[InstCombine] Remove requirement on `trunc` in `slt/sgt` case in `foldSelectICmpAndOr`

[goldstein.w.n]

Summary

AFAICT, the trunc is not needed for correctness/performance and just
blocks what should be handlable cases.

[nikic]

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

676

Unless I'm missing something, we're missing an assignment to the V variable in this branch now? Also, do we still need the NeedAnd?

We should really just be using decomposeBitTestICmp() and a pow2 check I think, but this code is doing some very weird things I don't fully get.

[goldstein.w.n]

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

676

Unless I'm missing something, we're missing an assignment to the V variable in this branch now?

Good catch not sure how that didn't get caught in the tests.

Also, do we still need the NeedAnd?

Yes. x >= 0 == x & SignBit == 0 but we don't have the actual x & SignBit expression.

We should really just be using decomposeBitTestICmp() and a pow2 check I think, but this code is doing some very weird things I don't fully get.

I think that works. Will make it a follow up patch to this.

[goldstein.w.n]

Fix uninit V

[goldstein.w.n]

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

676

Unless I'm missing something, we're missing an assignment to the V variable in this branch now? Also, do we still need the NeedAnd?

Done.

We should really just be using decomposeBitTestICmp() and a pow2 check I think, but this code is doing some very weird things I don't fully get.

See: D148744

[nikic]

llvm/test/Transforms/InstCombine/select-with-bitwise-ops.ll

772

Can you add a multi-use variant of this test? I think the reason why this code is checking for the trunc is that it replaces a one-use trunc with an and, which preserves the number of uses. If we drop the check, we'll have to take NeedsAnd into account when counting uses.

[goldstein.w.n]

Add NeedAnds to use profitability check

[goldstein.w.n]

llvm/test/Transforms/InstCombine/select-with-bitwise-ops.ll

772

Good point. Added NeedAnd to muiltiuse check.

[nikic]

LGTM

[goldstein.w.n]

Rebase

[goldstein.w.n]

Rebase

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

if (IC->isEquality()) {
     const APInt *C1;
     if (!match(CmpLHS, m_And(m_Value(), m_Power2(C1))))
       return nullptr;
 
     V = CmpLHS;
     C1Log = C1->logBase2();
     IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_EQ;
   } else {
-    // We also need to recognize (icmp slt (trunc (X)), 0) and
-    // (icmp sgt (trunc (X)), -1).
+    // We also need to recognize (icmp slt X, 0) and (icmp sgt X, -1).
     if (IC->getPredicate() == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes()))
       IsEqualZero = true;
     if (IC->getPredicate() == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero()))
       IsEqualZero = false;
     else
       return nullptr;
 
-    if (!match(CmpLHS, m_OneUse(m_Trunc(m_Value(V)))))
-      return nullptr;
-
     C1Log = CmpLHS->getType()->getScalarSizeInBits() - 1;
     NeedAnd = true;
   }
 
   Value *Or, *Y;
   const APInt *C2;
   bool NeedXor;
   if (match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2)))) {
     Y = TrueVal;
     Or = FalseVal;

llvm/test/Transforms/InstCombine/select-with-bitwise-ops.ll

;
   %cmp = icmp ne i32 %and, 0
   %and2 = and i32 %y, 2
   %select = select i1 %cmp, i32 %y, i32 %and2
   ret i32 %select
 }
 
 define i8 @test70(i8 %x, i8 %y) {
 ; CHECK-LABEL: @test70(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[X:%.*]], 0
-; CHECK-NEXT:    [[OR:%.*]] = or i8 [[Y:%.*]], 2
-; CHECK-NEXT:    [[SELECT:%.*]] = select i1 [[CMP]], i8 [[OR]], i8 [[Y]]
+; CHECK-NEXT:    [[TMP1:%.*]] = lshr i8 [[X:%.*]], 6
+; CHECK-NEXT:    [[TMP2:%.*]] = and i8 [[TMP1]], 2
+; CHECK-NEXT:    [[SELECT:%.*]] = or i8 [[TMP2]], [[Y:%.*]]
 ; CHECK-NEXT:    ret i8 [[SELECT]]
 ;
   %cmp = icmp slt i8 %x, 0
   %or = or i8 %y, 2
   %select = select i1 %cmp, i8 %or, i8 %y
   ret i8 %select
 }
 
 define i32 @shift_no_xor_multiuse_or(i32 %x, i32 %y) {
 ; CHECK-LABEL: @shift_no_xor_multiuse_or(
 ; CHECK-NEXT:    [[OR:%.*]] = or i32 [[Y:%.*]], 2
 ; CHECK-NEXT:    [[AND:%.*]] = shl i32 [[X:%.*]], 1
 ; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[AND]], 2
 ; CHECK-NEXT:    [[SELECT:%.*]] = or i32 [[TMP1]], [[Y]]
 ; CHECK-NEXT:    [[RES:%.*]] = mul i32 [[SELECT]], [[OR]]