70 lines · plain
1; RUN: opt %loadNPMPolly -polly-tiling=0 -polly-parallel -polly-opt-outer-coincidence=no '-passes=polly-custom<opt-isl;ast>' -polly-print-ast -disable-output < %s | FileCheck %s2; RUN: opt %loadNPMPolly -polly-tiling=0 -polly-parallel -polly-opt-outer-coincidence=yes '-passes=polly-custom<opt-isl;ast>' -polly-print-ast -disable-output < %s | FileCheck %s --check-prefix=OUTER3 4; By skewing, the diagonal can be made parallel. ISL does this when the Check5; the 'outer_coincidence' option is enabled.6;7; void func(int m, int n, float A[static const restrict m][n]) {8; for (int i = 1; i < m; i+=1)9; for (int j = 1; j < n; j+=1)10; A[i][j] = A[i-1][j] + A[i][j-1];11;}12 13define void @func(i64 %m, i64 %n, ptr noalias nonnull %A) #0 {14entry:15 br label %for.cond16 17for.cond: ; preds = %for.inc11, %entry18 %i.0 = phi i64 [ 1, %entry ], [ %add12, %for.inc11 ]19 %cmp = icmp slt i64 %i.0, %m20 br i1 %cmp, label %for.cond1.preheader, label %for.end1321 22for.cond1.preheader: ; preds = %for.cond23 br label %for.cond124 25for.cond1: ; preds = %for.cond1.preheader, %for.body326 %j.0 = phi i64 [ %add10, %for.body3 ], [ 1, %for.cond1.preheader ]27 %cmp2 = icmp slt i64 %j.0, %n28 br i1 %cmp2, label %for.body3, label %for.inc1129 30for.body3: ; preds = %for.cond131 %sub = add nsw i64 %i.0, -132 %tmp = mul nsw i64 %sub, %n33 %arrayidx = getelementptr inbounds float, ptr %A, i64 %tmp34 %arrayidx4 = getelementptr inbounds float, ptr %arrayidx, i64 %j.035 %tmp13 = load float, ptr %arrayidx4, align 436 %sub5 = add nsw i64 %j.0, -137 %tmp14 = mul nsw i64 %i.0, %n38 %arrayidx6 = getelementptr inbounds float, ptr %A, i64 %tmp1439 %arrayidx7 = getelementptr inbounds float, ptr %arrayidx6, i64 %sub540 %tmp15 = load float, ptr %arrayidx7, align 441 %add = fadd float %tmp13, %tmp1542 %tmp16 = mul nsw i64 %i.0, %n43 %arrayidx8 = getelementptr inbounds float, ptr %A, i64 %tmp1644 %arrayidx9 = getelementptr inbounds float, ptr %arrayidx8, i64 %j.045 store float %add, ptr %arrayidx9, align 446 %add10 = add nuw nsw i64 %j.0, 147 br label %for.cond148 49for.inc11: ; preds = %for.cond150 %add12 = add nuw nsw i64 %i.0, 151 br label %for.cond52 53for.end13: ; preds = %for.cond54 ret void55}56 57 58; CHECK: #pragma minimal dependence distance: 159; CHECK-NEXT: for (int c0 = 0; c0 < m - 1; c0 += 1)60; CHECK-NEXT: #pragma minimal dependence distance: 161; CHECK-NEXT: for (int c1 = 0; c1 < n - 1; c1 += 1)62; CHECK-NEXT: Stmt_for_body3(c0, c1);63 64; OUTER: #pragma minimal dependence distance: 165; OUTER-NEXT: for (int c0 = 0; c0 < m + n - 3; c0 += 1)66; OUTER-NEXT: #pragma simd67; OUTER-NEXT: #pragma known-parallel68; OUTER-NEXT: for (int c1 = max(0, -n + c0 + 2); c1 <= min(m - 2, c0); c1 += 1)69; OUTER-NEXT: Stmt_for_body3(c1, c0 - c1);70