[PATCH] D90145: [TargetLowering] Add i1 condition for bit comparison fold

[Erik Hogeman via Phabricator via llvm-commits]

ErikHogeman created this revision.
ErikHogeman added a reviewer: spatel.
Herald added subscribers: llvm-commits, hiraditya, jholewinski.
Herald added a project: LLVM.
ErikHogeman requested review of this revision.

For i1 types, boolean false is represented identically regardless of the boolean content, so we can allow optimizations that otherwise would not be correct for booleans with false represented as a negative one.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D90145

Files:
  llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
  llvm/test/CodeGen/NVPTX/pow2_mask_cmp.ll


Index: llvm/test/CodeGen/NVPTX/pow2_mask_cmp.ll
===================================================================
--- llvm/test/CodeGen/NVPTX/pow2_mask_cmp.ll
+++ llvm/test/CodeGen/NVPTX/pow2_mask_cmp.ll
@@ -3,14 +3,14 @@
 ; Tests the following pattern:
 ; (X & 8) != 0 --> (X & 8) >> 3
 
-; This produces incorrect code when boolean false is represented
-; as a negative one, and this test checks that the transform is
-; not triggered.
+; This produces incorrect code in general when boolean false is
+; represented as a negative one. There is however a special
+; case when the type has a bitsize of 1, for which the false
+; value will be identical regardless of the boolean representation.
+; Check that the optimization triggers in this case.
 
 ; CHECK-LABEL: @pow2_mask_cmp
-; CHECK: and.b32 [[AND:%r[0-9]+]], %r{{[0-9]+}}, 8
-; CHECK: setp.ne.s32 [[SETP:%p[0-9+]]], [[AND]], 0
-; CHECK: selp.u32 %r{{[0-9]+}}, 1, 0, [[SETP]]
+; CHECK: bfe.u32 {{%r[0-9]+}}, {{%r[0-9]+}}, 3, 1
 define i32 @pow2_mask_cmp(i32 %x) {
   %a = and i32 %x, 8
   %cmp = icmp ne i32 %a, 0
Index: llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
===================================================================
--- llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -3980,8 +3980,12 @@
     const APInt &C1 = N1C->getAPIntValue();
     EVT ShValTy = N0.getValueType();
 
-    // Fold bit comparisons when we can.
-    if (getBooleanContents(N0.getValueType()) == ZeroOrOneBooleanContent &&
+    // Fold bit comparisons when we can. This will result in an
+    // incorrect value when boolean false is negative one, unless
+    // the bitsize is 1 in which case the false value is the same
+    // in practice regardless of the representation.
+    if ((VT.getSizeInBits() == 1 ||
+         getBooleanContents(N0.getValueType()) == ZeroOrOneBooleanContent) &&
         (Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
         (VT == ShValTy || (isTypeLegal(VT) && VT.bitsLE(ShValTy))) &&
         N0.getOpcode() == ISD::AND) {


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D90145.300621.patch
Type: text/x-patch
Size: 2065 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20201026/5d8c1487/attachment.bin>