simple-ldstr-ptrs.ll 6.4 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134
  1. target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
  2. ; RUN: opt < %s -bb-vectorize -bb-vectorize-req-chain-depth=3 -instcombine -gvn -S | FileCheck %s
  3. ; RUN: opt < %s -bb-vectorize -bb-vectorize-req-chain-depth=3 -bb-vectorize-aligned-only -instcombine -gvn -S | FileCheck %s -check-prefix=CHECK-AO
  4. ; FIXME: re-enable this once pointer vectors work properly
  5. ; XFAIL: *
  6. ; Simple 3-pair chain also with loads and stores (using ptrs and gep)
  7. define double @test1(i64* %a, i64* %b, i64* %c) nounwind uwtable readonly {
  8. entry:
  9. %i0 = load i64, i64* %a, align 8
  10. %i1 = load i64, i64* %b, align 8
  11. %mul = mul i64 %i0, %i1
  12. %arrayidx3 = getelementptr inbounds i64, i64* %a, i64 1
  13. %i3 = load i64, i64* %arrayidx3, align 8
  14. %arrayidx4 = getelementptr inbounds i64, i64* %b, i64 1
  15. %i4 = load i64, i64* %arrayidx4, align 8
  16. %mul5 = mul i64 %i3, %i4
  17. %ptr = inttoptr i64 %mul to double*
  18. %ptr5 = inttoptr i64 %mul5 to double*
  19. %aptr = getelementptr inbounds double, double* %ptr, i64 2
  20. %aptr5 = getelementptr inbounds double, double* %ptr5, i64 3
  21. %av = load double, double* %aptr, align 16
  22. %av5 = load double, double* %aptr5, align 16
  23. %r = fmul double %av, %av5
  24. store i64 %mul, i64* %c, align 8
  25. %arrayidx5 = getelementptr inbounds i64, i64* %c, i64 1
  26. store i64 %mul5, i64* %arrayidx5, align 8
  27. ret double %r
  28. ; CHECK-LABEL: @test1(
  29. ; CHECK: %i0.v.i0 = bitcast i64* %a to <2 x i64>*
  30. ; CHECK: %i1.v.i0 = bitcast i64* %b to <2 x i64>*
  31. ; CHECK: %i0 = load <2 x i64>, <2 x i64>* %i0.v.i0, align 8
  32. ; CHECK: %i1 = load <2 x i64>, <2 x i64>* %i1.v.i0, align 8
  33. ; CHECK: %mul = mul <2 x i64> %i0, %i1
  34. ; CHECK: %ptr = inttoptr <2 x i64> %mul to <2 x double*>
  35. ; CHECK: %aptr = getelementptr inbounds double, <2 x double*> %ptr, <2 x i64> <i64 2, i64 3>
  36. ; CHECK: %aptr.v.r1 = extractelement <2 x double*> %aptr, i32 0
  37. ; CHECK: %aptr.v.r2 = extractelement <2 x double*> %aptr, i32 1
  38. ; CHECK: %av = load double, double* %aptr.v.r1, align 16
  39. ; CHECK: %av5 = load double, double* %aptr.v.r2, align 16
  40. ; CHECK: %r = fmul double %av, %av5
  41. ; CHECK: %0 = bitcast i64* %c to <2 x i64>*
  42. ; CHECK: store <2 x i64> %mul, <2 x i64>* %0, align 8
  43. ; CHECK: ret double %r
  44. ; CHECK-AO-LABEL: @test1(
  45. ; CHECK-AO-NOT: load <2 x
  46. }
  47. ; Simple 3-pair chain with loads and stores (using ptrs and gep)
  48. define void @test2(i64** %a, i64** %b, i64** %c) nounwind uwtable readonly {
  49. entry:
  50. %i0 = load i64*, i64** %a, align 8
  51. %i1 = load i64*, i64** %b, align 8
  52. %arrayidx3 = getelementptr inbounds i64*, i64** %a, i64 1
  53. %i3 = load i64*, i64** %arrayidx3, align 8
  54. %arrayidx4 = getelementptr inbounds i64*, i64** %b, i64 1
  55. %i4 = load i64*, i64** %arrayidx4, align 8
  56. %o1 = load i64, i64* %i1, align 8
  57. %o4 = load i64, i64* %i4, align 8
  58. %ptr0 = getelementptr inbounds i64, i64* %i0, i64 %o1
  59. %ptr3 = getelementptr inbounds i64, i64* %i3, i64 %o4
  60. store i64* %ptr0, i64** %c, align 8
  61. %arrayidx5 = getelementptr inbounds i64*, i64** %c, i64 1
  62. store i64* %ptr3, i64** %arrayidx5, align 8
  63. ret void
  64. ; CHECK-LABEL: @test2(
  65. ; CHECK: %i0.v.i0 = bitcast i64** %a to <2 x i64*>*
  66. ; CHECK: %i1 = load i64*, i64** %b, align 8
  67. ; CHECK: %i0 = load <2 x i64*>, <2 x i64*>* %i0.v.i0, align 8
  68. ; CHECK: %arrayidx4 = getelementptr inbounds i64*, i64** %b, i64 1
  69. ; CHECK: %i4 = load i64*, i64** %arrayidx4, align 8
  70. ; CHECK: %o1 = load i64, i64* %i1, align 8
  71. ; CHECK: %o4 = load i64, i64* %i4, align 8
  72. ; CHECK: %ptr0.v.i1.1 = insertelement <2 x i64> undef, i64 %o1, i32 0
  73. ; CHECK: %ptr0.v.i1.2 = insertelement <2 x i64> %ptr0.v.i1.1, i64 %o4, i32 1
  74. ; CHECK: %ptr0 = getelementptr inbounds i64, <2 x i64*> %i0, <2 x i64> %ptr0.v.i1.2
  75. ; CHECK: %0 = bitcast i64** %c to <2 x i64*>*
  76. ; CHECK: store <2 x i64*> %ptr0, <2 x i64*>* %0, align 8
  77. ; CHECK: ret void
  78. ; CHECK-AO-LABEL: @test2(
  79. ; CHECK-AO-NOT: <2 x
  80. }
  81. ; Simple 3-pair chain with loads and stores (using ptrs and gep)
  82. ; using pointer vectors.
  83. define void @test3(<2 x i64*>* %a, <2 x i64*>* %b, <2 x i64*>* %c) nounwind uwtable readonly {
  84. entry:
  85. %i0 = load <2 x i64*>, <2 x i64*>* %a, align 8
  86. %i1 = load <2 x i64*>, <2 x i64*>* %b, align 8
  87. %arrayidx3 = getelementptr inbounds <2 x i64*>, <2 x i64*>* %a, i64 1
  88. %i3 = load <2 x i64*>, <2 x i64*>* %arrayidx3, align 8
  89. %arrayidx4 = getelementptr inbounds <2 x i64*>, <2 x i64*>* %b, i64 1
  90. %i4 = load <2 x i64*>, <2 x i64*>* %arrayidx4, align 8
  91. %j1 = extractelement <2 x i64*> %i1, i32 0
  92. %j4 = extractelement <2 x i64*> %i4, i32 0
  93. %o1 = load i64, i64* %j1, align 8
  94. %o4 = load i64, i64* %j4, align 8
  95. %j0 = extractelement <2 x i64*> %i0, i32 0
  96. %j3 = extractelement <2 x i64*> %i3, i32 0
  97. %ptr0 = getelementptr inbounds i64, i64* %j0, i64 %o1
  98. %ptr3 = getelementptr inbounds i64, i64* %j3, i64 %o4
  99. %qtr0 = insertelement <2 x i64*> undef, i64* %ptr0, i32 0
  100. %rtr0 = insertelement <2 x i64*> %qtr0, i64* %ptr0, i32 1
  101. %qtr3 = insertelement <2 x i64*> undef, i64* %ptr3, i32 0
  102. %rtr3 = insertelement <2 x i64*> %qtr3, i64* %ptr3, i32 1
  103. store <2 x i64*> %rtr0, <2 x i64*>* %c, align 8
  104. %arrayidx5 = getelementptr inbounds <2 x i64*>, <2 x i64*>* %c, i64 1
  105. store <2 x i64*> %rtr3, <2 x i64*>* %arrayidx5, align 8
  106. ret void
  107. ; CHECK-LABEL: @test3(
  108. ; CHECK: %i0.v.i0 = bitcast <2 x i64*>* %a to <4 x i64*>*
  109. ; CHECK: %i1 = load <2 x i64*>, <2 x i64*>* %b, align 8
  110. ; CHECK: %i0 = load <4 x i64*>, <4 x i64*>* %i0.v.i0, align 8
  111. ; CHECK: %arrayidx4 = getelementptr inbounds <2 x i64*>, <2 x i64*>* %b, i64 1
  112. ; CHECK: %i4 = load <2 x i64*>, <2 x i64*>* %arrayidx4, align 8
  113. ; CHECK: %j1 = extractelement <2 x i64*> %i1, i32 0
  114. ; CHECK: %j4 = extractelement <2 x i64*> %i4, i32 0
  115. ; CHECK: %o1 = load i64, i64* %j1, align 8
  116. ; CHECK: %o4 = load i64, i64* %j4, align 8
  117. ; CHECK: %ptr0.v.i1.1 = insertelement <2 x i64> undef, i64 %o1, i32 0
  118. ; CHECK: %ptr0.v.i1.2 = insertelement <2 x i64> %ptr0.v.i1.1, i64 %o4, i32 1
  119. ; CHECK: %ptr0.v.i0 = shufflevector <4 x i64*> %i0, <4 x i64*> undef, <2 x i32> <i32 0, i32 2>
  120. ; CHECK: %ptr0 = getelementptr inbounds i64, <2 x i64*> %ptr0.v.i0, <2 x i64> %ptr0.v.i1.2
  121. ; CHECK: %rtr0 = shufflevector <2 x i64*> %ptr0, <2 x i64*> undef, <2 x i32> zeroinitializer
  122. ; CHECK: %rtr3 = shufflevector <2 x i64*> %ptr0, <2 x i64*> undef, <2 x i32> <i32 1, i32 1>
  123. ; CHECK: %0 = bitcast <2 x i64*>* %c to <4 x i64*>*
  124. ; CHECK: %1 = shufflevector <2 x i64*> %rtr0, <2 x i64*> %rtr3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  125. ; CHECK: store <4 x i64*> %1, <4 x i64*>* %0, align 8
  126. ; CHECK: ret void
  127. ; CHECK-AO-LABEL: @test3(
  128. ; CHECK-AO-NOT: <4 x
  129. }