[PATCH] D24955: [ValueTracking] Teach computeKnownBits and ComputeNumSignBits to look through ExtractElement.

[Bjorn Pettersson via llvm-commits]

bjope created this revision.
bjope added reviewers: majnemer, spatel.
bjope added a subscriber: llvm-commits.

The computeKnownBits and ComputeNumSignBits functions in ValueTracking can now do a simple look-through of ExtractElement.


https://reviews.llvm.org/D24955

Files:
  lib/Analysis/ValueTracking.cpp
  test/Analysis/ValueTracking/signbits-extract-elt.ll

Index: test/Analysis/ValueTracking/signbits-extract-elt.ll
===================================================================
--- /dev/null
+++ test/Analysis/ValueTracking/signbits-extract-elt.ll
@@ -0,0 +1,38 @@
+; RUN: opt < %s -instcombine -S | FileCheck %s
+
+
+; If computeKnownBits can do a simple look-thru for extractelement
+; then InstCombine will know that we are adding two positive number, and that
+; the addition is nsw.
+; (test case may seem overly complicated, but without two extractelement + add
+; this would be scalarized without testing computeKnownBits which is the
+; purpopse with the test).
+define i1 @test1(<4 x i16>* %in) {
+; CHECK-LABEL: @test1(
+; CHECK-NEXT:    ret i1 false
+  %vec2 = load <4 x i16>, <4 x i16>* %in, align 1
+  %vec3 = lshr <4 x i16> %vec2, <i16 2, i16 3, i16 2, i16 3>
+  %vec4 = sext <4 x i16> %vec3 to <4 x i32>
+  %elt0 = extractelement <4 x i32> %vec4, i32 0
+  %elt1 = extractelement <4 x i32> %vec4, i32 1
+  %sum = add i32 %elt0, %elt1
+  %bool = icmp slt i32 %sum, 0
+  ret i1 %bool
+}
+
+; This is to verify that computeKnownSignBits is doing a simple look-thru for
+; extractelement. It is detected as the shift of %elt0 becoming "nsw".
+define i32 @test2(<4 x i16>* %in) {
+; CHECK-LABEL: @test2(
+; CHECK:    %tmp1 = shl nsw i32 %elt0, 18
+; CHECK:    %tmp2 = shl i32 %elt1, 19
+  %vec2 = load <4 x i16>, <4 x i16>* %in, align 1
+  %vec3 = ashr <4 x i16> %vec2, <i16 2, i16 2, i16 2, i16 2>
+  %vec4 = sext <4 x i16> %vec3 to <4 x i32>
+  %elt0 = extractelement <4 x i32> %vec4, i32 0
+  %elt1 = extractelement <4 x i32> %vec4, i32 1
+  %tmp1 = shl i32 %elt0, 18
+  %tmp2 = shl i32 %elt1, 19
+  %r1 = add i32 %tmp1, %tmp2
+  ret i32 %r1
+}
Index: lib/Analysis/ValueTracking.cpp
===================================================================
--- lib/Analysis/ValueTracking.cpp
+++ lib/Analysis/ValueTracking.cpp
@@ -1389,6 +1389,14 @@
       }
     }
     break;
+  case Instruction::ExtractElement:
+    // Look through extract element.  At the moment we keep this simple and skip
+    // tracking the specific element. But at least we might find information
+    // valid for all elements of the vector (for example if vector is sign
+    // extended, shifted, etc).  It could be questioned if Depth should be
+    // increased or not here. But since it isn't increased when looking thru
+    // ExtractValue I decided to keep the same logic here.
+    return computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth, Q);
   case Instruction::ExtractValue:
     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
       const ExtractValueInst *EVI = cast<ExtractValueInst>(I);
@@ -2220,6 +2228,13 @@
     // FIXME: it's tricky to do anything useful for this, but it is an important
     // case for targets like X86.
     break;
+
+  case Instruction::ExtractElement:
+    // Look through extract element.  At the moment we keep this simple and skip
+    // tracking the specific element. But at least we might find information
+    // valid for all elements of the vector (for example if vector is sign
+    // extended, shifted, etc).
+    return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
   }
 
   // Finally, if we can prove that the top bits of the result are 0's or 1's,


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D24955.72494.patch
Type: text/x-patch
Size: 3282 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160927/9d517bd8/attachment.bin>