patch and regression test for problem with vzeroupper

llvm_patches/3_3_r193261_bug17631_win_vzeroupper.patch

@@ -0,0 +1,69 @@
+From b9b016cda57d8afc26a150de7ee329b54a994c85 Mon Sep 17 00:00:00 2001
+From: Michael Liao <michael.hliao@gmail.com>
+Date: Mon, 21 Oct 2013 17:47:58 -0700
+Subject: [PATCH] Fix PR17631
+
+- Skip instructions added in prolog. For specific targets, prolog may
+  insert helper function calls (e.g. _chkstk will be called when
+  there're more than 4K bytes allocated on stack). However, these
+  helpers don't use/def YMM/XMM registers.
+---
+ lib/Target/X86/X86VZeroUpper.cpp | 11 ++++++++++-
+ test/CodeGen/X86/pr17631.ll      | 22 ++++++++++++++++++++++
+ 2 files changed, 32 insertions(+), 1 deletion(-)
+ create mode 100644 test/CodeGen/X86/pr17631.ll
+
+diff --git a/lib/Target/X86/X86VZeroUpper.cpp b/lib/Target/X86/X86VZeroUpper.cpp
+index 477f75a..0d37a7d 100644
+--- lib/Target/X86/X86VZeroUpper.cpp
++++ lib/Target/X86/X86VZeroUpper.cpp
+@@ -231,8 +231,17 @@ bool VZeroUpperInserter::processBasicBlock(MachineFunction &MF,
+   bool BBHasCall = false;
+ 
+   for (MachineBasicBlock::iterator I = BB.begin(); I != BB.end(); ++I) {
+-    MachineInstr *MI = I;
+     DebugLoc dl = I->getDebugLoc();
++    MachineInstr *MI = I;
++
++    // Don't need to check instructions added in prolog.
++    // In prolog, special function calls may be added for specific targets
++    // (e.g. on Windows, a prolog helper '_chkstk' is called when the local
++    // variables exceed 4K bytes on stack.) These helpers won't use/def YMM/XMM
++    // registers.
++    if (MI->getFlag(MachineInstr::FrameSetup))
++      continue;
++
+     bool isControlFlow = MI->isCall() || MI->isReturn();
+ 
+     // Shortcut: don't need to check regular instructions in dirty state.
+diff --git a/test/CodeGen/X86/pr17631.ll b/test/CodeGen/X86/pr17631.ll
+new file mode 100644
+index 0000000..a572ff2
+--- /dev/null
++++ test/CodeGen/X86/pr17631.ll
+@@ -0,0 +1,22 @@
++; RUN: llc < %s -mcpu=core-avx-i -mtriple=i386-pc-win32 | FileCheck %s
++ 
++%struct_type = type { [64 x <8 x float>], <8 x float> }
++ 
++; Function Attrs: nounwind readnone
++declare i32 @llvm.x86.avx.movmsk.ps.256(<8 x float>)
++ 
++; Function Attrs: nounwind
++define i32 @equal(<8 x i32> %A) {
++allocas:
++  %first_alloc  = alloca [64 x <8 x i32>]
++  %second_alloc = alloca %struct_type
++ 
++  %A1 = bitcast <8 x i32> %A to <8 x float>
++  %A2 = call i32 @llvm.x86.avx.movmsk.ps.256(<8 x float> %A1)
++  ret i32 %A2
++}
++
++; CHECK: equal
++; CHECK-NOT: vzeroupper
++; CHECK: _chkstk
++; CHECK: ret
+-- 
+1.8.1.2
+

tests/chkstk.ispc

@@ -0,0 +1,49 @@
+//test for 17631 bug in LLVM.
+
+export uniform int width() { return programCount; }
+ 
+struct s_temp
+{
+    float temp[64];
+};
+ 
+int CompressBlockBC7(int A, uniform float b)
+{
+    // This declaration caused problem because LLVM inserted
+    // _chkstk after declaration and vzeroupper before it's call.
+    // A will be in ymm at avx, so we lose a half of it.
+    s_temp _state;
+    // These two loops are here to prevent elimination of declaration
+    for (int i=0; i<64; i++) {
+        float ii = i;
+        _state.temp[i] = b + sin(ii);
+    }
+    float r = 0;
+    for (int j=0; j<64; j+=9) {
+        r += _state.temp[j] + j;
+    }
+
+    // Here upper bits of A in ymm can be zeros. This will crash the test.
+    int B;
+    if (A!=0) {
+        B = 20;
+    }
+    else {
+        B = 30;
+    }
+    if(A == 1) {
+        B = r;
+    }
+    return B;
+}
+ 
+export void f_fu(uniform float RET[], uniform float aFOO[], uniform float b) {
+    int A = programIndex;
+    RET[programIndex] = CompressBlockBC7(A, b);
+}
+ 
+export void result(uniform float RET[]) {
+    RET[programIndex] = 20;
+    RET[0] = 30;
+    RET[1] = 292;
+}