[Mlir-commits] [mlir] fc5ca83 - [mlir][LLVM] Add support for `ptrtoaddr` (#185104)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Thu Mar 12 04:01:37 PDT 2026
Author: Markus Böck
Date: 2026-03-12T12:01:32+01:00
New Revision: fc5ca83970ac93fbb58f35e6c3c1d61a2093b599
URL: https://github.com/llvm/llvm-project/commit/fc5ca83970ac93fbb58f35e6c3c1d61a2093b599
DIFF: https://github.com/llvm/llvm-project/commit/fc5ca83970ac93fbb58f35e6c3c1d61a2093b599.diff
LOG: [mlir][LLVM] Add support for `ptrtoaddr` (#185104)
The `ptrtoaddr` op is akin to `ptrtoint` with some important
differences:
* It does not capture the provenance of the pointer, meaning a pointer
does not escape and subsequent `inttoptr` don't make a legal pointer.
LLVM can then assume the pointer never escaped, which helps alias
analysis.
* It does not support arbitrary integer types, but only exactly the
integer type that is equal in width to the pointer type as specified by
the data layout.
This PR adds the op the MLIR dialect and adds the corresponding
verification for the datalayout property.
Added:
Modified:
mlir/include/mlir/Dialect/LLVMIR/LLVMOps.td
mlir/lib/Dialect/LLVMIR/IR/LLVMDialect.cpp
mlir/test/Dialect/LLVMIR/invalid.mlir
mlir/test/Dialect/LLVMIR/roundtrip.mlir
mlir/test/Target/LLVMIR/Import/instructions.ll
mlir/test/Target/LLVMIR/llvmir.mlir
Removed:
################################################################################
diff --git a/mlir/include/mlir/Dialect/LLVMIR/LLVMOps.td b/mlir/include/mlir/Dialect/LLVMIR/LLVMOps.td
index 5e1fd69e0dc1c..0b22a3da5c3f0 100644
--- a/mlir/include/mlir/Dialect/LLVMIR/LLVMOps.td
+++ b/mlir/include/mlir/Dialect/LLVMIR/LLVMOps.td
@@ -629,6 +629,50 @@ def LLVM_IntToPtrOp : LLVM_DereferenceableCastOp<"inttoptr", "IntToPtr",
def LLVM_PtrToIntOp : LLVM_CastOp<"ptrtoint", "PtrToInt",
LLVM_ScalarOrVectorOf<LLVM_AnyPointer>,
LLVM_ScalarOrVectorOf<AnySignlessInteger>>;
+
+def LLVM_PtrToAddrOp : LLVM_CastOp<"ptrtoaddr", "PtrToAddr",
+ LLVM_ScalarOrVectorOf<LLVM_AnyPointer>,
+ LLVM_ScalarOrVectorOf<AnySignlessInteger>> {
+ // Overwrite the base class llvmBuilder since the instruction doesn't take a
+ // destination type argument.
+ let llvmBuilder = "$res = builder.CreatePtrToAddr($arg);";
+ let hasVerifier = 1;
+
+ let description = [{
+ Operation mirroring LLVM's `ptrtoaddr` operation.
+
+ This operation casts a pointer (or a vector of pointers) to an integer
+ (or a vector of integers) without capturing the provenance of the pointer.
+ Therefore, an integer returned or derived from `llvm.ptrtoaddr` does not
+ create a legal-to-access pointer when used in `llvm.inttoptr`.
+ Code that only cares about the address value of a pointer
+ (e.g. pointer subtraction) should prefer `llvm.ptrtoaddr` over
+ `llvm.ptrtoint`.
+
+ The integer type used as the result type is required to be equal in width
+ to the pointer type as specified in the data layout.
+ Use the `llvm-target-to-data-layout` pass to derive an MLIR datalayout from
+ an LLVM datalayout.
+
+ Examples:
+ ```
+ llvm.func @default_64_bit_ptrtoaddr(%arg0 : !llvm.ptr) -> i64 {
+ %0 = llvm.ptrtoaddr %arg0 : !llvm.ptr to i64
+ llvm.return i64
+ }
+
+ module attributes { dlti.dl_spec = #dlti.dl_spec<
+ #dlti.dl_entry<!llvm.ptr, dense<[/*size=*/32, 32, 64]> : vector<3xi64>>
+ >} {
+ llvm.func @datalayout_32_bit(%arg0 : !llvm.ptr) -> i32 {
+ %0 = llvm.ptrtoaddr %arg0 : !llvm.ptr to i32
+ llvm.return %0 : i32
+ }
+ }
+ ```
+ }];
+}
+
def LLVM_SExtOp : LLVM_CastOp<"sext", "SExt",
LLVM_ScalarOrVectorOf<AnySignlessInteger>,
LLVM_ScalarOrVectorOf<AnySignlessInteger>> {
diff --git a/mlir/lib/Dialect/LLVMIR/IR/LLVMDialect.cpp b/mlir/lib/Dialect/LLVMIR/IR/LLVMDialect.cpp
index e5bafea345700..d7e844b98dc92 100644
--- a/mlir/lib/Dialect/LLVMIR/IR/LLVMDialect.cpp
+++ b/mlir/lib/Dialect/LLVMIR/IR/LLVMDialect.cpp
@@ -3858,6 +3858,22 @@ LogicalResult LLVM::BitcastOp::verify() {
return success();
}
+LogicalResult LLVM::PtrToAddrOp::verify() {
+ auto pointerType =
+ cast<LLVM::LLVMPointerType>(extractVectorElementType(getArg().getType()));
+ auto integerType = cast<IntegerType>(extractVectorElementType(getType()));
+
+ auto dataLayout = DataLayout::closest(*this);
+ std::optional<unsigned> width = dataLayout.getTypeIndexBitwidth(pointerType);
+ assert(width && "pointers always return an index bitwidth");
+ if (width != integerType.getWidth())
+ return emitOpError("bit-width of integer result type ")
+ << integerType << " must match the pointer bitwidth (" << *width
+ << ") specified in the datalayout";
+
+ return success();
+}
+
//===----------------------------------------------------------------------===//
// Folder for LLVM::AddrSpaceCastOp
//===----------------------------------------------------------------------===//
diff --git a/mlir/test/Dialect/LLVMIR/invalid.mlir b/mlir/test/Dialect/LLVMIR/invalid.mlir
index 5068ddc42e1e5..7f3eb20ce00fa 100644
--- a/mlir/test/Dialect/LLVMIR/invalid.mlir
+++ b/mlir/test/Dialect/LLVMIR/invalid.mlir
@@ -2072,3 +2072,21 @@ module {
// expected-error at +1 {{'llvm.blockaddress' op expects an existing block label target in the referenced function}}
%0 = llvm.blockaddress <function = @missing_func, tag = <id = 1>> : !llvm.ptr
}
+
+// -----
+
+llvm.func @too_narrow_ptr_to_addr(%arg0 : !llvm.ptr) -> () {
+ // expected-error at +1 {{'llvm.ptrtoaddr' op bit-width of integer result type 'i32' must match the pointer bitwidth (64) specified in the datalayout}}
+ %0 = llvm.ptrtoaddr %arg0 : !llvm.ptr to i32
+}
+
+// -----
+
+module attributes { dlti.dl_spec = #dlti.dl_spec<
+ #dlti.dl_entry<!llvm.ptr, dense<[32, 32, 64]> : vector<3xi64>>
+>} {
+ llvm.func @too_narrow_ptr_to_addr(%arg0 : !llvm.ptr) -> () {
+ // expected-error at +1 {{'llvm.ptrtoaddr' op bit-width of integer result type 'i64' must match the pointer bitwidth (32) specified in the datalayout}}
+ %0 = llvm.ptrtoaddr %arg0 : !llvm.ptr to i64
+ }
+}
diff --git a/mlir/test/Dialect/LLVMIR/roundtrip.mlir b/mlir/test/Dialect/LLVMIR/roundtrip.mlir
index f3f1fb69054f3..437d06841f7e4 100644
--- a/mlir/test/Dialect/LLVMIR/roundtrip.mlir
+++ b/mlir/test/Dialect/LLVMIR/roundtrip.mlir
@@ -246,8 +246,12 @@ func.func @ops(%arg0: i32, %arg1: f32,
//
// CHECK: %[[PTR:.*]] = llvm.inttoptr %[[I32]] : i32 to !llvm.ptr
// CHECK: %{{.*}} = llvm.ptrtoint %[[PTR]] : !llvm.ptr to i32
+// CHECK: %{{.*}} = llvm.ptrtoaddr %[[PTR]] : !llvm.ptr to i64
%25 = llvm.inttoptr %arg0 : i32 to !llvm.ptr
%26 = llvm.ptrtoint %25 : !llvm.ptr to i32
+ %a26 = llvm.ptrtoaddr %25 : !llvm.ptr to i64
+ %vp = llvm.mlir.poison : vector<3 x !llvm.ptr>
+ %va26 = llvm.ptrtoaddr %vp : vector<3 x !llvm.ptr> to vector<3 x i64>
// Extended and Quad floating point
//
diff --git a/mlir/test/Target/LLVMIR/Import/instructions.ll b/mlir/test/Target/LLVMIR/Import/instructions.ll
index e171d5ddd474f..6fae74dcf9895 100644
--- a/mlir/test/Target/LLVMIR/Import/instructions.ll
+++ b/mlir/test/Target/LLVMIR/Import/instructions.ll
@@ -181,13 +181,18 @@ define void @integer_extension_and_truncation(i32 %arg1) {
; CHECK-LABEL: @pointer_casts
; CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]
; CHECK-SAME: %[[ARG2:[a-zA-Z0-9]+]]
-define ptr @pointer_casts(ptr %arg1, i64 %arg2) {
+; CHECK-SAME: %[[ARG3:[a-zA-Z0-9]+]]
+define ptr @pointer_casts(ptr %arg1, i64 %arg2, <3 x ptr> %arg3) {
; CHECK: %[[NULL:[0-9]+]] = llvm.mlir.zero : !llvm.ptr
; CHECK: llvm.ptrtoint %[[ARG1]] : !llvm.ptr to i64
+ ; CHECK: llvm.ptrtoaddr %[[ARG1]] : !llvm.ptr to i64
+ ; CHECK: llvm.ptrtoaddr %[[ARG3]] : vector<3x!llvm.ptr> to vector<3xi64>
; CHECK: llvm.inttoptr %[[ARG2]] : i64 to !llvm.ptr
; CHECK: llvm.bitcast %[[ARG1]] : !llvm.ptr to !llvm.ptr
; CHECK: llvm.return %[[NULL]] : !llvm.ptr
%1 = ptrtoint ptr %arg1 to i64
+ %p = ptrtoaddr ptr %arg1 to i64
+ %vp = ptrtoaddr <3 x ptr> %arg3 to <3 x i64>
%2 = inttoptr i64 %arg2 to ptr
%3 = bitcast ptr %arg1 to ptr
ret ptr null
diff --git a/mlir/test/Target/LLVMIR/llvmir.mlir b/mlir/test/Target/LLVMIR/llvmir.mlir
index 8f8a3ee9ef840..26cf2fd1a9853 100644
--- a/mlir/test/Target/LLVMIR/llvmir.mlir
+++ b/mlir/test/Target/LLVMIR/llvmir.mlir
@@ -1283,6 +1283,24 @@ llvm.func @intpointerconversion(%arg0 : i32) -> i32 {
llvm.return %2 : i32
}
+// CHECK-LABEL: @addrpointerconversion_scalar
+// CHECK-SAME: %[[ARG0:[[:alnum:]]+]]
+llvm.func @addrpointerconversion_scalar(%arg0 : !llvm.ptr) -> i64 {
+// CHECK: %[[PTR:.*]] = ptrtoaddr ptr %[[ARG0]] to i64
+// CHECK-NEXT: ret i64 %[[PTR]]
+ %1 = llvm.ptrtoaddr %arg0 : !llvm.ptr to i64
+ llvm.return %1 : i64
+}
+
+// CHECK-LABEL: @addrpointerconversion_vector
+// CHECK-SAME: %[[ARG0:[[:alnum:]]+]]
+llvm.func @addrpointerconversion_vector(%arg0 : vector<3x!llvm.ptr>) -> vector<3x i64> {
+// CHECK: %[[PTR:.*]] = ptrtoaddr <3 x ptr> %[[ARG0]] to <3 x i64>
+// CHECK-NEXT: ret <3 x i64> %[[PTR]]
+ %1 = llvm.ptrtoaddr %arg0 : vector<3x!llvm.ptr> to vector<3x i64>
+ llvm.return %1 : vector<3x i64>
+}
+
llvm.func @fpconversion(%arg0 : i32) -> i32 {
// CHECK: %2 = sitofp i32 %0 to float
// CHECK-NEXT: %3 = fptosi float %2 to i32
More information about the Mlir-commits
mailing list