From 9dada83d6c533376b919eaf93cea99772f180814 Mon Sep 17 00:00:00 2001
From: Sanjay Patel <spatel@rotateright.com>
Date: Tue, 26 Feb 2019 18:26:56 +0000
Subject: [InstSimplify] remove zero-shift-guard fold for general funnel shift

As discussed on llvm-dev:
http://lists.llvm.org/pipermail/llvm-dev/2019-February/130491.html

We can't remove the compare+select in the general case because
we are treating funnel shift like a standard instruction (as
opposed to a special instruction like select/phi).

That means that if one of the operands of the funnel shift is
poison, the result is poison regardless of whether we know that
the operand is actually unused based on the instruction's
particular semantics.

The motivating case for this transform is the more specific
rotate op (rather than funnel shift), and we are preserving the
fold for that case because there is no chance of introducing
extra poison when there is no anonymous extra operand to the
funnel shift.

llvm-svn: 354905
---
 llvm/lib/Analysis/InstructionSimplify.cpp | 41 ++++++++++++++++++++++---------
 1 file changed, 29 insertions(+), 12 deletions(-)

(limited to 'llvm/lib')

diff --git a/llvm/lib/Analysis/InstructionSimplify.cpp b/llvm/lib/Analysis/InstructionSimplify.cpp
index ed857ddbc78..baf72a0bf37 100644
--- a/llvm/lib/Analysis/InstructionSimplify.cpp
+++ b/llvm/lib/Analysis/InstructionSimplify.cpp
@@ -3908,27 +3908,44 @@ static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
                                            Pred == ICmpInst::ICMP_EQ))
         return V;
 
-    // Test for zero-shift-guard-ops around funnel shifts. These are used to
-    // avoid UB from oversized shifts in raw IR rotate patterns, but the
-    // intrinsics do not have that problem.
+    // Test for a bogus zero-shift-guard-op around funnel-shift or rotate.
     Value *ShAmt;
     auto isFsh = m_CombineOr(m_Intrinsic<Intrinsic::fshl>(m_Value(X), m_Value(),
                                                           m_Value(ShAmt)),
                              m_Intrinsic<Intrinsic::fshr>(m_Value(), m_Value(X),
                                                           m_Value(ShAmt)));
-    // (ShAmt != 0) ? fshl(X, *, ShAmt) : X --> fshl(X, *, ShAmt)
-    // (ShAmt != 0) ? fshr(*, X, ShAmt) : X --> fshr(*, X, ShAmt)
     // (ShAmt == 0) ? fshl(X, *, ShAmt) : X --> X
     // (ShAmt == 0) ? fshr(*, X, ShAmt) : X --> X
-    if (match(TrueVal, isFsh) && FalseVal == X && CmpLHS == ShAmt)
-      return Pred == ICmpInst::ICMP_NE ? TrueVal : X;
-
-    // (ShAmt == 0) ? X : fshl(X, *, ShAmt) --> fshl(X, *, ShAmt)
-    // (ShAmt == 0) ? X : fshr(*, X, ShAmt) --> fshr(*, X, ShAmt)
+    if (match(TrueVal, isFsh) && FalseVal == X && CmpLHS == ShAmt &&
+        Pred == ICmpInst::ICMP_EQ)
+      return X;
     // (ShAmt != 0) ? X : fshl(X, *, ShAmt) --> X
     // (ShAmt != 0) ? X : fshr(*, X, ShAmt) --> X
-    if (match(FalseVal, isFsh) && TrueVal == X && CmpLHS == ShAmt)
-      return Pred == ICmpInst::ICMP_EQ ? FalseVal : X;
+    if (match(FalseVal, isFsh) && TrueVal == X && CmpLHS == ShAmt &&
+        Pred == ICmpInst::ICMP_NE)
+      return X;
+
+    // Test for a zero-shift-guard-op around rotates. These are used to
+    // avoid UB from oversized shifts in raw IR rotate patterns, but the
+    // intrinsics do not have that problem.
+    // We do not allow this transform for the general funnel shift case because
+    // that would not preserve the poison safety of the original code.
+    auto isRotate = m_CombineOr(m_Intrinsic<Intrinsic::fshl>(m_Value(X),
+                                                             m_Deferred(X),
+                                                             m_Value(ShAmt)),
+                                m_Intrinsic<Intrinsic::fshr>(m_Value(X),
+                                                             m_Deferred(X),
+                                                             m_Value(ShAmt)));
+    // (ShAmt != 0) ? fshl(X, X, ShAmt) : X --> fshl(X, X, ShAmt)
+    // (ShAmt != 0) ? fshr(X, X, ShAmt) : X --> fshr(X, X, ShAmt)
+    if (match(TrueVal, isRotate) && FalseVal == X && CmpLHS == ShAmt &&
+        Pred == ICmpInst::ICMP_NE)
+      return TrueVal;
+    // (ShAmt == 0) ? X : fshl(X, X, ShAmt) --> fshl(X, X, ShAmt)
+    // (ShAmt == 0) ? X : fshr(X, X, ShAmt) --> fshr(X, X, ShAmt)
+    if (match(FalseVal, isRotate) && TrueVal == X && CmpLHS == ShAmt &&
+        Pred == ICmpInst::ICMP_EQ)
+      return FalseVal;
   }
 
   // Check for other compares that behave like bit test.
-- 
cgit v1.2.3