[x86, AVX] allow explicit calls to VZERO* to modify state in VZeroUpperInserter pass (PR27823)
As noted in the review, there are still problems, so this doesn't the bug completely.
Differential Revision: http://reviews.llvm.org/D20529
llvm-svn: 270718
diff --git a/llvm/lib/Target/X86/X86VZeroUpper.cpp b/llvm/lib/Target/X86/X86VZeroUpper.cpp
index 4465c0f..944c76a 100644
--- a/llvm/lib/Target/X86/X86VZeroUpper.cpp
+++ b/llvm/lib/Target/X86/X86VZeroUpper.cpp
@@ -188,16 +188,17 @@
bool IsReturnFromX86INTR = IsX86INTR && MI->isReturn();
bool IsControlFlow = MI->isCall() || MI->isReturn();
+ // An existing VZERO* instruction resets the state.
+ if (MI->getOpcode() == X86::VZEROALL ||
+ MI->getOpcode() == X86::VZEROUPPER) {
+ CurState = EXITS_CLEAN;
+ continue;
+ }
+
// Shortcut: don't need to check regular instructions in dirty state.
if ((!IsControlFlow || IsReturnFromX86INTR) && CurState == EXITS_DIRTY)
continue;
- // Ignore existing VZERO* instructions.
- // FIXME: The existence of these instructions should be used to modify the
- // current state and/or used when deciding whether we need to create a VZU.
- if (MI->getOpcode() == X86::VZEROALL || MI->getOpcode() == X86::VZEROUPPER)
- continue;
-
if (hasYmmReg(MI)) {
// We found a ymm-using instruction; this could be an AVX instruction,
// or it could be control flow.