, p, q] -> { Stmt_for_body[i0] -> MemRef_I[p + q] };11; CHECK-NEXT: Execution Context: [N, p, q] -> { : N > 0 and -2147483648 - p <= q <= 2147483647 - p }12; CHECK-NEXT: ReadAccess := [Reduction Type: NONE] [Scalar: 0]13; CHECK-NEXT: [N, p, q] -> { Stmt_for_body[i0] -> MemRef_tmp1[0] };14; CHECK-NEXT: Execution Context: [N, p, q] -> { : N > 0 }15; CHECK-NEXT: }16;17; IR: polly.preload.merge:18; IR-NEXT: %polly.preload.tmp1.merge = phi ptr [ %polly.access.I.load, %polly.preload.exec ], [ null, %polly.preload.cond ]19; IR-NEXT: store ptr %polly.preload.tmp1.merge, ptr %tmp1.preload.s2a20; IR-NEXT: %12 = sext i32 %N to i6421; IR-NEXT: %13 = icmp sge i64 %12, 122; IR-NEXT: %14 = sext i32 %q to i6423; IR-NEXT: %15 = sext i32 %p to i6424; IR-NEXT: %16 = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %15, i64 %14)25; IR-NEXT: %.obit4 = extractvalue { i64, i1 } %16, 126; IR-NEXT: %polly.overflow.state5 = or i1 false, %.obit427; IR-NEXT: %.res6 = extractvalue { i64, i1 } %16, 028; IR-NEXT: %17 = icmp sle i64 %.res6, 214748364729; IR-NEXT: %18 = and i1 %13, %1730; IR-NEXT: %19 = sext i32 %q to i6431; IR-NEXT: %20 = sext i32 %p to i6432; IR-NEXT: %21 = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %20, i64 %19)33; IR-NEXT: %.obit7 = extractvalue { i64, i1 } %21, 134; IR-NEXT: %polly.overflow.state8 = or i1 %polly.overflow.state5, %.obit735; IR-NEXT: %.res9 = extractvalue { i64, i1 } %21, 036; IR-NEXT: %22 = icmp sge i64 %.res9, -214748364837; IR-NEXT: %23 = and i1 %18, %2238; IR-NEXT: %polly.preload.cond.overflown10 = xor i1 %polly.overflow.state8, true39; IR-NEXT: %polly.preload.cond.result11 = and i1 %23, %polly.preload.cond.overflown1040; IR-NEXT: br label %polly.preload.cond1241;42; IR: polly.preload.cond12:43; IR-NEXT: br i1 %polly.preload.cond.result1144;45; IR: polly.preload.exec14:46; IR-NEXT: %polly.access.polly.preload.tmp1.merge = getelementptr i32, ptr %polly.preload.tmp1.merge, i64 047; IR-NEXT: %polly.access.polly.preload.tmp1.merge.load = load i32, ptr %polly.access.polly.preload.tmp1.merge, align 448;49; IRA: polly.preload.merge:50; IRA-NEXT: %polly.preload.tmp1.merge = phi ptr [ %polly.access.I.load, %polly.preload.exec ], [ null, %polly.preload.cond ]51; IRA-NEXT: store ptr %polly.preload.tmp1.merge, ptr %tmp1.preload.s2a52; IRA-NEXT: %12 = sext i32 %N to i6453; IRA-NEXT: %13 = icmp sge i64 %12, 154; IRA-NEXT: %14 = sext i32 %q to i6455; IRA-NEXT: %15 = sext i32 %p to i6456; IRA-NEXT: %16 = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %15, i64 %14)57; IRA-NEXT: %.obit5 = extractvalue { i64, i1 } %16, 158; IRA-NEXT: %.res6 = extractvalue { i64, i1 } %16, 059; IRA-NEXT: %17 = icmp sle i64 %.res6, 214748364760; IRA-NEXT: %18 = and i1 %13, %1761; IRA-NEXT: %19 = sext i32 %q to i6462; IRA-NEXT: %20 = sext i32 %p to i6463; IRA-NEXT: %21 = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %20, i64 %19)64; IRA-NEXT: %.obit7 = extractvalue { i64, i1 } %21, 165; IRA-NEXT: %.res8 = extractvalue { i64, i1 } %21, 066; IRA-NEXT: %22 = icmp sge i64 %.res8, -214748364867; IRA-NEXT: %23 = and i1 %18, %2268; IRA-NEXT: %polly.preload.cond.overflown9 = xor i1 %.obit7, true69; IRA-NEXT: %polly.preload.cond.result10 = and i1 %23, %polly.preload.cond.overflown970; IRA-NEXT: br label %polly.preload.cond1171;72; IRA: polly.preload.cond11:73; IRA-NEXT: br i1 %polly.preload.cond.result1074;75; IRA: polly.preload.exec13:76; IRA-NEXT: %polly.access.polly.preload.tmp1.merge = getelementptr i32, ptr %polly.preload.tmp1.merge, i64 077; IRA-NEXT: %polly.access.polly.preload.tmp1.merge.load = load i32, ptr %polly.access.polly.preload.tmp1.merge, align 478;79; void f(int **I, int *A, int N, int p, int q) {80; for (int i = 0; i < N; i++)81; A[i] = *(I[p + q]);82; }83;84target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"85 86define void @f(ptr %I, ptr %A, i32 %N, i32 %p, i32 %q) {87entry:88 %tmp = sext i32 %N to i6489 br label %for.cond90 91for.cond: ; preds = %for.inc, %entry92 %indvars.iv = phi i64 [ %indvars.iv.next, %for.inc ], [ 0, %entry ]93 %cmp = icmp slt i64 %indvars.iv, %tmp94 br i1 %cmp, label %for.body, label %for.end95 96for.body: ; preds = %for.cond97 %add = add i32 %p, %q98 %idxprom = sext i32 %add to i6499 %arrayidx = getelementptr inbounds ptr, ptr %I, i64 %idxprom100 %tmp1 = load ptr, ptr %arrayidx, align 8101 %tmp2 = load i32, ptr %tmp1, align 4102 %arrayidx2 = getelementptr inbounds i32, ptr %A, i64 %indvars.iv103 store i32 %tmp2, ptr %arrayidx2, align 4104 br label %for.inc105 106for.inc: ; preds = %for.body107 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1108 br label %for.cond109 110for.end: ; preds = %for.cond111 ret void112}113