Introduce a pass to insert vzeroupper instructions to avoid AVX to

SSE transition penalty. The pass is enabled through the "x86-use-vzeroupper"
llc command line option. This is only the first step (very naive and
conservative one) to sketch out the idea, but proper DFA is coming next
to allow smarter decisions. Comments and ideas now and in further commits
will be very appreciated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@138317 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 26 insertions, 0 deletions

diff --git a/test/CodeGen/X86/avx-vzeroupper.ll b/test/CodeGen/X86/avx-vzeroupper.ll
new file mode 100644
index 0000000000..eaf236c6c7
--- /dev/null
+++ b/test/CodeGen/X86/avx-vzeroupper.ll
@@ -0,0 +1,26 @@
+; RUN: llc < %s -x86-use-vzeroupper -mtriple=x86_64-apple-darwin -mcpu=corei7-avx -mattr=+avx | FileCheck %s
+
+define <4 x float> @do_sse_local(<4 x float> %a) nounwind uwtable readnone ssp {
+entry:
+  %add.i = fadd <4 x float> %a, %a
+  ret <4 x float> %add.i
+}
+
+; CHECK: _test00
+define <4 x float> @test00(<4 x float> %a, <4 x float> %b) nounwind uwtable ssp {
+entry:
+  %add.i = fadd <4 x float> %a, %b
+  ; CHECK: vzeroupper
+  ; CHECK-NEXT: callq _do_sse
+  %call3 = tail call <4 x float> @do_sse(<4 x float> %add.i) nounwind
+  %sub.i = fsub <4 x float> %call3, %add.i
+  ; CHECK-NOT: vzeroupper
+  ; CHECK: callq _do_sse_local
+  %call8 = tail call <4 x float> @do_sse_local(<4 x float> %sub.i)
+  ; CHECK: vzeroupper
+  ; CHECK-NEXT: jmp _do_sse
+  %call10 = tail call <4 x float> @do_sse(<4 x float> %call8) nounwind
+  ret <4 x float> %call10
+}
+
+declare <4 x float> @do_sse(<4 x float>)