[llvm-commits] [PATCH] [llvm] Transform (x&C)>V into (x&C)!=0 where possible.

[Schoedel, Kevin P]

When the least bit of C is greater than V, (x&C) must be greater than V
if it is not zero, so the comparison can be simplified.

Although this was suggested in Target/X86/README.txt, it benefits any
architecture with a directly testable form of AND.
---
 lib/Target/X86/README.txt                          |   37 --------------------
 lib/Transforms/InstCombine/InstCombineCompares.cpp |    9 +++++
 test/Transforms/InstCombine/icmp.ll                |   16 +++++++++
 3 files changed, 25 insertions(+), 37 deletions(-)

diff --git a/lib/Target/X86/README.txt b/lib/Target/X86/README.txt
index 6a8a4fd..b4285a0 100644
--- a/lib/Target/X86/README.txt
+++ b/lib/Target/X86/README.txt
@@ -1568,43 +1568,6 @@ The second one is done for: Atom, Pentium Pro, all AMDs, Pentium 4, Nocona,
   Core 2, and "Generic"
 
 //===---------------------------------------------------------------------===//
-
-Testcase:
-int a(int x) { return (x & 127) > 31; }
-
-Current output:
-	movl	4(%esp), %eax
-	andl	$127, %eax
-	cmpl	$31, %eax
-	seta	%al
-	movzbl	%al, %eax
-	ret
-
-Ideal output:
-	xorl	%eax, %eax
-	testl	$96, 4(%esp)
-	setne	%al
-	ret
-
-This should definitely be done in instcombine, canonicalizing the range
-condition into a != condition.  We get this IR:
-
-define i32 @a(i32 %x) nounwind readnone {
-entry:
-	%0 = and i32 %x, 127		; <i32> [#uses=1]
-	%1 = icmp ugt i32 %0, 31		; <i1> [#uses=1]
-	%2 = zext i1 %1 to i32		; <i32> [#uses=1]
-	ret i32 %2
-}
-
-Instcombine prefers to strength reduce relational comparisons to equality
-comparisons when possible, this should be another case of that.  This could
-be handled pretty easily in InstCombiner::visitICmpInstWithInstAndIntCst, but it
-looks like InstCombiner::visitICmpInstWithInstAndIntCst should really already
-be redesigned to use ComputeMaskedBits and friends.
-
-
-//===---------------------------------------------------------------------===//
 Testcase:
 int x(int a) { return (a&0xf0)>>4; }
 
diff --git a/lib/Transforms/InstCombine/InstCombineCompares.cpp b/lib/Transforms/InstCombine/InstCombineCompares.cpp
index 1b96c3c..ce1066a 100644
--- a/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -1226,6 +1226,15 @@ Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
         ICI.setOperand(0, NewAnd);
         return &ICI;
       }
+
+      // If any bit set in (X & AndCST) will produce a result greater than RHSV,
+      // replace ((X & AndCST) > RHSV) with ((X & AndCST) != 0).
+      unsigned NTZ = AndCST->getValue().countTrailingZeros();
+      if ((ICI.getPredicate() == ICmpInst::ICMP_UGT) &&
+          (NTZ < AndCST->getBitWidth()) &&
+          APInt::getOneBitSet(AndCST->getBitWidth(), NTZ).ugt(RHSV))
+        return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
+                            Constant::getNullValue(RHS->getType()));
     }
 
     // Try to optimize things like "A[i]&42 == 0" to index computations.
diff --git a/test/Transforms/InstCombine/icmp.ll b/test/Transforms/InstCombine/icmp.ll
index 8e064a4..c265c08 100644
--- a/test/Transforms/InstCombine/icmp.ll
+++ b/test/Transforms/InstCombine/icmp.ll
@@ -677,3 +677,19 @@ define i1 @test66(i64 %A, i64 %B) {
 ; CHECK-NEXT: ret i1 true
   ret i1 %cmp
 }
+
+; CHECK: @test67
+; CHECK: %cmp = icmp ne i32 %and, 0
+define i1 @test67(i32 %x) nounwind uwtable {
+  %and = and i32 %x, 127
+  %cmp = icmp sgt i32 %and, 31
+  ret i1 %cmp
+}
+
+; CHECK: @test68
+; CHECK: %cmp = icmp ugt i32 %and, 30
+define i1 @test68(i32 %x) nounwind uwtable {
+  %and = and i32 %x, 127
+  %cmp = icmp sgt i32 %and, 30
+  ret i1 %cmp
+}
-- 
1.7.10.4