Transform (x&C)>V into (x&C)!=0 where possible

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.

Patch by Kevin Schoedel

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@170576 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/README.txt b/lib/Target/X86/README.txt
index 6a8a4fdf2520d81a8b1f5e9828f69cc6b2bcd024..b4285a0718799c379eb2a281f159d2c1a42784ed 100644 (file)
--- a/lib/Target/X86/README.txt
+++ b/lib/Target/X86/README.txt
@@ -1567,43 +1567,6 @@ The first one is done for all AMDs, Core2, and "Generic"
 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 1b96c3cca4ee30142a8106e8f588de9676dea23a..13516d15c4d21e8e9a68830102b7cdc3c8c40e29 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -1226,6 +1226,16 @@ Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
         ICI.setOperand(0, NewAnd);
         return &ICI;
       }
+
+      // Replace ((X & AndCST) > RHSV) with ((X & AndCST) != 0), if any
+      // bit set in (X & AndCST) will produce a result greater than RHSV.
+      if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
+        unsigned NTZ = AndCST->getValue().countTrailingZeros();
+        if ((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 8e064a4f2fc940e2fb49c406a952b556fb2ec187..138b10eaa0276d25368287d38e8c152141b25acc 100644 (file)
--- a/test/Transforms/InstCombine/icmp.ll
+++ b/test/Transforms/InstCombine/icmp.ll
@@ -677,3 +677,20 @@ define i1 @test66(i64 %A, i64 %B) {
 ; CHECK-NEXT: ret i1 true
   ret i1 %cmp
 }
+
+; CHECK: @test67
+; CHECK: %and = and i32 %x, 96
+; 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
+}