Improve 'tail' call marking in TRE. A bootstrap of clang goes from 375k calls marked tail in the IR to 470k, however this improvement does not carry into an improvement of the call/jmp ratio on x86. The most common pattern is a tail call + br to a block with nothing but a 'ret'.

The number of tail call to loop conversions remains the same (1618 by my count).

The new algorithm does a local scan over the use-def chains to identify local "alloca-derived" values, as well as points where the alloca could escape. Then, a visit over the CFG marks blocks as being before or after the allocas have escaped, and annotates the calls accordingly.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208017 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 23 insertions, 0 deletions

diff --git a/test/Transforms/TailCallElim/basic.ll b/test/Transforms/TailCallElim/basic.ll
index 5582ee33ed..341736d48e 100644
--- a/test/Transforms/TailCallElim/basic.ll
+++ b/test/Transforms/TailCallElim/basic.ll
@@ -151,3 +151,26 @@ define void @test9(i32* byval %a) {
   call void @use(i32* %a)
   ret void
 }
+
+%struct.X = type { i8* }
+
+declare void @ctor(%struct.X*)
+define void @test10(%struct.X* noalias sret %agg.result, i1 zeroext %b) {
+; CHECK-LABEL @test10
+entry:
+  %x = alloca %struct.X, align 8
+  br i1 %b, label %if.then, label %if.end
+
+if.then:                                          ; preds = %entry
+  call void @ctor(%struct.X* %agg.result)
+; CHECK: tail call void @ctor
+  br label %return
+
+if.end:
+  call void @ctor(%struct.X* %x)
+; CHECK: call void @ctor
+  br label %return
+
+return:
+  ret void
+}