[LSR] Complexity tradeoff in GenerateReassociationsImpl

[mkazantsev]

Summary

During its work, GenerateReassociationsImpl may end up trying calculating a sum
of a huge amount of huge SCEVs. Attempts to simplify them lead to huge compile
time problem. This patch applies a heuristic that limits simplification recursion depth
depending on number of arguments in this method to get an acceptible compile time
in corner cases while it should not affect most of real situations.

[mkazantsev]

Ping

[sanjoy]

lib/Transforms/Scalar/LoopStrengthReduce.cpp

3497

This worries me from a design perspective -- the Depth parameter definitely should not be part of the public API of SCEV (I had not noticed that we've added these parameters to the public API). We can't guarantee its meaning or even its presence to passes that use SCEV.

Can we instead make SCEV smarter to avoid spending a lot of cycles in pathological cases like this?

[mkazantsev]

The other way is limiting SCEV transformation not by depth only, but also by the size of operands list, if it makes sense.

[sanjoy]

In D40718#964316, @mkazantsev wrote:

The other way is limiting SCEV transformation not by depth only, but also by the size of operands list, if it makes sense.

I think your current approach is fine (penalize depth for high "breadth"), but it just needs to somehow happen within SCEV (so that "depth" remains a SCEV implementation detail).

lib/Transforms/Scalar/LoopStrengthReduce.cpp

for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
 
     // Don't leave just a constant behind in a register if the constant could
     // be folded into an immediate field.
     if (InnerAddOps.size() == 1 &&
         isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
                          LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
       continue;
 
-    const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
+    // In case if InnerAddOps has a lot of operands, and if they are huge, SCEV
+    // may spend a really long time trying to simplify it. Here we compromise
+    // between the number of operands and depth of recursive simplification to
+    // be done for them. In most real cases, the starting Depth is effectively
+    // zero, and nothing changes for them. But in corner cases (with hundreds of
+    // operands in InnerAddOps) it may save us from being stalled.
+    // Function 'Depth = InnerAddOps.size() / 4' is chosen arbitrarily and may
+    // be reconsidered in future.
+    unsigned Depth = InnerAddOps.size() / 4;
+    const SCEV *InnerSum = SE.getAddExpr(InnerAddOps, SCEV::FlagAnyWrap, Depth);
     if (InnerSum->isZero())
       continue;
     Formula F = Base;
 
     // Add the remaining pieces of the add back into the new formula.
     const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
     if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
         TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +

test/Transforms/LoopStrengthReduce/2017-12-01-long_compilation_corner_case.ll

+; RUN: opt < %s -loop-reduce -S | FileCheck %s
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128-ni:1"
+target triple = "x86_64-unknown-linux-gnu"
+
+; Just make sure that this test can be compiled within reasonable timeframe.
+; If it stalls, then we have an exponential behavior with huge SCEVs.
+
+define void @test(i8* %p) {
+
+; CHECK-LABEL: @test(
+
+entry:
+  %var.addr37 = getelementptr inbounds i8, i8* %p, i64 136
+  %var.addr.typed38 = bitcast i8* %var.addr37 to i32*
+  %var39 = load i32, i32* %var.addr.typed38, align 4
+  %var.addr115 = getelementptr inbounds i8, i8* %p, i64 392
+  %var.addr.typed116 = bitcast i8* %var.addr115 to i32*
+  %var117 = load i32, i32* %var.addr.typed116, align 4
+  %var.addr118 = getelementptr inbounds i8, i8* %p, i64 400
+  %var.addr.typed119 = bitcast i8* %var.addr118 to i32*
+  %var120 = load i32, i32* %var.addr.typed119, align 4
+  %var.addr130 = getelementptr inbounds i8, i8* %p, i64 432
+  %var.addr.typed131 = bitcast i8* %var.addr130 to i32*
+  %var132 = load i32, i32* %var.addr.typed131, align 4
+  %var.addr139 = getelementptr inbounds i8, i8* %p, i64 456
+  %tmp0 = bitcast i8* %var.addr139 to i32*
+  %var1411 = load i32, i32* %tmp0, align 4
+  %var.addr145 = getelementptr inbounds i8, i8* %p, i64 472
+  %var.addr.typed146 = bitcast i8* %var.addr145 to i32*
+  %var147 = load i32, i32* %var.addr.typed146, align 4
+  %var.addr148 = getelementptr inbounds i8, i8* %p, i64 488
+  %var.addr.typed149 = bitcast i8* %var.addr148 to i64*
+  %var150 = load i64, i64* %var.addr.typed149, align 8
+  %var.addr151 = getelementptr inbounds i8, i8* %p, i64 496
+  %var.addr.typed152 = bitcast i8* %var.addr151 to i32*
+  %var153 = load i32, i32* %var.addr.typed152, align 4
+  %var.addr154 = getelementptr inbounds i8, i8* %p, i64 512
+  %var.addr.typed155 = bitcast i8* %var.addr154 to i64*
+  %var156 = load i64, i64* %var.addr.typed155, align 8
+  %var.addr157 = getelementptr inbounds i8, i8* %p, i64 528
+  %var.addr.typed158 = bitcast i8* %var.addr157 to i64*
+  %var159 = load i64, i64* %var.addr.typed158, align 8
+  %var.addr160 = getelementptr inbounds i8, i8* %p, i64 536
+  %var.addr.typed161 = bitcast i8* %var.addr160 to i32*
+  %var162 = load i32, i32* %var.addr.typed161, align 4
+  %var.addr163 = getelementptr inbounds i8, i8* %p, i64 544
+  %var.addr.typed164 = bitcast i8* %var.addr163 to i32*
+  %var165 = load i32, i32* %var.addr.typed164, align 4
+  %var.addr169 = getelementptr inbounds i8, i8* %p, i64 568
+  %tmp1 = bitcast i8* %var.addr169 to i64*
+  %var1712 = load i64, i64* %tmp1, align 8
+  %var.addr172 = getelementptr inbounds i8, i8* %p, i64 584
+  %var.addr.typed173 = bitcast i8* %var.addr172 to i64*
+  %var174 = load i64, i64* %var.addr.typed173, align 8
+  %var.addr175 = getelementptr inbounds i8, i8* %p, i64 592
+  %var.addr.typed176 = bitcast i8* %var.addr175 to i32*
+  %var177 = load i32, i32* %var.addr.typed176, align 4
+  %var.addr178 = getelementptr inbounds i8, i8* %p, i64 600
+  %var.addr.typed179 = bitcast i8* %var.addr178 to i32*
+  %var180 = load i32, i32* %var.addr.typed179, align 4
+  %var.addr181 = getelementptr inbounds i8, i8* %p, i64 608
+  %var.addr.typed182 = bitcast i8* %var.addr181 to i32*
+  %var183 = load i32, i32* %var.addr.typed182, align 4
+  %var.addr184 = getelementptr inbounds i8, i8* %p, i64 616
+  %tmp2 = bitcast i8* %var.addr184 to i32*
+  %var1863 = load i32, i32* %tmp2, align 4
+  %var.addr187 = getelementptr inbounds i8, i8* %p, i64 632
+  %var.addr.typed188 = bitcast i8* %var.addr187 to i64*
+  %var189 = load i64, i64* %var.addr.typed188, align 8
+  %var.addr190 = getelementptr inbounds i8, i8* %p, i64 640
+  %var.addr.typed191 = bitcast i8* %var.addr190 to i32*
+  %var192 = load i32, i32* %var.addr.typed191, align 4
+  %var.addr193 = getelementptr inbounds i8, i8* %p, i64 656
+  %var.addr.typed194 = bitcast i8* %var.addr193 to i64*
+  %var195 = load i64, i64* %var.addr.typed194, align 8
+  %var.addr196 = getelementptr inbounds i8, i8* %p, i64 664
+  %var.addr.typed197 = bitcast i8* %var.addr196 to i32*
+  %var198 = load i32, i32* %var.addr.typed197, align 4
+  %var.addr199 = getelementptr inbounds i8, i8* %p, i64 680
+  %var.addr.typed200 = bitcast i8* %var.addr199 to i64*
+  %var201 = load i64, i64* %var.addr.typed200, align 8
+  %var.addr202 = getelementptr inbounds i8, i8* %p, i64 688
+  %var.addr.typed203 = bitcast i8* %var.addr202 to i32*
+  %var204 = load i32, i32* %var.addr.typed203, align 4
+  %var.addr205 = getelementptr inbounds i8, i8* %p, i64 696
+  %var.addr.typed206 = bitcast i8* %var.addr205 to i32*
+  %var207 = load i32, i32* %var.addr.typed206, align 4
+  %var.addr208 = getelementptr inbounds i8, i8* %p, i64 704
+  %var.addr.typed209 = bitcast i8* %var.addr208 to i32*
+  %var210 = load i32, i32* %var.addr.typed209, align 4
+  %var.addr211 = getelementptr inbounds i8, i8* %p, i64 720
+  %tmp3 = bitcast i8* %var.addr211 to i64*
+  %var2134 = load i64, i64* %tmp3, align 8
+  %var.addr217 = getelementptr inbounds i8, i8* %p, i64 736
+  %tmp4 = bitcast i8* %var.addr217 to i32*
+  %var2195 = load i32, i32* %tmp4, align 4
+  %var.addr220 = getelementptr inbounds i8, i8* %p, i64 744
+  %var.addr.typed221 = bitcast i8* %var.addr220 to i32*
+  %var222 = load i32, i32* %var.addr.typed221, align 4
+  %var.addr223 = getelementptr inbounds i8, i8* %p, i64 752
+  %var.addr.typed224 = bitcast i8* %var.addr223 to i32*
+  %var225 = load i32, i32* %var.addr.typed224, align 4
+  %var.addr226 = getelementptr inbounds i8, i8* %p, i64 760
+  %var.addr.typed227 = bitcast i8* %var.addr226 to i32*
+  %var228 = load i32, i32* %var.addr.typed227, align 4
+  %var.addr229 = getelementptr inbounds i8, i8* %p, i64 768
+  %var.addr.typed230 = bitcast i8* %var.addr229 to i32*
+  %var231 = load i32, i32* %var.addr.typed230, align 4
+  %var.addr232 = getelementptr inbounds i8, i8* %p, i64 776
+  %var.addr.typed233 = bitcast i8* %var.addr232 to i32*
+  %var234 = load i32, i32* %var.addr.typed233, align 4
+  %var.addr235 = getelementptr inbounds i8, i8* %p, i64 792
+  %tmp5 = bitcast i8* %var.addr235 to i64*
+  %var2376 = load i64, i64* %tmp5, align 8
+  %var.addr238 = getelementptr inbounds i8, i8* %p, i64 800
+  %var.addr.typed239 = bitcast i8* %var.addr238 to i32*
+  %var240 = load i32, i32* %var.addr.typed239, align 4
+  %var.addr241 = getelementptr inbounds i8, i8* %p, i64 808
+  %var.addr.typed242 = bitcast i8* %var.addr241 to i32*
+  %var243 = load i32, i32* %var.addr.typed242, align 4
+  %var.addr244 = getelementptr inbounds i8, i8* %p, i64 824
+  %var.addr.typed245 = bitcast i8* %var.addr244 to i64*
+  %var246 = load i64, i64* %var.addr.typed245, align 8
+  %var.addr247 = getelementptr inbounds i8, i8* %p, i64 832
+  %var.addr.typed248 = bitcast i8* %var.addr247 to i32*
+  %var249 = load i32, i32* %var.addr.typed248, align 4
+  %var.addr250 = getelementptr inbounds i8, i8* %p, i64 840
+  %var.addr.typed251 = bitcast i8* %var.addr250 to i32*
+  %var252 = load i32, i32* %var.addr.typed251, align 4
+  %var.addr253 = getelementptr inbounds i8, i8* %p, i64 848
+  %var.addr.typed254 = bitcast i8* %var.addr253 to i32*
+  %var255 = load i32, i32* %var.addr.typed254, align 4
+  %var.addr256 = getelementptr inbounds i8, i8* %p, i64 856
+  %tmp6 = bitcast i8* %var.addr256 to i32*
+  %var2587 = load i32, i32* %tmp6, align 4
+  %var.addr259 = getelementptr inbounds i8, i8* %p, i64 864
+  %var.addr.typed260 = bitcast i8* %var.addr259 to i32*
+  %var261 = load i32, i32* %var.addr.typed260, align 4
+  %var.addr265 = getelementptr inbounds i8, i8* %p, i64 888
+  %tmp7 = bitcast i8* %var.addr265 to i64*
+  %var2678 = load i64, i64* %tmp7, align 8
+  %var.addr268 = getelementptr inbounds i8, i8* %p, i64 904
+  %var.addr.typed269 = bitcast i8* %var.addr268 to i64*
+  %var270 = load i64, i64* %var.addr.typed269, align 8
+  %var.addr271 = getelementptr inbounds i8, i8* %p, i64 912
+  %var.addr.typed272 = bitcast i8* %var.addr271 to i32*
+  %var273 = load i32, i32* %var.addr.typed272, align 4
+  %var.addr277 = getelementptr inbounds i8, i8* %p, i64 928
+  %var.addr.typed278 = bitcast i8* %var.addr277 to i32*
+  %var279 = load i32, i32* %var.addr.typed278, align 4
+  %var.addr280 = getelementptr inbounds i8, i8* %p, i64 936
+  %tmp8 = bitcast i8* %var.addr280 to i32*
+  %var2829 = load i32, i32* %tmp8, align 4
+  %var.addr283 = getelementptr inbounds i8, i8* %p, i64 944
+  %var.addr.typed284 = bitcast i8* %var.addr283 to i32*
+  %var285 = load i32, i32* %var.addr.typed284, align 4
+  %var.addr286 = getelementptr inbounds i8, i8* %p, i64 952
+  %var.addr.typed287 = bitcast i8* %var.addr286 to i32*
+  %var288 = load i32, i32* %var.addr.typed287, align 4
+  %var.addr289 = getelementptr inbounds i8, i8* %p, i64 960
+  %var.addr.typed290 = bitcast i8* %var.addr289 to i32*
+  %var291 = load i32, i32* %var.addr.typed290, align 4
+  %var.addr295 = getelementptr inbounds i8, i8* %p, i64 976
+  %var.addr.typed296 = bitcast i8* %var.addr295 to i32*
+  %var297 = load i32, i32* %var.addr.typed296, align 4
+  %var.addr298 = getelementptr inbounds i8, i8* %p, i64 984
+  %tmp9 = bitcast i8* %var.addr298 to i32*
+  %var30010 = load i32, i32* %tmp9, align 4
+  %var.addr301 = getelementptr inbounds i8, i8* %p, i64 992
+  %var.addr.typed302 = bitcast i8* %var.addr301 to i32*
+  %var303 = load i32, i32* %var.addr.typed302, align 4
+  %var.addr307 = getelementptr inbounds i8, i8* %p, i64 1008
+  %var.addr.typed308 = bitcast i8* %var.addr307 to i32*
+  %var309 = load i32, i32* %var.addr.typed308, align 4
+  %var.addr310 = getelementptr inbounds i8, i8* %p, i64 1016
+  %var.addr.typed311 = bitcast i8* %var.addr310 to i32*
+  %var312 = load i32, i32* %var.addr.typed311, align 4
+  %var.addr313 = getelementptr inbounds i8, i8* %p, i64 1024
+  %var.addr.typed314 = bitcast i8* %var.addr313 to i32*
+  %var315 = load i32, i32* %var.addr.typed314, align 4
+  %var.addr316 = getelementptr inbounds i8, i8* %p, i64 1032
+  %var.addr.typed317 = bitcast i8* %var.addr316 to i32*
+  %var318 = load i32, i32* %var.addr.typed317, align 4
+  %var.addr319 = getelementptr inbounds i8, i8* %p, i64 1040
+  %var.addr.typed320 = bitcast i8* %var.addr319 to i32*
+  %var321 = load i32, i32* %var.addr.typed320, align 4
+  %var.addr322 = getelementptr inbounds i8, i8* %p, i64 1048
+  %var.addr.typed323 = bitcast i8* %var.addr322 to i32*
+  %var324 = load i32, i32* %var.addr.typed323, align 4
+  %var.addr325 = getelementptr inbounds i8, i8* %p, i64 1056
+  %var.addr.typed326 = bitcast i8* %var.addr325 to i32*
+  %var327 = load i32, i32* %var.addr.typed326, align 4
+  %var.addr328 = getelementptr inbounds i8, i8* %p, i64 1064
+  %var.addr.typed329 = bitcast i8* %var.addr328 to i32*
+  %var330 = load i32, i32* %var.addr.typed329, align 4
+  %var.addr331 = getelementptr inbounds i8, i8* %p, i64 1072
+  %var.addr.typed332 = bitcast i8* %var.addr331 to i32*
+  %var333 = load i32, i32* %var.addr.typed332, align 4
+  %var.addr334 = getelementptr inbounds i8, i8* %p, i64 1080
+  %var.addr.typed335 = bitcast i8* %var.addr334 to i32*
+  %var336 = load i32, i32* %var.addr.typed335, align 4
+  %var.addr337 = getelementptr inbounds i8, i8* %p, i64 1088
+  %var.addr.typed338 = bitcast i8* %var.addr337 to i32*
+  %var339 = load i32, i32* %var.addr.typed338, align 4
+  %var.addr340 = getelementptr inbounds i8, i8* %p, i64 1096
+  %var.addr.typed341 = bitcast i8* %var.addr340 to i32*
+  %var342 = load i32, i32* %var.addr.typed341, align 4
+  %var.addr343 = getelementptr inbounds i8, i8* %p, i64 1104
+  %var.addr.typed344 = bitcast i8* %var.addr343 to i32*
+  %var345 = load i32, i32* %var.addr.typed344, align 4
+  %var.addr346 = getelementptr inbounds i8, i8* %p, i64 1112
+  %var.addr.typed347 = bitcast i8* %var.addr346 to i32*
+  %var348 = load i32, i32* %var.addr.typed347, align 4
+  %var.addr349 = getelementptr inbounds i8, i8* %p, i64 1120
+  %var.addr.typed350 = bitcast i8* %var.addr349 to i32*
+  %var351 = load i32, i32* %var.addr.typed350, align 4
+  %var.addr352 = getelementptr inbounds i8, i8* %p, i64 1136
+  %var.addr.typed353 = bitcast i8* %var.addr352 to i64*
+  %var354 = load i64, i64* %var.addr.typed353, align 8
+  %var.addr355 = getelementptr inbounds i8, i8* %p, i64 1152
+  %var.addr.typed356 = bitcast i8* %var.addr355 to i64*
+  %var357 = load i64, i64* %var.addr.typed356, align 8
+  %var.addr358 = getelementptr inbounds i8, i8* %p, i64 1168
+  %var.addr.typed359 = bitcast i8* %var.addr358 to i64*
+  %var360 = load i64, i64* %var.addr.typed359, align 8
+  %var.addr361 = getelementptr inbounds i8, i8* %p, i64 1184
+  %var.addr.typed362 = bitcast i8* %var.addr361 to i64*
+  %var363 = load i64, i64* %var.addr.typed362, align 8
+  %var.addr370 = getelementptr inbounds i8, i8* %p, i64 1208
+  %tmp10 = bitcast i8* %var.addr370 to i32*
+  %var37211 = load i32, i32* %tmp10, align 4
+  %var.addr373 = getelementptr inbounds i8, i8* %p, i64 1216
+  %tmp11 = bitcast i8* %var.addr373 to i32*
+  %var37512 = load i32, i32* %tmp11, align 4
+  %var.addr376 = getelementptr inbounds i8, i8* %p, i64 1224
+  %tmp12 = bitcast i8* %var.addr376 to i32*
+  %var3782 = load i32, i32* %tmp12, align 4
+  %var.addr382 = getelementptr inbounds i8, i8* %p, i64 1240
+  %var.addr.typed383 = bitcast i8* %var.addr382 to i32*
+  %var384 = load i32, i32* %var.addr.typed383, align 4
+  br label %loop
+
+loop:                                        ; preds = %backedge, %entry
+  %phi_124_486 = phi i32 [ %var303, %entry ], [ %tmp13, %backedge ]
+  %phi_125_487 = phi i64 [ 0, %entry ], [ 2, %backedge ]
+  %tmp13 = add i32 %phi_124_486, 1
+  %tmp14 = icmp sgt i32 %tmp13, 117
+  br i1 %tmp14, label %exit, label %backedge
+
+exit:                                        ; preds = %loop
+  %tmp15 = sext i32 %var39 to i64
+  %tmp16 = sext i32 %var117 to i64
+  %tmp17 = sext i32 %var120 to i64
+  %tmp18 = sext i32 %var132 to i64
+  %tmp19 = sext i32 %var1411 to i64
+  %tmp20 = sext i32 %var147 to i64
+  %tmp21 = sext i32 %var153 to i64
+  %tmp22 = sext i32 %var162 to i64
+  %tmp23 = sext i32 %var165 to i64
+  %tmp24 = sext i32 %var177 to i64
+  %tmp25 = sext i32 %var180 to i64
+  %tmp26 = sext i32 %var183 to i64
+  %tmp27 = sext i32 %var1863 to i64
+  %tmp28 = sext i32 %var192 to i64
+  %tmp29 = sext i32 %var198 to i64
+  %tmp30 = sext i32 %var204 to i64
+  %tmp31 = sext i32 %var207 to i64
+  %tmp32 = sext i32 %var210 to i64
+  %tmp33 = sext i32 %var2195 to i64
+  %tmp34 = sext i32 %var222 to i64
+  %tmp35 = icmp ne i32 %var225, 0
+  %.42831 = zext i1 %tmp35 to i64
+  %tmp36 = sext i32 %var228 to i64
+  %tmp37 = sext i32 %var231 to i64
+  %tmp38 = sext i32 %var234 to i64
+  %tmp39 = sext i32 %var240 to i64
+  %tmp40 = sext i32 %var243 to i64
+  %tmp41 = sext i32 %var249 to i64
+  %tmp42 = sext i32 %var252 to i64
+  %tmp43 = sext i32 %var255 to i64
+  %tmp44 = sext i32 %var2587 to i64
+  %tmp45 = sext i32 %var261 to i64
+  %tmp46 = sext i32 %var273 to i64
+  %tmp47 = sext i32 %var279 to i64
+  %phitmp42823 = sext i32 %var2829 to i64
+  %tmp48 = sext i32 %var285 to i64
+  %tmp49 = sext i32 %var288 to i64
+  %tmp50 = sext i32 %var291 to i64
+  %tmp51 = sext i32 %var297 to i64
+  %phitmp42824 = sext i32 %var30010 to i64
+  %tmp52 = sext i32 %tmp13 to i64
+  %tmp53 = sext i32 %var309 to i64
+  %tmp54 = icmp ne i32 %var348, 0
+  %.42832 = zext i1 %tmp54 to i64
+  %tmp55 = sext i32 %var312 to i64
+  %tmp56 = sext i32 %var315 to i64
+  %tmp57 = sext i32 %var318 to i64
+  %tmp58 = sext i32 %var321 to i64
+  %tmp59 = sext i32 %var324 to i64
+  %tmp60 = sext i32 %var327 to i64
+  %tmp61 = sext i32 %var330 to i64
+  %phitmp42825 = sext i32 %var37211 to i64
+  %tmp62 = icmp ne i32 %var351, 0
+  %.42833 = zext i1 %tmp62 to i64
+  %tmp63 = sext i32 %var333 to i64
+  %phitmp42826 = sext i32 %var37512 to i64
+  %tmp64 = sext i32 %var336 to i64
+  %tmp65 = sext i32 %var339 to i64
+  %tmp66 = sext i32 %var342 to i64
+  %tmp67 = sext i32 %var345 to i64
+  %tmp68 = sext i32 %var384 to i64
+  %phitmp42827 = sext i32 %var3782 to i64
+  %.op = shl nsw i64 %tmp19, 1
+  %tmp69 = add i64 %.op, %var150
+  %tmp70 = add i64 %tmp69, %tmp21
+  %tmp71 = add i64 %tmp70, %var156
+  %tmp72 = add i64 %tmp71, %var159
+  %tmp73 = add i64 %tmp72, %tmp22
+  %tmp74 = add i64 %tmp73, %tmp23
+  %tmp75 = add i64 %tmp74, %var1712
+  %tmp76 = add i64 %tmp75, %tmp25
+  %tmp77 = add i64 %tmp76, %tmp26
+  %tmp78 = add i64 %tmp77, %var189
+  %tmp79 = add i64 %tmp78, %tmp27
+  %tmp80 = add i64 %tmp79, %tmp28
+  %tmp81 = add i64 %tmp80, %var195
+  %tmp82 = add i64 %tmp81, %tmp29
+  %tmp83 = add i64 %tmp82, %var201
+  %tmp84 = add i64 %tmp83, %tmp30
+  %tmp85 = add i64 %tmp84, %tmp31
+  %tmp86 = add i64 %tmp85, %tmp32
+  %tmp87 = add i64 %tmp86, %var2134
+  %tmp88 = add i64 %tmp87, 216
+  %tmp89 = add i64 %tmp88, %tmp34
+  %tmp90 = add i64 %tmp89, %tmp33
+  %tmp91 = add i64 %tmp90, %.42831
+  %tmp92 = add i64 %tmp91, %tmp36
+  %tmp93 = add i64 %tmp92, %tmp37
+  %tmp94 = add i64 %tmp93, %tmp38
+  %tmp95 = add i64 %tmp94, %var2376
+  %tmp96 = add i64 %tmp95, %tmp39
+  %tmp97 = add i64 %tmp96, %tmp40
+  %tmp98 = add i64 %tmp97, %var246
+  %tmp99 = add i64 %tmp98, %tmp41
+  %tmp100 = add i64 %tmp99, %tmp42
+  %tmp101 = add i64 %tmp100, %tmp43
+  %tmp102 = add i64 %tmp101, %tmp45
+  %tmp103 = add i64 %tmp102, %tmp44
+  %tmp104 = add i64 %tmp103, %var270
+  %tmp105 = add i64 %tmp104, %tmp46
+  %tmp106 = add i64 %tmp105, %tmp47
+  %tmp107 = add i64 %tmp106, %tmp48
+  %tmp108 = add i64 %tmp107, %phitmp42823
+  %tmp109 = add i64 %tmp108, %tmp49
+  %tmp110 = add i64 %tmp109, %tmp50
+  %tmp111 = add i64 %tmp110, %tmp51
+  %tmp112 = add i64 %tmp111, %phitmp42824
+  %tmp113 = add i64 %tmp112, %tmp53
+  %tmp114 = add i64 %tmp113, %tmp56
+  %tmp115 = add i64 %tmp114, %tmp58
+  %tmp116 = add i64 %tmp115, %tmp59
+  %tmp117 = add i64 %tmp116, %tmp60
+  %tmp118 = add i64 %tmp117, %tmp61
+  %tmp119 = add i64 %tmp118, %tmp63
+  %tmp120 = add i64 %tmp119, %tmp64
+  %tmp121 = add i64 %tmp120, %tmp65
+  %tmp122 = add i64 %tmp121, %tmp66
+  %tmp123 = add i64 %tmp122, %tmp67
+  %tmp124 = add i64 %tmp123, %.42833
+  %tmp125 = add i64 %tmp124, %var357
+  %tmp126 = add i64 %tmp125, %var360
+  %tmp127 = add i64 %tmp126, %var363
+  %tmp128 = add i64 %tmp127, %phitmp42825
+  %tmp129 = add i64 %tmp128, %phitmp42826
+  %tmp130 = add i64 %tmp129, %tmp68
+  %tmp131 = add i64 %tmp130, %phitmp42827
+  %tmp132 = add i64 %tmp131, %tmp16
+  %tmp133 = add i64 %tmp132, %tmp17
+  %tmp134 = add i64 %tmp133, %tmp18
+  %tmp135 = add i64 %tmp134, %tmp20
+  %tmp136 = add i64 %tmp135, %tmp24
+  %tmp137 = add i64 %tmp136, %var2678
+  %tmp138 = add i64 %tmp137, %tmp55
+  %tmp139 = add i64 %tmp138, %tmp57
+  %tmp140 = add i64 %tmp139, %.42832
+  %tmp141 = add i64 %tmp140, 2143289440
+  %tmp142 = add i64 %tmp141, %var174
+  %tmp143 = add i64 %tmp142, %tmp15
+  %tmp144 = add i64 %tmp143, %var354
+  %tmp145 = add i64 %tmp144, %tmp52
+  %tmp146 = add i64 %tmp145, %phi_125_487
+  store volatile i64 %tmp146, i64* null, align 536870912
+  ret void
+
+backedge:                                          ; preds = %loop
+  br label %loop
+}