1; RUN: opt %loadNPMPolly -passes=polly-opt-isl -polly-pattern-matching-based-opts=true \ 2; RUN: -polly-tc-opt=true -debug -disable-output < %s 2>&1 | FileCheck %s 3; REQUIRES: asserts 4; 5; /* C := alpha*A*B + beta*C */ 6; for (i = 0; i < _PB_NI; i++) 7; for (j = 0; j < _PB_NJ; j += 2) 8; { 9; C[i][j] *= beta; 10; for (k = 0; k < _PB_NK; ++k) 11; C[i][j] += alpha * A[i][k] * B[k][j]; 12; } 13; 14; Check that we won’t detect the matrix multiplication pattern, 15; if, for example, there are memory accesses that have stride 2 16; after the interchanging of loops. 17; 18; CHECK-NOT: The matrix multiplication pattern was detected 19; CHECK-NOT: The tensor contraction pattern was detected 20; 21target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128" 22target triple = "x86_64-unknown-unknown" 23 24define internal void @kernel_gemm(i32 %arg, i32 %arg1, i32 %arg2, double %arg3, double %arg4, ptr %arg5, ptr %arg6, ptr %arg7) #0 { 25bb: 26 br label %bb8 27 28bb8: ; preds = %bb29, %bb 29 %tmp = phi i64 [ 0, %bb ], [ %tmp30, %bb29 ] 30 br label %bb9 31 32bb9: ; preds = %bb26, %bb8 33 %tmp10 = phi i64 [ 0, %bb8 ], [ %tmp27, %bb26 ] 34 %tmp11 = getelementptr inbounds [1056 x double], ptr %arg5, i64 %tmp, i64 %tmp10 35 %tmp12 = load double, ptr %tmp11, align 8 36 %tmp13 = fmul double %tmp12, %arg4 37 store double %tmp13, ptr %tmp11, align 8 38 br label %Copy_0 39 40Copy_0: ; preds = %Copy_0, %bb9 41 %tmp15 = phi i64 [ 0, %bb9 ], [ %tmp24, %Copy_0 ] 42 %tmp16 = getelementptr inbounds [1024 x double], ptr %arg6, i64 %tmp, i64 %tmp15 43 %tmp17 = load double, ptr %tmp16, align 8 44 %tmp18 = fmul double %tmp17, %arg3 45 %tmp19 = getelementptr inbounds [1056 x double], ptr %arg7, i64 %tmp15, i64 %tmp10 46 %tmp20 = load double, ptr %tmp19, align 8 47 %tmp21 = fmul double %tmp18, %tmp20 48 %tmp22 = load double, ptr %tmp11, align 8 49 %tmp23 = fadd double %tmp22, %tmp21 50 store double %tmp23, ptr %tmp11, align 8 51 %tmp24 = add nuw nsw i64 %tmp15, 1 52 %tmp25 = icmp ne i64 %tmp24, 1024 53 br i1 %tmp25, label %Copy_0, label %bb26 54 55bb26: ; preds = %Copy_0 56 %tmp27 = add nuw nsw i64 %tmp10, 2 57 %tmp28 = icmp ne i64 %tmp27, 1056 58 br i1 %tmp28, label %bb9, label %bb29 59 60bb29: ; preds = %bb26 61 %tmp30 = add nuw nsw i64 %tmp, 1 62 %tmp31 = icmp ne i64 %tmp30, 1056 63 br i1 %tmp31, label %bb8, label %bb32 64 65bb32: ; preds = %bb29 66 ret void 67} 68