[llvm] 2ba1428 - Revert "[SelectionDAG] Handle NSW for ADD/SUB in computeKnownBits()"

[Nikita Popov via llvm-commits]

Author: Nikita Popov
Date: 2023-05-25T11:13:51+02:00
New Revision: 2ba14283cdce08e03c62e8a1d895f1743300236b

URL: https://github.com/llvm/llvm-project/commit/2ba14283cdce08e03c62e8a1d895f1743300236b
DIFF: https://github.com/llvm/llvm-project/commit/2ba14283cdce08e03c62e8a1d895f1743300236b.diff

LOG: Revert "[SelectionDAG] Handle NSW for ADD/SUB in computeKnownBits()"

This reverts commit b66551370fdfc6f357ae0d77237119d2b1077b62.

This has exposed a pre-existing miscompile, reported in
https://reviews.llvm.org/D150769#4370467.

Added: 
    

Modified: 
    llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
    llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
    llvm/test/CodeGen/Thumb2/mve-blockplacement.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 11f873400fc0d..ffa9998949b40 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -3631,15 +3631,6 @@ KnownBits SelectionDAG::computeKnownBits(SDValue Op, const APInt &DemandedElts,
     // All bits are zero except the low bit.
     Known.Zero.setBitsFrom(1);
     break;
-  case ISD::ADD:
-  case ISD::SUB: {
-    SDNodeFlags Flags = Op.getNode()->getFlags();
-    Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
-    Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
-    Known = KnownBits::computeForAddSub(Op.getOpcode() == ISD::ADD,
-                                        Flags.hasNoSignedWrap(), Known, Known2);
-    break;
-  }
   case ISD::USUBO:
   case ISD::SSUBO:
   case ISD::USUBO_CARRY:
@@ -3653,6 +3644,7 @@ KnownBits SelectionDAG::computeKnownBits(SDValue Op, const APInt &DemandedElts,
       break;
     }
     [[fallthrough]];
+  case ISD::SUB:
   case ISD::SUBC: {
     assert(Op.getResNo() == 0 &&
            "We only compute knownbits for the 
diff erence here.");
@@ -3680,6 +3672,7 @@ KnownBits SelectionDAG::computeKnownBits(SDValue Op, const APInt &DemandedElts,
       break;
     }
     [[fallthrough]];
+  case ISD::ADD:
   case ISD::ADDC:
   case ISD::ADDE: {
     assert(Op.getResNo() == 0 && "We only compute knownbits for the sum here.");

diff  --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 58f6e2a36b212..35197dc199f60 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -2694,9 +2694,11 @@ bool TargetLowering::SimplifyDemandedBits(
     if (Op.getOpcode() == ISD::MUL) {
       Known = KnownBits::mul(KnownOp0, KnownOp1);
     } else { // Op.getOpcode() is either ISD::ADD or ISD::SUB.
-      Known = KnownBits::computeForAddSub(Op.getOpcode() == ISD::ADD,
-                                          Flags.hasNoSignedWrap(), KnownOp0,
-                                          KnownOp1);
+      // TODO: Update `computeForAddCarry` to handle the NSW flag as well so
+      //       that `Flags.hasNoSignedWrap()` can be passed through here
+      //       instead of false.
+      Known = KnownBits::computeForAddSub(Op.getOpcode() == ISD::ADD, false,
+                                          KnownOp0, KnownOp1);
     }
     break;
   }

diff  --git a/llvm/test/CodeGen/Thumb2/mve-blockplacement.ll b/llvm/test/CodeGen/Thumb2/mve-blockplacement.ll
index 41fd74d99356a..2d7126db12fab 100644
--- a/llvm/test/CodeGen/Thumb2/mve-blockplacement.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-blockplacement.ll
@@ -366,16 +366,18 @@ define i32 @d(i64 %e, i32 %f, i64 %g, i32 %h) {
 ; CHECK-NEXT:    str r0, [sp, #12] @ 4-byte Spill
 ; CHECK-NEXT:    movs r1, #4
 ; CHECK-NEXT:    strd r2, r12, [sp, #4] @ 8-byte Folded Spill
-; CHECK-NEXT:    add.w r1, r1, r4, lsr #1
 ; CHECK-NEXT:    add.w r3, r3, r4, lsr #1
-; CHECK-NEXT:    bic r7, r1, #3
+; CHECK-NEXT:    add.w r1, r1, r4, lsr #1
+; CHECK-NEXT:    movw r4, #65532
+; CHECK-NEXT:    vdup.32 q6, r3
+; CHECK-NEXT:    movt r4, #32767
+; CHECK-NEXT:    and.w r7, r1, r4
 ; CHECK-NEXT:    adr r1, .LCPI1_0
+; CHECK-NEXT:    vdup.32 q7, r3
 ; CHECK-NEXT:    vldrw.u32 q0, [r1]
 ; CHECK-NEXT:    adr r1, .LCPI1_1
 ; CHECK-NEXT:    vldrw.u32 q5, [r1]
-; CHECK-NEXT:    vdup.32 q6, r3
 ; CHECK-NEXT:    vadd.i32 q4, q0, lr
-; CHECK-NEXT:    vdup.32 q7, r3
 ; CHECK-NEXT:    b .LBB1_4
 ; CHECK-NEXT:  .LBB1_2: @ %for.body6.preheader
 ; CHECK-NEXT:    @ in Loop: Header=BB1_4 Depth=1