[llvm] [AMDGPU] Do not copy the range attribute onto a widened kernarg load (PR #207655)

[Arseniy Obolenskiy via llvm-commits]

https://github.com/aobolensk created https://github.com/llvm/llvm-project/pull/207655

Sub-dword args (e.g. i8) are loaded as a full i32 and truncated. Copying the i8 range onto the i32 load is invalid IR, and its bounds are wrong since the upper bytes hold other args

Only copy the range when the load type matches the argument type

>From 618301fdfcac564b6df3d4d44b9d256e333e9cf0 Mon Sep 17 00:00:00 2001
From: Arseniy Obolenskiy <arseniy.obolenskiy at amd.com>
Date: Mon, 6 Jul 2026 07:36:18 +0200
Subject: [PATCH] [AMDGPU] Do not copy the range attribute onto a widened
 kernarg load

Sub-dword args (e.g. i8) are loaded as a full i32 and truncated. Copying the i8 range onto the i32 load is invalid IR, and its bounds are wrong since the upper bytes hold other args

Only copy the range when the load type matches the argument type
---
 .../AMDGPU/AMDGPULowerKernelArguments.cpp     |  2 +-
 llvm/test/CodeGen/AMDGPU/lower-kernargs.ll    | 21 +++++++++++++++++++
 2 files changed, 22 insertions(+), 1 deletion(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp b/llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp
index b6168a52bb2ef..cb86728eb70ef 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp
@@ -319,7 +319,7 @@ static bool lowerKernelArguments(Function &F, const TargetMachine &TM,
     if (Arg.hasAttribute(Attribute::NoUndef))
       Load->setMetadata(LLVMContext::MD_noundef, MDNode::get(Ctx, {}));
 
-    if (Arg.hasAttribute(Attribute::Range)) {
+    if (Arg.hasAttribute(Attribute::Range) && AdjustedArgTy == ArgTy) {
       const ConstantRange &Range =
           Arg.getAttribute(Attribute::Range).getValueAsConstantRange();
       Load->setMetadata(LLVMContext::MD_range,
diff --git a/llvm/test/CodeGen/AMDGPU/lower-kernargs.ll b/llvm/test/CodeGen/AMDGPU/lower-kernargs.ll
index e4a7d9dcc837c..ce361e2bc19e2 100644
--- a/llvm/test/CodeGen/AMDGPU/lower-kernargs.ll
+++ b/llvm/test/CodeGen/AMDGPU/lower-kernargs.ll
@@ -274,6 +274,27 @@ define amdgpu_kernel void @kern_range_noundef_i32(i32 noundef range(i32 0, 8) %a
   ret void
 }
 
+define amdgpu_kernel void @kern_range_i8(i8 range(i8 0, 4) %arg0) {
+; HSA-LABEL: @kern_range_i8(
+; HSA-NEXT:    [[KERN_RANGE_I8_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(264) ptr addrspace(4) @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[ARG0_KERNARG_OFFSET_ALIGN_DOWN:%.*]] = getelementptr inbounds i8, ptr addrspace(4) [[KERN_RANGE_I8_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[TMP1:%.*]] = load i32, ptr addrspace(4) [[ARG0_KERNARG_OFFSET_ALIGN_DOWN]], align 16, !invariant.load [[META0]]
+; HSA-NEXT:    [[TMP2:%.*]] = trunc i32 [[TMP1]] to i8
+; HSA-NEXT:    call void (...) @llvm.fake.use(i8 [[TMP2]])
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @kern_range_i8(
+; MESA-NEXT:    [[KERN_RANGE_I8_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(260) ptr addrspace(4) @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[ARG0_KERNARG_OFFSET_ALIGN_DOWN:%.*]] = getelementptr inbounds i8, ptr addrspace(4) [[KERN_RANGE_I8_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[TMP1:%.*]] = load i32, ptr addrspace(4) [[ARG0_KERNARG_OFFSET_ALIGN_DOWN]], align 4, !invariant.load [[META0]]
+; MESA-NEXT:    [[TMP2:%.*]] = trunc i32 [[TMP1]] to i8
+; MESA-NEXT:    call void (...) @llvm.fake.use(i8 [[TMP2]])
+; MESA-NEXT:    ret void
+;
+  call void (...) @llvm.fake.use(i8 %arg0)
+  ret void
+}
+
 define amdgpu_kernel void @kern_f32(float %arg0) {
 ; HSA-LABEL: @kern_f32(
 ; HSA-NEXT:    [[KERN_F32_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(264) ptr addrspace(4) @llvm.amdgcn.kernarg.segment.ptr()