[IR] Handle trunc for ptrtoaddr(inttoptr) cast pair

[nikic]

For ptrtoint(inttoptr) and ptrtoaddr(inttoptr), handle the case where the source and destination size do not match and convert to either zext or trunc. We can't do this if the middle size is smaller than both src/dest, because we'd have to perform an additional masking operation in that case.

Most of these cases are handled by dint of ptrtoint/inttoptr size canonicalization (so I added some unit tests instead). However, the ptrtoaddr(inttoptr) case where the pointer size and address size differ is relevant, as in that case the mismatch in integer sizes is canonical.

[llvmbot]

@llvm/pr-subscribers-llvm-ir

@llvm/pr-subscribers-llvm-transforms

Author: Nikita Popov (nikic)

Changes

For ptrtoint(inttoptr) and ptrtoaddr(inttoptr), handle the case where the source and destination size do not match and convert to either zext or trunc. We can't do this if the middle size is smaller than both src/dest, because we'd have to perform an additional masking operation in that case.

Most of these cases are handled by dint of ptrtoint/inttoptr size canonicalization (so I added some unit tests instead). However, the ptrtoaddr(inttoptr) case where the pointer size and address size differ is relevant, as in that case the mismatch in integer sizes is canonical.

---

Full diff: https://github.com/llvm/llvm-project/pull/162842.diff

3 Files Affected:

• (modified) llvm/lib/IR/Instructions.cpp (+10-6)
• (modified) llvm/test/Transforms/InstCombine/ptrtoaddr.ll (+1-4)
• (modified) llvm/unittests/IR/InstructionsTest.cpp (+24)

diff --git a/llvm/lib/IR/Instructions.cpp b/llvm/lib/IR/Instructions.cpp
index 88e7c44a8b885..9060a89bb15ff 100644
--- a/llvm/lib/IR/Instructions.cpp
+++ b/llvm/lib/IR/Instructions.cpp
@@ -2965,8 +2965,7 @@ unsigned CastInst::isEliminableCastPair(Instruction::CastOps firstOp,
       // zext, sext -> zext, because sext can't sign extend after zext
       return Instruction::ZExt;
     case 11: {
-      // inttoptr, ptrtoint/ptrtoaddr -> bitcast if SrcSize<=PtrSize/AddrSize
-      // and SrcSize==DstSize
+      // inttoptr, ptrtoint/ptrtoaddr -> integer cast
       if (!DL)
         return 0;
       unsigned MidSize = secondOp == Instruction::PtrToAddr
@@ -2974,10 +2973,15 @@ unsigned CastInst::isEliminableCastPair(Instruction::CastOps firstOp,
                              : DL->getPointerTypeSizeInBits(MidTy);
       unsigned SrcSize = SrcTy->getScalarSizeInBits();
       unsigned DstSize = DstTy->getScalarSizeInBits();
-      // TODO: Could also produce zext or trunc here.
-      if (SrcSize <= MidSize && SrcSize == DstSize)
-        return Instruction::BitCast;
-      return 0;
+      // If the middle size is smaller than both source and destination,
+      // an additional masking operation would be required.
+      if (MidSize < SrcSize && MidSize < DstSize)
+        return 0;
+      if (DstSize < SrcSize)
+        return Instruction::Trunc;
+      if (DstSize > SrcSize)
+        return Instruction::ZExt;
+      return Instruction::BitCast;
     }
     case 12:
       // addrspacecast, addrspacecast -> bitcast,       if SrcAS == DstAS
diff --git a/llvm/test/Transforms/InstCombine/ptrtoaddr.ll b/llvm/test/Transforms/InstCombine/ptrtoaddr.ll
index 7b0b152990216..ffaa8b110e32c 100644
--- a/llvm/test/Transforms/InstCombine/ptrtoaddr.ll
+++ b/llvm/test/Transforms/InstCombine/ptrtoaddr.ll
@@ -23,10 +23,7 @@ define i64 @ptrtoaddr_inttoptr_arg(i64 %a) {
 define i32 @ptrtoaddr_inttoptr_arg_addrsize(i32 %a) {
 ; CHECK-LABEL: define i32 @ptrtoaddr_inttoptr_arg_addrsize(
 ; CHECK-SAME: i32 [[A:%.*]]) {
-; CHECK-NEXT:    [[TMP1:%.*]] = zext i32 [[A]] to i64
-; CHECK-NEXT:    [[TOPTR:%.*]] = inttoptr i64 [[TMP1]] to ptr addrspace(1)
-; CHECK-NEXT:    [[TOADDR:%.*]] = ptrtoaddr ptr addrspace(1) [[TOPTR]] to i32
-; CHECK-NEXT:    ret i32 [[TOADDR]]
+; CHECK-NEXT:    ret i32 [[A]]
 ;
   %toptr = inttoptr i32 %a to ptr addrspace(1)
   %toaddr = ptrtoaddr ptr addrspace(1) %toptr to i32
diff --git a/llvm/unittests/IR/InstructionsTest.cpp b/llvm/unittests/IR/InstructionsTest.cpp
index fe9e7e8228490..573cdecb55bc9 100644
--- a/llvm/unittests/IR/InstructionsTest.cpp
+++ b/llvm/unittests/IR/InstructionsTest.cpp
@@ -637,6 +637,30 @@ TEST(InstructionsTest, isEliminableCastPair) {
                                            Int64Ty, &DL1),
             0U);
 
+  // Destination larger than source. Pointer type same as destination.
+  EXPECT_EQ(CastInst::isEliminableCastPair(CastInst::IntToPtr,
+                                           CastInst::PtrToInt, Int16Ty, PtrTy64,
+                                           Int64Ty, &DL1),
+            CastInst::ZExt);
+
+  // Destination larger than source. Pointer type different from destination.
+  EXPECT_EQ(CastInst::isEliminableCastPair(CastInst::IntToPtr,
+                                           CastInst::PtrToInt, Int16Ty, PtrTy32,
+                                           Int64Ty, &DL1),
+            CastInst::ZExt);
+
+  // Destination smaller than source. Pointer type same as source.
+  EXPECT_EQ(CastInst::isEliminableCastPair(CastInst::IntToPtr,
+                                           CastInst::PtrToInt, Int64Ty, PtrTy64,
+                                           Int16Ty, &DL1),
+            CastInst::Trunc);
+
+  // Destination smaller than source. Pointer type different from source.
+  EXPECT_EQ(CastInst::isEliminableCastPair(CastInst::IntToPtr,
+                                           CastInst::PtrToInt, Int64Ty, PtrTy32,
+                                           Int16Ty, &DL1),
+            CastInst::Trunc);
+
   // Test that we don't eliminate bitcasts between different address spaces,
   // or if we don't have available pointer size information.
   DataLayout DL2("e-p:32:32:32-p1:16:16:16-p2:64:64:64-i1:8:8-i8:8:8-i16:16:16"

[dtcxzyw]

LG

[arichardson]

Could you also add a ptrtoaddr test here with e.g. p2:64:64:64:32 data layout? Showing it only works for the address bits?

[XChy]

LGTM