summaryrefslogtreecommitdiff
path: root/test/Transforms/LoopStrengthReduce
diff options
context:
space:
mode:
authorAndrew Trick <atrick@apple.com>2012-03-22 17:10:07 +0000
committerAndrew Trick <atrick@apple.com>2012-03-22 17:10:07 +0000
commit0c0a3effdeb8d90b314a48f6306ce7dbd05b6255 (patch)
tree1576d2587dcd05da674cda3d49c843b1049f54f4 /test/Transforms/LoopStrengthReduce
parentf6399a65f011a64e8d57c3ebe51350c4aecc7cfd (diff)
downloadllvm-0c0a3effdeb8d90b314a48f6306ce7dbd05b6255.tar.gz
llvm-0c0a3effdeb8d90b314a48f6306ce7dbd05b6255.tar.bz2
llvm-0c0a3effdeb8d90b314a48f6306ce7dbd05b6255.tar.xz
Convert -indvars tests that rely on SCEV expansion to -loop-reduce tests.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153259 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'test/Transforms/LoopStrengthReduce')
-rw-r--r--test/Transforms/LoopStrengthReduce/addrec-gep.ll82
-rw-r--r--test/Transforms/LoopStrengthReduce/preserve-gep-loop-variant.ll42
2 files changed, 124 insertions, 0 deletions
diff --git a/test/Transforms/LoopStrengthReduce/addrec-gep.ll b/test/Transforms/LoopStrengthReduce/addrec-gep.ll
new file mode 100644
index 0000000000..3e4e369657
--- /dev/null
+++ b/test/Transforms/LoopStrengthReduce/addrec-gep.ll
@@ -0,0 +1,82 @@
+; RUN: opt < %s -loop-reduce -S | FileCheck %s
+; CHECK: bb1:
+; CHECK: load double* [[IV:%[^,]+]]
+; CHECK: store double {{.*}}, double* [[IV]]
+; CHECK: getelementptr double*
+; CHECK-NOT: cast
+; CHECK: br {{.*}} label %bb1
+
+; This test tests several things. The load and store should use the
+; same address instead of having it computed twice, and SCEVExpander should
+; be able to reconstruct the full getelementptr, despite it having a few
+; obstacles set in its way.
+; We only check that the inner loop (bb1-bb2) is "reduced" because LSR
+; currently only operates on inner loops.
+
+target datalayout = "e-p:64:64:64-n32:64"
+
+define void @foo(i64 %n, i64 %m, i64 %o, i64 %q, double* nocapture %p) nounwind {
+entry:
+ %tmp = icmp sgt i64 %n, 0 ; <i1> [#uses=1]
+ br i1 %tmp, label %bb.nph3, label %return
+
+bb.nph: ; preds = %bb2.preheader
+ %tmp1 = mul i64 %tmp16, %i.02 ; <i64> [#uses=1]
+ %tmp2 = mul i64 %tmp19, %i.02 ; <i64> [#uses=1]
+ br label %bb1
+
+bb1: ; preds = %bb2, %bb.nph
+ %j.01 = phi i64 [ %tmp9, %bb2 ], [ 0, %bb.nph ] ; <i64> [#uses=3]
+ %tmp3 = add i64 %j.01, %tmp1 ; <i64> [#uses=1]
+ %tmp4 = add i64 %j.01, %tmp2 ; <i64> [#uses=1]
+ %z0 = add i64 %tmp3, 5203
+ %tmp5 = getelementptr double* %p, i64 %z0 ; <double*> [#uses=1]
+ %tmp6 = load double* %tmp5, align 8 ; <double> [#uses=1]
+ %tmp7 = fdiv double %tmp6, 2.100000e+00 ; <double> [#uses=1]
+ %z1 = add i64 %tmp4, 5203
+ %tmp8 = getelementptr double* %p, i64 %z1 ; <double*> [#uses=1]
+ store double %tmp7, double* %tmp8, align 8
+ %tmp9 = add i64 %j.01, 1 ; <i64> [#uses=2]
+ br label %bb2
+
+bb2: ; preds = %bb1
+ %tmp10 = icmp slt i64 %tmp9, %m ; <i1> [#uses=1]
+ br i1 %tmp10, label %bb1, label %bb2.bb3_crit_edge
+
+bb2.bb3_crit_edge: ; preds = %bb2
+ br label %bb3
+
+bb3: ; preds = %bb2.preheader, %bb2.bb3_crit_edge
+ %tmp11 = add i64 %i.02, 1 ; <i64> [#uses=2]
+ br label %bb4
+
+bb4: ; preds = %bb3
+ %tmp12 = icmp slt i64 %tmp11, %n ; <i1> [#uses=1]
+ br i1 %tmp12, label %bb2.preheader, label %bb4.return_crit_edge
+
+bb4.return_crit_edge: ; preds = %bb4
+ br label %bb4.return_crit_edge.split
+
+bb4.return_crit_edge.split: ; preds = %bb.nph3, %bb4.return_crit_edge
+ br label %return
+
+bb.nph3: ; preds = %entry
+ %tmp13 = icmp sgt i64 %m, 0 ; <i1> [#uses=1]
+ %tmp14 = mul i64 %n, 37 ; <i64> [#uses=1]
+ %tmp15 = mul i64 %tmp14, %o ; <i64> [#uses=1]
+ %tmp16 = mul i64 %tmp15, %q ; <i64> [#uses=1]
+ %tmp17 = mul i64 %n, 37 ; <i64> [#uses=1]
+ %tmp18 = mul i64 %tmp17, %o ; <i64> [#uses=1]
+ %tmp19 = mul i64 %tmp18, %q ; <i64> [#uses=1]
+ br i1 %tmp13, label %bb.nph3.split, label %bb4.return_crit_edge.split
+
+bb.nph3.split: ; preds = %bb.nph3
+ br label %bb2.preheader
+
+bb2.preheader: ; preds = %bb.nph3.split, %bb4
+ %i.02 = phi i64 [ %tmp11, %bb4 ], [ 0, %bb.nph3.split ] ; <i64> [#uses=3]
+ br i1 true, label %bb.nph, label %bb3
+
+return: ; preds = %bb4.return_crit_edge.split, %entry
+ ret void
+}
diff --git a/test/Transforms/LoopStrengthReduce/preserve-gep-loop-variant.ll b/test/Transforms/LoopStrengthReduce/preserve-gep-loop-variant.ll
new file mode 100644
index 0000000000..f90d030800
--- /dev/null
+++ b/test/Transforms/LoopStrengthReduce/preserve-gep-loop-variant.ll
@@ -0,0 +1,42 @@
+; RUN: opt < %s -loop-reduce -S | FileCheck %s
+; CHECK-NOT: {{inttoptr|ptrtoint}}
+; CHECK: scevgep
+; CHECK-NOT: {{inttoptr|ptrtoint}}
+target datalayout = "E-p:64:64:64-a0:0:8-f32:32:32-f64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-v64:64:64-v128:128:128-n32:64"
+
+; Indvars shouldn't need inttoptr/ptrtoint to expand an address here.
+
+define void @foo(i8* %p) nounwind {
+entry:
+ br i1 true, label %bb.nph, label %for.end
+
+for.cond:
+ %phitmp = icmp slt i64 %inc, 20
+ br i1 %phitmp, label %for.body, label %for.cond.for.end_crit_edge
+
+for.cond.for.end_crit_edge:
+ br label %for.end
+
+bb.nph:
+ br label %for.body
+
+for.body:
+ %storemerge1 = phi i64 [ %inc, %for.cond ], [ 0, %bb.nph ]
+ %call = tail call i64 @bar() nounwind
+ %call2 = tail call i64 @car() nounwind
+ %conv = trunc i64 %call2 to i8
+ %conv3 = sext i8 %conv to i64
+ %add = add nsw i64 %call, %storemerge1
+ %add4 = add nsw i64 %add, %conv3
+ %arrayidx = getelementptr inbounds i8* %p, i64 %add4
+ store i8 0, i8* %arrayidx
+ %inc = add nsw i64 %storemerge1, 1
+ br label %for.cond
+
+for.end:
+ ret void
+}
+
+declare i64 @bar()
+
+declare i64 @car()