[PATCH] D36944: [InstCombine] Fix a weakness in canEvaluateZExtd around 'and' instructions

[Craig Topper via Phabricator via llvm-commits]

craig.topper created this revision.

If the bitsToClear from the LHS of an 'and' comes back non-zero, but all of those bits are known zero on the RHS, we can reset bitsToClear.

Without this, the 'or' in the modified test case blocks the transform because it has non-zero bits in its RHS in those bits.


https://reviews.llvm.org/D36944

Files:
  lib/Transforms/InstCombine/InstCombineCasts.cpp
  test/Transforms/InstCombine/cast.ll


Index: test/Transforms/InstCombine/cast.ll
===================================================================
--- test/Transforms/InstCombine/cast.ll
+++ test/Transforms/InstCombine/cast.ll
@@ -1591,12 +1591,10 @@
 ; TODO: This is currently blocked because we don't realize the 'and' has cleared the extra bits that would be shifted in widening the lshr.
 define i32 @test95(i32 %x) {
 ; CHECK-LABEL: @test95(
-; CHECK-NEXT:    [[TMP1:%.*]] = trunc i32 [[X:%.*]] to i8
-; CHECK-NEXT:    [[TMP2:%.*]] = lshr i8 [[TMP1]], 6
-; CHECK-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 2
-; CHECK-NEXT:    [[TMP4:%.*]] = or i8 [[TMP3]], 40
-; CHECK-NEXT:    [[TMP5:%.*]] = zext i8 [[TMP4]] to i32
-; CHECK-NEXT:    ret i32 [[TMP5]]
+; CHECK-NEXT:    [[TMP1:%.*]] = lshr i32 [[X:%.*]], 6
+; CHECK-NEXT:    [[TMP2:%.*]] = and i32 [[TMP1]], 2
+; CHECK-NEXT:    [[TMP3:%.*]] = or i32 [[TMP2]], 40
+; CHECK-NEXT:    ret i32 [[TMP3]]
 ;
   %1 = trunc i32 %x to i8
   %2 = lshr i8 %1, 6
Index: lib/Transforms/InstCombine/InstCombineCasts.cpp
===================================================================
--- lib/Transforms/InstCombine/InstCombineCasts.cpp
+++ lib/Transforms/InstCombine/InstCombineCasts.cpp
@@ -950,8 +950,13 @@
       unsigned VSize = V->getType()->getScalarSizeInBits();
       if (IC.MaskedValueIsZero(I->getOperand(1),
                                APInt::getHighBitsSet(VSize, BitsToClear),
-                               0, CxtI))
+                               0, CxtI)) {
+        // If this is an And instruction and all of the BitsToClear are
+        // known to be zero we can reset BitsToClear.
+        if (Opc == Instruction::And)
+          BitsToClear = 0;
         return true;
+      }
     }
 
     // Otherwise, we don't know how to analyze this BitsToClear case yet.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D36944.111911.patch
Type: text/x-patch
Size: 1796 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170820/58739c56/attachment.bin>