|
| 1 | +; RUN: opt < %s -separate-const-offset-from-gep -dce -S | FileCheck %s |
| 2 | + |
| 3 | +; Several unit tests for -separate-const-offset-from-gep. The transformation |
| 4 | +; heavily relies on TargetTransformInfo, so we put these tests under |
| 5 | +; target-specific folders. |
| 6 | + |
| 7 | +target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128" |
| 8 | +; target triple is necessary; otherwise TargetTransformInfo rejects any |
| 9 | +; addressing mode. |
| 10 | +target triple = "nvptx64-unknown-unknown" |
| 11 | + |
| 12 | +%struct.S = type { float, double } |
| 13 | + |
| 14 | +@struct_array = global [1024 x %struct.S] zeroinitializer, align 16 |
| 15 | +@float_2d_array = global [32 x [32 x float]] zeroinitializer, align 4 |
| 16 | + |
| 17 | +; We should not extract any struct field indices, because fields in a struct |
| 18 | +; may have different types. |
| 19 | +define double* @struct(i32 %i) { |
| 20 | +entry: |
| 21 | + %add = add nsw i32 %i, 5 |
| 22 | + %idxprom = sext i32 %add to i64 |
| 23 | + %p = getelementptr inbounds [1024 x %struct.S]* @struct_array, i64 0, i64 %idxprom, i32 1 |
| 24 | + ret double* %p |
| 25 | +} |
| 26 | +; CHECK-LABEL: @struct |
| 27 | +; CHECK: getelementptr [1024 x %struct.S]* @struct_array, i64 0, i32 %i, i32 1 |
| 28 | + |
| 29 | +; We should be able to trace into sext/zext if it's directly used as a GEP |
| 30 | +; index. |
| 31 | +define float* @sext_zext(i32 %i, i32 %j) { |
| 32 | +entry: |
| 33 | + %i1 = add i32 %i, 1 |
| 34 | + %j2 = add i32 %j, 2 |
| 35 | + %i1.ext = sext i32 %i1 to i64 |
| 36 | + %j2.ext = zext i32 %j2 to i64 |
| 37 | + %p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i1.ext, i64 %j2.ext |
| 38 | + ret float* %p |
| 39 | +} |
| 40 | +; CHECK-LABEL: @sext_zext |
| 41 | +; CHECK: getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i32 %i, i32 %j |
| 42 | +; CHECK: add i64 %{{[0-9]+}}, 136 |
| 43 | + |
| 44 | +; We should be able to trace into sext/zext if it can be distributed to both |
| 45 | +; operands, e.g., sext (add nsw a, b) == add nsw (sext a), (sext b) |
| 46 | +define float* @ext_add_no_overflow(i64 %a, i32 %b, i64 %c, i32 %d) { |
| 47 | + %b1 = add nsw i32 %b, 1 |
| 48 | + %b2 = sext i32 %b1 to i64 |
| 49 | + %i = add i64 %a, %b2 |
| 50 | + %d1 = add nuw i32 %d, 1 |
| 51 | + %d2 = zext i32 %d1 to i64 |
| 52 | + %j = add i64 %c, %d2 |
| 53 | + %p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i, i64 %j |
| 54 | + ret float* %p |
| 55 | +} |
| 56 | +; CHECK-LABEL: @ext_add_no_overflow |
| 57 | +; CHECK: [[BASE_PTR:%[0-9]+]] = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %{{[0-9]+}}, i64 %{{[0-9]+}} |
| 58 | +; CHECK: [[BASE_INT:%[0-9]+]] = ptrtoint float* [[BASE_PTR]] to i64 |
| 59 | +; CHECK: add i64 [[BASE_INT]], 132 |
| 60 | + |
| 61 | +; We should treat "or" with no common bits (%k) as "add", and leave "or" with |
| 62 | +; potentially common bits (%l) as is. |
| 63 | +define float* @or(i64 %i) { |
| 64 | +entry: |
| 65 | + %j = shl i64 %i, 2 |
| 66 | + %k = or i64 %j, 3 ; no common bits |
| 67 | + %l = or i64 %j, 4 ; potentially common bits |
| 68 | + %p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %k, i64 %l |
| 69 | + ret float* %p |
| 70 | +} |
| 71 | +; CHECK-LABEL: @or |
| 72 | +; CHECK: getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %j, i64 %l |
| 73 | +; CHECK: add i64 %{{[0-9]+}}, 384 |
| 74 | + |
| 75 | +; The subexpression (b + 5) is used in both "i = a + (b + 5)" and "*out = b + |
| 76 | +; 5". When extracting the constant offset 5, make sure "*out = b + 5" isn't |
| 77 | +; affected. |
| 78 | +define float* @expr(i64 %a, i64 %b, i64* %out) { |
| 79 | +entry: |
| 80 | + %b5 = add i64 %b, 5 |
| 81 | + %i = add i64 %b5, %a |
| 82 | + %p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i, i64 0 |
| 83 | + store i64 %b5, i64* %out |
| 84 | + ret float* %p |
| 85 | +} |
| 86 | +; CHECK-LABEL: @expr |
| 87 | +; CHECK: getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %0, i64 0 |
| 88 | +; CHECK: add i64 %{{[0-9]+}}, 640 |
| 89 | +; CHECK: store i64 %b5, i64* %out |
| 90 | + |
| 91 | +; Verifies we handle "sub" correctly. |
| 92 | +define float* @sub(i64 %i, i64 %j) { |
| 93 | + %i2 = sub i64 %i, 5 ; i - 5 |
| 94 | + %j2 = sub i64 5, %j ; 5 - i |
| 95 | + %p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i2, i64 %j2 |
| 96 | + ret float* %p |
| 97 | +} |
| 98 | +; CHECK-LABEL: @sub |
| 99 | +; CHECK: %[[j2:[0-9]+]] = sub i64 0, %j |
| 100 | +; CHECK: getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i, i64 %[[j2]] |
| 101 | +; CHECK: add i64 %{{[0-9]+}}, -620 |
0 commit comments