This is an archive of the discontinued LLVM Phabricator instance.

Differential D26962

[LoadStoreVectorizer] Split the chain if the prefix is empty
AbandonedPublic

Authored by volkan on Nov 22 2016, 4:03 AM.

Details

Reviewers
asbirlea
• tstellarAMD
jlebar
arsenm
Summary

Currently, the vectorizer discards all of the instructions in the chain if
the vectorizable prefix is empty. Instead, we can try to vectorize smaller
chunks in order to increase the number of vector instructions generated.

Diff Detail

Event Timeline

volkan updated this revision to Diff 78847.Nov 22 2016, 4:03 AM
volkan retitled this revision from to [LoadStoreVectorizer] Split the chain if the prefix is empty.
volkan updated this object.
volkan added reviewers: jlebar, asbirlea, arsenm.
volkan added a subscriber: llvm-commits.
Herald added subscribers: wdng, mzolotukhin. · View Herald TranscriptNov 22 2016, 4:03 AM
anna added a subscriber: anna.Nov 22 2016, 5:16 AM

Could you please add some test cases?

volkan updated this revision to Diff 78864.Nov 22 2016, 6:10 AM

Added a test case.

Herald added a reviewer: • tstellarAMD. · View Herald TranscriptNov 22 2016, 6:10 AM
Herald added a subscriber: nhaehnle. · View Herald Transcript
volkan updated this revision to Diff 78877.Nov 22 2016, 8:12 AM
volkan edited edge metadata.

Updated the test case.

Herald edited edge metadata. · View Herald TranscriptNov 22 2016, 8:12 AM
jlebar added inline comments.Nov 22 2016, 9:55 AM
lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
741

"less than or equal to VF" is redundant with <= VF in the code. Can we say "is too small" or "is smaller than the vector width"?

749

Don't we know a priori that Left will not vectorize, because its vectorizable prefix equals Chain's vectorizable prefix?

Which leads to another question: Why do we chop off VF elements rather than just one?

750

These return bools, so can we use ||?

test/Transforms/LoadStoreVectorizer/AMDGPU/empty-prefix.ll
4

Can we write this comment in the form of

  1. "Check that we X"
  2. Something that does not rely on implementation details of the LSV? That is, can we avoid the term "vectorizable prefix", or, if we must use it, can we define it here?
asbirlea edited edge metadata.Nov 22 2016, 11:58 AM

The problem you found is real, but this is not the right solution for it.
The issue is in "getVectorizablePrefix". When a load that aliases the store is found, it gives up trying to find a prefix. In reality, all stores that precede the aliasing load (you can view this as a memory barrier) could be safely vectorized.
I'll work on a patch to fix this.

Side note: The testcase you added is fairly large. The problem can be showcased with a much smaller example, along the lines: 4 stores, 4 loads, 4 stores, with the 8 stores forming a chain. "getVectorizablePrefix" should not give up after the first store, but allow all 4 stores that precede the aliasing load to be vectorized.

asbirlea mentioned this in D27008: [LoadStoreVectorizer] Enable vectorization of stores in the presence of an aliasing load.Nov 22 2016, 3:32 PM
volkan abandoned this revision.Nov 23 2016, 3:31 AM
asbirlea mentioned this in rL287781: [LoadStoreVectorizer] Enable vectorization of stores in the presence of an….Nov 23 2016, 9:53 AM

Revision Contents

PathSize
lib/
Transforms/
Vectorize/
30 lines
test/
Transforms/
LoadStoreVectorizer/
AMDGPU/
330 lines

Diff 78877

lib/Transforms/Vectorize/LoadStoreVectorizer.cpp

Show First 20 Lines • Show All 731 Lines • ▼ Show 20 Linesbool Vectorizer::vectorizeStoreChain(
if (!isPowerOf2_32(Sz) || VF < 2 || ChainSize < 2) { if (!isPowerOf2_32(Sz) || VF < 2 || ChainSize < 2) {
InstructionsProcessed->insert(Chain.begin(), Chain.end()); InstructionsProcessed->insert(Chain.begin(), Chain.end());
return false; return false;
} }
ArrayRef<Instruction *> NewChain = getVectorizablePrefix(Chain); ArrayRef<Instruction *> NewChain = getVectorizablePrefix(Chain);
if (NewChain.empty()) { if (NewChain.empty()) {
// No vectorization possible. // No vectorization possible if the number of instructions
// is less than or equal to VF.
jlebarUnsubmitted
Not Done
ReplyInline Actions

"less than or equal to VF" is redundant with <= VF in the code. Can we say "is too small" or "is smaller than the vector width"?

jlebar: "less than or equal to VF" is redundant with `<= VF` in the code. Can we say "is too small"…
if (ChainSize <= VF) {
InstructionsProcessed->insert(Chain.begin(), Chain.end()); InstructionsProcessed->insert(Chain.begin(), Chain.end());
return false; return false;
} }
// Split the chain and try each slice seperately in order
// to increase the number of vector instructions generated.
ArrayRef<Instruction *> Left = Chain.slice(0, VF);
ArrayRef<Instruction *> Right = Chain.slice(VF);
jlebarUnsubmitted
Not Done
ReplyInline Actions

Don't we know a priori that Left will not vectorize, because its vectorizable prefix equals Chain's vectorizable prefix?

Which leads to another question: Why do we chop off VF elements rather than just one?

jlebar: Don't we know a priori that Left will not vectorize, because its vectorizable prefix equals…
return vectorizeStoreChain(Left, InstructionsProcessed) |
jlebarUnsubmitted
Not Done
ReplyInline Actions

These return bools, so can we use ||?

jlebar: These return bools, so can we use `||`?
vectorizeStoreChain(Right, InstructionsProcessed);
}
if (NewChain.size() == 1) { if (NewChain.size() == 1) {
// Failed after the first instruction. Discard it and try the smaller chain. // Failed after the first instruction. Discard it and try the smaller chain.
InstructionsProcessed->insert(NewChain.front()); InstructionsProcessed->insert(NewChain.front());
return false; return false;
} }
// Update Chain to the valid vectorizable subchain. // Update Chain to the valid vectorizable subchain.
Chain = NewChain; Chain = NewChain;
▲ Show 20 Lines • Show All 128 Lines • ▼ Show 20 Linesbool Vectorizer::vectorizeLoadChain(
if (!isPowerOf2_32(Sz) || VF < 2 || ChainSize < 2) { if (!isPowerOf2_32(Sz) || VF < 2 || ChainSize < 2) {
InstructionsProcessed->insert(Chain.begin(), Chain.end()); InstructionsProcessed->insert(Chain.begin(), Chain.end());
return false; return false;
} }
ArrayRef<Instruction *> NewChain = getVectorizablePrefix(Chain); ArrayRef<Instruction *> NewChain = getVectorizablePrefix(Chain);
if (NewChain.empty()) { if (NewChain.empty()) {
// No vectorization possible. // No vectorization possible if the number of instructions
// is less than or equal to VF.
if (ChainSize <= VF) {
InstructionsProcessed->insert(Chain.begin(), Chain.end()); InstructionsProcessed->insert(Chain.begin(), Chain.end());
return false; return false;
} }
// Split the chain and try each slice seperately in order
// to increase the number of vector instructions generated.
ArrayRef<Instruction *> Left = Chain.slice(0, VF);
ArrayRef<Instruction *> Right = Chain.slice(VF);
return vectorizeLoadChain(Left, InstructionsProcessed) |
vectorizeLoadChain(Right, InstructionsProcessed);
}
if (NewChain.size() == 1) { if (NewChain.size() == 1) {
// Failed after the first instruction. Discard it and try the smaller chain. // Failed after the first instruction. Discard it and try the smaller chain.
InstructionsProcessed->insert(NewChain.front()); InstructionsProcessed->insert(NewChain.front());
return false; return false;
} }
// Update Chain to the valid vectorizable subchain. // Update Chain to the valid vectorizable subchain.
Chain = NewChain; Chain = NewChain;
▲ Show 20 Lines • Show All 135 Lines • Show Last 20 Lines

test/Transforms/LoadStoreVectorizer/AMDGPU/empty-prefix.ll

  • This file was added.
; RUN: opt -mtriple=amdgcn-amd-amdhsa -load-store-vectorizer -S -o - %s | FileCheck %s
; Split the chain and try again if the vectorizable
; prefix is empty. This will vectorize 48 more scalars.
jlebarUnsubmitted
Not Done
ReplyInline Actions

Can we write this comment in the form of

  1. "Check that we X"
  2. Something that does not rely on implementation details of the LSV? That is, can we avoid the term "vectorizable prefix", or, if we must use it, can we define it here?
jlebar: Can we write this comment in the form of 1) "Check that we X" 2) Something that does not rely…
; CHECK-LABEL: foo
; CHECK: load <4 x float>
; CHECK: load <4 x float>
; CHECK: load <4 x float>
; CHECK: load <4 x float>
; CHECK: load <4 x float>
; CHECK: load <4 x float>
; CHECK: load <4 x float>
; CHECK: load <4 x float>
; CHECK: store <4 x float>
; CHECK: store <4 x float>
; CHECK: store <4 x float>
; CHECK: store <4 x float>
; CHECK: store <4 x float>
; CHECK: store <4 x float>
; CHECK: store <4 x float>
; CHECK: store <4 x float>
; CHECK: load <4 x float>
; CHECK: load <4 x float>
; CHECK: load <4 x float>
; CHECK: load <4 x float>
; CHECK: store <4 x float>
; CHECK: store <4 x float>
; CHECK: store <4 x float>
; CHECK: store <4 x float>
; Function Attrs: nounwind
define void @foo(i8 addrspace(1)* %a, i8 addrspace(1)* %b) #0 {
bb:
%tmp = bitcast i8 addrspace(1)* %b to float addrspace(1)*
%tmp1 = load float, float addrspace(1)* %tmp, align 4
%tmp2 = getelementptr i8, i8 addrspace(1)* %b, i64 4
%tmp3 = bitcast i8 addrspace(1)* %tmp2 to float addrspace(1)*
%tmp4 = load float, float addrspace(1)* %tmp3, align 4
%tmp5 = getelementptr i8, i8 addrspace(1)* %b, i64 8
%tmp6 = bitcast i8 addrspace(1)* %tmp5 to float addrspace(1)*
%tmp7 = load float, float addrspace(1)* %tmp6, align 4
%tmp8 = getelementptr i8, i8 addrspace(1)* %b, i64 12
%tmp9 = bitcast i8 addrspace(1)* %tmp8 to float addrspace(1)*
%tmp10 = load float, float addrspace(1)* %tmp9, align 4
%tmp11 = getelementptr i8, i8 addrspace(1)* %b, i64 16
%tmp12 = bitcast i8 addrspace(1)* %tmp11 to float addrspace(1)*
%tmp13 = load float, float addrspace(1)* %tmp12, align 4
%tmp14 = getelementptr i8, i8 addrspace(1)* %b, i64 20
%tmp15 = bitcast i8 addrspace(1)* %tmp14 to float addrspace(1)*
%tmp16 = load float, float addrspace(1)* %tmp15, align 4
%tmp17 = getelementptr i8, i8 addrspace(1)* %b, i64 24
%tmp18 = bitcast i8 addrspace(1)* %tmp17 to float addrspace(1)*
%tmp19 = load float, float addrspace(1)* %tmp18, align 4
%tmp20 = getelementptr i8, i8 addrspace(1)* %b, i64 28
%tmp21 = bitcast i8 addrspace(1)* %tmp20 to float addrspace(1)*
%tmp22 = load float, float addrspace(1)* %tmp21, align 4
%tmp23 = getelementptr i8, i8 addrspace(1)* %b, i64 32
%tmp24 = bitcast i8 addrspace(1)* %tmp23 to float addrspace(1)*
%tmp25 = load float, float addrspace(1)* %tmp24, align 4
%tmp26 = getelementptr i8, i8 addrspace(1)* %b, i64 36
%tmp27 = bitcast i8 addrspace(1)* %tmp26 to float addrspace(1)*
%tmp28 = load float, float addrspace(1)* %tmp27, align 4
%tmp29 = getelementptr i8, i8 addrspace(1)* %b, i64 40
%tmp30 = bitcast i8 addrspace(1)* %tmp29 to float addrspace(1)*
%tmp31 = load float, float addrspace(1)* %tmp30, align 4
%tmp32 = getelementptr i8, i8 addrspace(1)* %b, i64 44
%tmp33 = bitcast i8 addrspace(1)* %tmp32 to float addrspace(1)*
%tmp34 = load float, float addrspace(1)* %tmp33, align 4
%tmp35 = getelementptr i8, i8 addrspace(1)* %b, i64 48
%tmp36 = bitcast i8 addrspace(1)* %tmp35 to float addrspace(1)*
%tmp37 = load float, float addrspace(1)* %tmp36, align 4
%tmp38 = getelementptr i8, i8 addrspace(1)* %b, i64 52
%tmp39 = bitcast i8 addrspace(1)* %tmp38 to float addrspace(1)*
%tmp40 = load float, float addrspace(1)* %tmp39, align 4
%tmp41 = getelementptr i8, i8 addrspace(1)* %b, i64 56
%tmp42 = bitcast i8 addrspace(1)* %tmp41 to float addrspace(1)*
%tmp43 = load float, float addrspace(1)* %tmp42, align 4
%tmp44 = getelementptr i8, i8 addrspace(1)* %b, i64 60
%tmp45 = bitcast i8 addrspace(1)* %tmp44 to float addrspace(1)*
%tmp46 = load float, float addrspace(1)* %tmp45, align 4
%tmp47 = getelementptr i8, i8 addrspace(1)* %b, i64 64
%tmp48 = bitcast i8 addrspace(1)* %tmp47 to float addrspace(1)*
%tmp49 = load float, float addrspace(1)* %tmp48, align 4
%tmp50 = getelementptr i8, i8 addrspace(1)* %b, i64 68
%tmp51 = bitcast i8 addrspace(1)* %tmp50 to float addrspace(1)*
%tmp52 = load float, float addrspace(1)* %tmp51, align 4
%tmp53 = getelementptr i8, i8 addrspace(1)* %b, i64 72
%tmp54 = bitcast i8 addrspace(1)* %tmp53 to float addrspace(1)*
%tmp55 = load float, float addrspace(1)* %tmp54, align 4
%tmp56 = getelementptr i8, i8 addrspace(1)* %b, i64 76
%tmp57 = bitcast i8 addrspace(1)* %tmp56 to float addrspace(1)*
%tmp58 = load float, float addrspace(1)* %tmp57, align 4
%tmp59 = getelementptr i8, i8 addrspace(1)* %b, i64 80
%tmp60 = bitcast i8 addrspace(1)* %tmp59 to float addrspace(1)*
%tmp61 = load float, float addrspace(1)* %tmp60, align 4
%tmp62 = getelementptr i8, i8 addrspace(1)* %b, i64 84
%tmp63 = bitcast i8 addrspace(1)* %tmp62 to float addrspace(1)*
%tmp64 = load float, float addrspace(1)* %tmp63, align 4
%tmp65 = getelementptr i8, i8 addrspace(1)* %b, i64 88
%tmp66 = bitcast i8 addrspace(1)* %tmp65 to float addrspace(1)*
%tmp67 = load float, float addrspace(1)* %tmp66, align 4
%tmp68 = getelementptr i8, i8 addrspace(1)* %b, i64 92
%tmp69 = bitcast i8 addrspace(1)* %tmp68 to float addrspace(1)*
%tmp70 = load float, float addrspace(1)* %tmp69, align 4
%tmp71 = getelementptr i8, i8 addrspace(1)* %b, i64 96
%tmp72 = bitcast i8 addrspace(1)* %tmp71 to float addrspace(1)*
%tmp73 = load float, float addrspace(1)* %tmp72, align 4
%tmp74 = getelementptr i8, i8 addrspace(1)* %b, i64 100
%tmp75 = bitcast i8 addrspace(1)* %tmp74 to float addrspace(1)*
%tmp76 = load float, float addrspace(1)* %tmp75, align 4
%tmp77 = getelementptr i8, i8 addrspace(1)* %b, i64 104
%tmp78 = bitcast i8 addrspace(1)* %tmp77 to float addrspace(1)*
%tmp79 = load float, float addrspace(1)* %tmp78, align 4
%tmp80 = getelementptr i8, i8 addrspace(1)* %b, i64 108
%tmp81 = bitcast i8 addrspace(1)* %tmp80 to float addrspace(1)*
%tmp82 = load float, float addrspace(1)* %tmp81, align 4
%tmp83 = getelementptr i8, i8 addrspace(1)* %b, i64 112
%tmp84 = bitcast i8 addrspace(1)* %tmp83 to float addrspace(1)*
%tmp85 = load float, float addrspace(1)* %tmp84, align 4
%tmp86 = getelementptr i8, i8 addrspace(1)* %b, i64 116
%tmp87 = bitcast i8 addrspace(1)* %tmp86 to float addrspace(1)*
%tmp88 = load float, float addrspace(1)* %tmp87, align 4
%tmp89 = getelementptr i8, i8 addrspace(1)* %b, i64 120
%tmp90 = bitcast i8 addrspace(1)* %tmp89 to float addrspace(1)*
%tmp91 = load float, float addrspace(1)* %tmp90, align 4
%tmp92 = getelementptr i8, i8 addrspace(1)* %b, i64 124
%tmp93 = bitcast i8 addrspace(1)* %tmp92 to float addrspace(1)*
%tmp94 = load float, float addrspace(1)* %tmp93, align 4
%tmp95 = bitcast i8 addrspace(1)* %a to float addrspace(1)*
store float %tmp1, float addrspace(1)* %tmp95, align 4
%tmp96 = getelementptr i8, i8 addrspace(1)* %a, i64 4
%tmp97 = bitcast i8 addrspace(1)* %tmp96 to float addrspace(1)*
store float %tmp4, float addrspace(1)* %tmp97, align 4
%tmp98 = getelementptr i8, i8 addrspace(1)* %a, i64 8
%tmp99 = bitcast i8 addrspace(1)* %tmp98 to float addrspace(1)*
store float %tmp7, float addrspace(1)* %tmp99, align 4
%tmp100 = getelementptr i8, i8 addrspace(1)* %a, i64 12
%tmp101 = bitcast i8 addrspace(1)* %tmp100 to float addrspace(1)*
store float %tmp10, float addrspace(1)* %tmp101, align 4
%tmp102 = getelementptr i8, i8 addrspace(1)* %a, i64 16
%tmp103 = bitcast i8 addrspace(1)* %tmp102 to float addrspace(1)*
store float %tmp13, float addrspace(1)* %tmp103, align 4
%tmp104 = getelementptr i8, i8 addrspace(1)* %a, i64 20
%tmp105 = bitcast i8 addrspace(1)* %tmp104 to float addrspace(1)*
store float %tmp16, float addrspace(1)* %tmp105, align 4
%tmp106 = getelementptr i8, i8 addrspace(1)* %a, i64 24
%tmp107 = bitcast i8 addrspace(1)* %tmp106 to float addrspace(1)*
store float %tmp19, float addrspace(1)* %tmp107, align 4
%tmp108 = getelementptr i8, i8 addrspace(1)* %a, i64 28
%tmp109 = bitcast i8 addrspace(1)* %tmp108 to float addrspace(1)*
store float %tmp22, float addrspace(1)* %tmp109, align 4
%tmp110 = getelementptr i8, i8 addrspace(1)* %a, i64 32
%tmp111 = bitcast i8 addrspace(1)* %tmp110 to float addrspace(1)*
store float %tmp25, float addrspace(1)* %tmp111, align 4
%tmp112 = getelementptr i8, i8 addrspace(1)* %a, i64 36
%tmp113 = bitcast i8 addrspace(1)* %tmp112 to float addrspace(1)*
store float %tmp28, float addrspace(1)* %tmp113, align 4
%tmp114 = getelementptr i8, i8 addrspace(1)* %a, i64 40
%tmp115 = bitcast i8 addrspace(1)* %tmp114 to float addrspace(1)*
store float %tmp31, float addrspace(1)* %tmp115, align 4
%tmp116 = getelementptr i8, i8 addrspace(1)* %a, i64 44
%tmp117 = bitcast i8 addrspace(1)* %tmp116 to float addrspace(1)*
store float %tmp34, float addrspace(1)* %tmp117, align 4
%tmp118 = getelementptr i8, i8 addrspace(1)* %a, i64 48
%tmp119 = bitcast i8 addrspace(1)* %tmp118 to float addrspace(1)*
store float %tmp37, float addrspace(1)* %tmp119, align 4
%tmp120 = getelementptr i8, i8 addrspace(1)* %a, i64 52
%tmp121 = bitcast i8 addrspace(1)* %tmp120 to float addrspace(1)*
store float %tmp40, float addrspace(1)* %tmp121, align 4
%tmp122 = getelementptr i8, i8 addrspace(1)* %a, i64 56
%tmp123 = bitcast i8 addrspace(1)* %tmp122 to float addrspace(1)*
store float %tmp43, float addrspace(1)* %tmp123, align 4
%tmp124 = getelementptr i8, i8 addrspace(1)* %a, i64 60
%tmp125 = bitcast i8 addrspace(1)* %tmp124 to float addrspace(1)*
store float %tmp46, float addrspace(1)* %tmp125, align 4
%tmp126 = getelementptr i8, i8 addrspace(1)* %a, i64 64
%tmp127 = bitcast i8 addrspace(1)* %tmp126 to float addrspace(1)*
store float %tmp49, float addrspace(1)* %tmp127, align 4
%tmp128 = getelementptr i8, i8 addrspace(1)* %a, i64 68
%tmp129 = bitcast i8 addrspace(1)* %tmp128 to float addrspace(1)*
store float %tmp52, float addrspace(1)* %tmp129, align 4
%tmp130 = getelementptr i8, i8 addrspace(1)* %a, i64 72
%tmp131 = bitcast i8 addrspace(1)* %tmp130 to float addrspace(1)*
store float %tmp55, float addrspace(1)* %tmp131, align 4
%tmp132 = getelementptr i8, i8 addrspace(1)* %a, i64 76
%tmp133 = bitcast i8 addrspace(1)* %tmp132 to float addrspace(1)*
store float %tmp58, float addrspace(1)* %tmp133, align 4
%tmp134 = getelementptr i8, i8 addrspace(1)* %a, i64 80
%tmp135 = bitcast i8 addrspace(1)* %tmp134 to float addrspace(1)*
store float %tmp61, float addrspace(1)* %tmp135, align 4
%tmp136 = getelementptr i8, i8 addrspace(1)* %a, i64 84
%tmp137 = bitcast i8 addrspace(1)* %tmp136 to float addrspace(1)*
store float %tmp64, float addrspace(1)* %tmp137, align 4
%tmp138 = getelementptr i8, i8 addrspace(1)* %a, i64 88
%tmp139 = bitcast i8 addrspace(1)* %tmp138 to float addrspace(1)*
store float %tmp67, float addrspace(1)* %tmp139, align 4
%tmp140 = getelementptr i8, i8 addrspace(1)* %a, i64 92
%tmp141 = bitcast i8 addrspace(1)* %tmp140 to float addrspace(1)*
store float %tmp70, float addrspace(1)* %tmp141, align 4
%tmp142 = getelementptr i8, i8 addrspace(1)* %a, i64 96
%tmp143 = bitcast i8 addrspace(1)* %tmp142 to float addrspace(1)*
store float %tmp73, float addrspace(1)* %tmp143, align 4
%tmp144 = getelementptr i8, i8 addrspace(1)* %a, i64 100
%tmp145 = bitcast i8 addrspace(1)* %tmp144 to float addrspace(1)*
store float %tmp76, float addrspace(1)* %tmp145, align 4
%tmp146 = getelementptr i8, i8 addrspace(1)* %a, i64 104
%tmp147 = bitcast i8 addrspace(1)* %tmp146 to float addrspace(1)*
store float %tmp79, float addrspace(1)* %tmp147, align 4
%tmp148 = getelementptr i8, i8 addrspace(1)* %a, i64 108
%tmp149 = bitcast i8 addrspace(1)* %tmp148 to float addrspace(1)*
store float %tmp82, float addrspace(1)* %tmp149, align 4
%tmp150 = getelementptr i8, i8 addrspace(1)* %a, i64 112
%tmp151 = bitcast i8 addrspace(1)* %tmp150 to float addrspace(1)*
store float %tmp85, float addrspace(1)* %tmp151, align 4
%tmp152 = getelementptr i8, i8 addrspace(1)* %a, i64 116
%tmp153 = bitcast i8 addrspace(1)* %tmp152 to float addrspace(1)*
store float %tmp88, float addrspace(1)* %tmp153, align 4
%tmp154 = getelementptr i8, i8 addrspace(1)* %a, i64 120
%tmp155 = bitcast i8 addrspace(1)* %tmp154 to float addrspace(1)*
store float %tmp91, float addrspace(1)* %tmp155, align 4
%tmp156 = getelementptr i8, i8 addrspace(1)* %a, i64 124
%tmp157 = bitcast i8 addrspace(1)* %tmp156 to float addrspace(1)*
store float %tmp94, float addrspace(1)* %tmp157, align 4
%tmp158 = getelementptr i8, i8 addrspace(1)* %b, i64 128
%tmp159 = bitcast i8 addrspace(1)* %tmp158 to float addrspace(1)*
%tmp160 = load float, float addrspace(1)* %tmp159, align 4
%tmp161 = getelementptr i8, i8 addrspace(1)* %b, i64 132
%tmp162 = bitcast i8 addrspace(1)* %tmp161 to float addrspace(1)*
%tmp163 = load float, float addrspace(1)* %tmp162, align 4
%tmp164 = getelementptr i8, i8 addrspace(1)* %b, i64 136
%tmp165 = bitcast i8 addrspace(1)* %tmp164 to float addrspace(1)*
%tmp166 = load float, float addrspace(1)* %tmp165, align 4
%tmp167 = getelementptr i8, i8 addrspace(1)* %b, i64 140
%tmp168 = bitcast i8 addrspace(1)* %tmp167 to float addrspace(1)*
%tmp169 = load float, float addrspace(1)* %tmp168, align 4
%tmp170 = getelementptr i8, i8 addrspace(1)* %b, i64 144
%tmp171 = bitcast i8 addrspace(1)* %tmp170 to float addrspace(1)*
%tmp172 = load float, float addrspace(1)* %tmp171, align 4
%tmp173 = getelementptr i8, i8 addrspace(1)* %b, i64 148
%tmp174 = bitcast i8 addrspace(1)* %tmp173 to float addrspace(1)*
%tmp175 = load float, float addrspace(1)* %tmp174, align 4
%tmp176 = getelementptr i8, i8 addrspace(1)* %b, i64 152
%tmp177 = bitcast i8 addrspace(1)* %tmp176 to float addrspace(1)*
%tmp178 = load float, float addrspace(1)* %tmp177, align 4
%tmp179 = getelementptr i8, i8 addrspace(1)* %b, i64 156
%tmp180 = bitcast i8 addrspace(1)* %tmp179 to float addrspace(1)*
%tmp181 = load float, float addrspace(1)* %tmp180, align 4
%tmp182 = getelementptr i8, i8 addrspace(1)* %b, i64 160
%tmp183 = bitcast i8 addrspace(1)* %tmp182 to float addrspace(1)*
%tmp184 = load float, float addrspace(1)* %tmp183, align 4
%tmp185 = getelementptr i8, i8 addrspace(1)* %b, i64 164
%tmp186 = bitcast i8 addrspace(1)* %tmp185 to float addrspace(1)*
%tmp187 = load float, float addrspace(1)* %tmp186, align 4
%tmp188 = getelementptr i8, i8 addrspace(1)* %b, i64 168
%tmp189 = bitcast i8 addrspace(1)* %tmp188 to float addrspace(1)*
%tmp190 = load float, float addrspace(1)* %tmp189, align 4
%tmp191 = getelementptr i8, i8 addrspace(1)* %b, i64 172
%tmp192 = bitcast i8 addrspace(1)* %tmp191 to float addrspace(1)*
%tmp193 = load float, float addrspace(1)* %tmp192, align 4
%tmp194 = getelementptr i8, i8 addrspace(1)* %b, i64 176
%tmp195 = bitcast i8 addrspace(1)* %tmp194 to float addrspace(1)*
%tmp196 = load float, float addrspace(1)* %tmp195, align 4
%tmp197 = getelementptr i8, i8 addrspace(1)* %b, i64 180
%tmp198 = bitcast i8 addrspace(1)* %tmp197 to float addrspace(1)*
%tmp199 = load float, float addrspace(1)* %tmp198, align 4
%tmp200 = getelementptr i8, i8 addrspace(1)* %b, i64 184
%tmp201 = bitcast i8 addrspace(1)* %tmp200 to float addrspace(1)*
%tmp202 = load float, float addrspace(1)* %tmp201, align 4
%tmp203 = getelementptr i8, i8 addrspace(1)* %b, i64 188
%tmp204 = bitcast i8 addrspace(1)* %tmp203 to float addrspace(1)*
%tmp205 = load float, float addrspace(1)* %tmp204, align 4
%tmp206 = getelementptr i8, i8 addrspace(1)* %b, i64 192
%tmp207 = bitcast i8 addrspace(1)* %tmp206 to half addrspace(1)*
%tmp208 = load half, half addrspace(1)* %tmp207, align 2
%tmp209 = getelementptr i8, i8 addrspace(1)* %a, i64 128
%tmp210 = bitcast i8 addrspace(1)* %tmp209 to float addrspace(1)*
store float %tmp160, float addrspace(1)* %tmp210, align 4
%tmp211 = getelementptr i8, i8 addrspace(1)* %a, i64 132
%tmp212 = bitcast i8 addrspace(1)* %tmp211 to float addrspace(1)*
store float %tmp163, float addrspace(1)* %tmp212, align 4
%tmp213 = getelementptr i8, i8 addrspace(1)* %a, i64 136
%tmp214 = bitcast i8 addrspace(1)* %tmp213 to float addrspace(1)*
store float %tmp166, float addrspace(1)* %tmp214, align 4
%tmp215 = getelementptr i8, i8 addrspace(1)* %a, i64 140
%tmp216 = bitcast i8 addrspace(1)* %tmp215 to float addrspace(1)*
store float %tmp169, float addrspace(1)* %tmp216, align 4
%tmp217 = getelementptr i8, i8 addrspace(1)* %a, i64 144
%tmp218 = bitcast i8 addrspace(1)* %tmp217 to float addrspace(1)*
store float %tmp172, float addrspace(1)* %tmp218, align 4
%tmp219 = getelementptr i8, i8 addrspace(1)* %a, i64 148
%tmp220 = bitcast i8 addrspace(1)* %tmp219 to float addrspace(1)*
store float %tmp175, float addrspace(1)* %tmp220, align 4
%tmp221 = getelementptr i8, i8 addrspace(1)* %a, i64 152
%tmp222 = bitcast i8 addrspace(1)* %tmp221 to float addrspace(1)*
store float %tmp178, float addrspace(1)* %tmp222, align 4
%tmp223 = getelementptr i8, i8 addrspace(1)* %a, i64 156
%tmp224 = bitcast i8 addrspace(1)* %tmp223 to float addrspace(1)*
store float %tmp181, float addrspace(1)* %tmp224, align 4
%tmp225 = getelementptr i8, i8 addrspace(1)* %a, i64 160
%tmp226 = bitcast i8 addrspace(1)* %tmp225 to float addrspace(1)*
store float %tmp184, float addrspace(1)* %tmp226, align 4
%tmp227 = getelementptr i8, i8 addrspace(1)* %a, i64 164
%tmp228 = bitcast i8 addrspace(1)* %tmp227 to float addrspace(1)*
store float %tmp187, float addrspace(1)* %tmp228, align 4
%tmp229 = getelementptr i8, i8 addrspace(1)* %a, i64 168
%tmp230 = bitcast i8 addrspace(1)* %tmp229 to float addrspace(1)*
store float %tmp190, float addrspace(1)* %tmp230, align 4
%tmp231 = getelementptr i8, i8 addrspace(1)* %a, i64 172
%tmp232 = bitcast i8 addrspace(1)* %tmp231 to float addrspace(1)*
store float %tmp193, float addrspace(1)* %tmp232, align 4
%tmp233 = getelementptr i8, i8 addrspace(1)* %a, i64 176
%tmp234 = bitcast i8 addrspace(1)* %tmp233 to float addrspace(1)*
store float %tmp196, float addrspace(1)* %tmp234, align 4
%tmp235 = getelementptr i8, i8 addrspace(1)* %a, i64 180
%tmp236 = bitcast i8 addrspace(1)* %tmp235 to float addrspace(1)*
store float %tmp199, float addrspace(1)* %tmp236, align 4
%tmp237 = getelementptr i8, i8 addrspace(1)* %a, i64 184
%tmp238 = bitcast i8 addrspace(1)* %tmp237 to float addrspace(1)*
store float %tmp202, float addrspace(1)* %tmp238, align 4
%tmp239 = getelementptr i8, i8 addrspace(1)* %a, i64 188
%tmp240 = bitcast i8 addrspace(1)* %tmp239 to float addrspace(1)*
store float %tmp205, float addrspace(1)* %tmp240, align 4
%tmp241 = getelementptr i8, i8 addrspace(1)* %a, i64 192
%tmp242 = bitcast i8 addrspace(1)* %tmp241 to half addrspace(1)*
store half %tmp208, half addrspace(1)* %tmp242, align 2
ret void
}
attributes #0 = { argmemonly nounwind }