87 lines · plain
1; RUN: opt %loadNPMPolly '-passes=polly-custom<import-jscop;codegen>' -S < %s | FileCheck %s2; RUN: opt %loadNPMPolly '-passes=polly-custom<import-jscop;codegen>' -polly-import-jscop-postfix=pow2 -S < %s | FileCheck %s -check-prefix=POW23;4; void exprModDiv(float *A, float *B, float *C, long N, long p) {5; for (long i = 0; i < N; i++)6; C[i] += A[i] + B[i] + A[i] + B[i + p];7; }8;9;10; This test case changes the access functions such that the resulting index11; expressions are modulo or division operations. We test that the code we12; generate takes advantage of knowledge about unsigned numerators. This is13; useful as LLVM will translate urem and udiv operations with power-of-two14; denominators to fast bitwise and or shift operations.15 16; A[i % 127]17; CHECK: %pexp.pdiv_r = urem i64 %polly.indvar, 12718; CHECK: %polly.access.A9 = getelementptr float, ptr %A, i64 %pexp.pdiv_r19 20; A[floor(i / 127)]21;22; Note: without the floor, we would create a map i -> i/127, which only contains23; values of i that are divisible by 127. All other values of i would not24; be mapped to any value. However, to generate correct code we require25; each value of i to indeed be mapped to a value.26;27; CHECK: %pexp.p_div_q = udiv i64 %polly.indvar, 12728; CHECK: %polly.access.B10 = getelementptr float, ptr %B, i64 %pexp.p_div_q29 30; A[p % 128]31 32; A[p / 127]33; CHECK: %pexp.div = sdiv exact i64 %p, 12734; CHECK: %polly.access.B12 = getelementptr float, ptr %B, i64 %pexp.div35 36; A[i % 128]37; POW2: %pexp.pdiv_r = urem i64 %polly.indvar, 12838; POW2: %polly.access.A9 = getelementptr float, ptr %A, i64 %pexp.pdiv_r39 40; A[floor(i / 128)]41; POW2: %pexp.p_div_q = udiv i64 %polly.indvar, 12842; POW2: %polly.access.B10 = getelementptr float, ptr %B, i64 %pexp.p_div_q43 44; A[p % 128]45 46; A[p / 128]47; POW2: %pexp.div = sdiv exact i64 %p, 12848; POW2: %polly.access.B12 = getelementptr float, ptr %B, i64 %pexp.div49 50target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"51 52define void @exprModDiv(ptr %A, ptr %B, ptr %C, i64 %N, i64 %p) {53entry:54 br label %for.cond55 56for.cond: ; preds = %for.inc, %entry57 %i.0 = phi i64 [ 0, %entry ], [ %inc, %for.inc ]58 %cmp = icmp slt i64 %i.0, %N59 br i1 %cmp, label %for.body, label %for.end60 61for.body: ; preds = %for.cond62 %arrayidx = getelementptr inbounds float, ptr %A, i64 %i.063 %tmp = load float, ptr %arrayidx, align 464 %arrayidx1 = getelementptr inbounds float, ptr %B, i64 %i.065 %tmp1 = load float, ptr %arrayidx1, align 466 %add = fadd float %tmp, %tmp167 %arrayidx2 = getelementptr inbounds float, ptr %A, i64 %i.068 %tmp2 = load float, ptr %arrayidx2, align 469 %add3 = fadd float %add, %tmp270 %padd = add nsw i64 %p, %i.071 %arrayidx4 = getelementptr inbounds float, ptr %B, i64 %padd72 %tmp3 = load float, ptr %arrayidx4, align 473 %add5 = fadd float %add3, %tmp374 %arrayidx6 = getelementptr inbounds float, ptr %C, i64 %i.075 %tmp4 = load float, ptr %arrayidx6, align 476 %add7 = fadd float %tmp4, %add577 store float %add7, ptr %arrayidx6, align 478 br label %for.inc79 80for.inc: ; preds = %for.body81 %inc = add nuw nsw i64 %i.0, 182 br label %for.cond83 84for.end: ; preds = %for.cond85 ret void86}87