Fix known zero bits for addrspacecast

[yaxunl]

Summary

Currently LLVM assumes that a pointer addrspacecasted to a different addr space is equivalent to trunc or zext bitwise, which is not true. For example, in amdgcn target, when a null pointer is addrspacecasted from addr space 4 to 0, its value is changed from i64 0 to i32 -1.

This patch teaches LLVM not to assume known bits of addrspacecast instruction to its operand.

[yaxunl]

Simplify the test.

[arsenm]

lib/Analysis/ValueTracking.cpp

1028–1029

AddrSpaceCast should probably just be removed from the switch

[yaxunl]

lib/Analysis/ValueTracking.cpp

1028–1029

Some optimizations will be lost if it is removed.

[arsenm]

lib/Analysis/ValueTracking.cpp

1028–1029

Those optimizations are also broken

[yaxunl]

lib/Analysis/ValueTracking.cpp

1028–1029

But you are right. We cannot assume anything about the bits after addrspacecast.

[yaxunl]

we are going to lose alignment info after addrspacecast, e.g., the following test will fail:

; CHECK-LABEL: @static_hem_addrspacecast(
; CHECK: , align 16
define <2 x i64> @static_hem_addrspacecast() {
  %t = getelementptr <2 x i64>, <2 x i64>* @x, i32 7
  %t.asc = addrspacecast <2 x i64>* %t to <2 x i64> addrspace(1)*
  %tmp1 = load <2 x i64>, <2 x i64> addrspace(1)* %t.asc, align 1
  ret <2 x i64> %tmp1
}

I guess that's a price we need to pay for correctness.

[mcrosier]

test/Analysis/ValueTracking/knownzero-addrspacecast.ll

14

Please add a newline.

[arsenm]

LGTM. I think it should be ok to preserve the alignment though

[yaxunl]

In D26803#599068, @arsenm wrote:

LGTM. I think it should be ok to preserve the alignment though

I think we may need another patch for that. Basically we may need a dedicated parameter for the known zero bits from alignment and pass it through addrspacecast.

test/Analysis/ValueTracking/knownzero-addrspacecast.ll

14

will do when committing.

lib/Analysis/ValueTracking.cpp

   case Instruction::FPExt:
   case Instruction::FPToUI:
   case Instruction::FPToSI:
   case Instruction::SIToFP:
   case Instruction::UIToFP:
     break; // Can't work with floating point.
   case Instruction::PtrToInt:
   case Instruction::IntToPtr:
   case Instruction::AddrSpaceCast: // Pointers could be different sizes.
     // Fall through and handle them the same as zext/trunc.
     LLVM_FALLTHROUGH;
   case Instruction::ZExt:
   case Instruction::Trunc: {
     Type *SrcTy = I->getOperand(0)->getType();
 
     unsigned SrcBitWidth;
     // Note that we handle pointer operands here because of inttoptr/ptrtoint
     // which fall through here.
     SrcBitWidth = Q.DL.getTypeSizeInBits(SrcTy->getScalarType());
 
     assert(SrcBitWidth && "SrcBitWidth can't be zero");
     KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
     KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
     computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth + 1, Q);
     KnownZero = KnownZero.zextOrTrunc(BitWidth);
     KnownOne = KnownOne.zextOrTrunc(BitWidth);
-    // Any top bits are known to be zero.
-    if (BitWidth > SrcBitWidth)
+    // AddrSpaceCasting to a wider pointer does not guarantee newly added higher
+    // bits to be zero. For other instructions it is true.
+    if (BitWidth > SrcBitWidth && I->getOpcode() != Instruction::AddrSpaceCast)
       KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
     break;
   }
   case Instruction::BitCast: {
     Type *SrcTy = I->getOperand(0)->getType();
     if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
         // TODO: For now, not handling conversions like:
         // (bitcast i64 %x to <2 x i32>)

test/Analysis/ValueTracking/knownzero-addrspacecast.ll

+; RUN: opt -instsimplify -S < %s | FileCheck %s
+
+target datalayout = "p:32:32-p4:64:64"
+
+; When a pointer is addrspacecasted to a wider pointer, there is no guarantee
+; that the newly added high bits are zero.
+; CHECK-NOT: ret i64 0
+define i64 @test(i8* %p) {
+  %g = addrspacecast i8* %p to i8 addrspace(4)*
+  %i = ptrtoint i8 addrspace(4)* %g to i64
+  %shift = lshr i64 %i, 32
+  ret i64 %shift
+}
 No newline at end of file