[llvm] 6fe6b29 - AMDGPU: Fix assertion in performSHLPtrCombine for 64-bit pointers

[Matt Arsenault via llvm-commits]

Author: Matt Arsenault
Date: 2020-08-10T13:46:52-04:00
New Revision: 6fe6b29c294e8ad7eb9842e2a5e9b6334ecf9483

URL: https://github.com/llvm/llvm-project/commit/6fe6b29c294e8ad7eb9842e2a5e9b6334ecf9483
DIFF: https://github.com/llvm/llvm-project/commit/6fe6b29c294e8ad7eb9842e2a5e9b6334ecf9483.diff

LOG: AMDGPU: Fix assertion in performSHLPtrCombine for 64-bit pointers

Added: 
    llvm/test/CodeGen/AMDGPU/shl_add_ptr_global.ll

Modified: 
    llvm/lib/Target/AMDGPU/SIISelLowering.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 0edcb38137c1..86bd1a2c7fef 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -8552,7 +8552,7 @@ SDValue SITargetLowering::performSHLPtrCombine(SDNode *N,
   EVT VT = N->getValueType(0);
 
   SDValue ShlX = DAG.getNode(ISD::SHL, SL, VT, N0.getOperand(0), N1);
-  SDValue COffset = DAG.getConstant(Offset, SL, MVT::i32);
+  SDValue COffset = DAG.getConstant(Offset, SL, VT);
 
   SDNodeFlags Flags;
   Flags.setNoUnsignedWrap(N->getFlags().hasNoUnsignedWrap() &&

diff  --git a/llvm/test/CodeGen/AMDGPU/shl_add_ptr_global.ll b/llvm/test/CodeGen/AMDGPU/shl_add_ptr_global.ll
new file mode 100644
index 000000000000..2260de0b74f3
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/shl_add_ptr_global.ll
@@ -0,0 +1,21 @@
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX9 %s
+
+; GCN-LABEL: {{^}}shl_base_global_ptr:
+; GCN: v_add_co_u32_e32 v[[EXTRA_LO:[0-9]+]], vcc, 0x80, v4
+; GCN: v_addc_co_u32_e32 v[[EXTRA_HI:[0-9]+]], vcc, 0, v5, vcc
+; GCN: v_lshlrev_b64 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}}, 2, v[4:5]
+; GCN: v_mov_b32_e32 [[THREE:v[0-9]+]], 3
+; GCN: global_atomic_and v{{\[}}[[LO]]:[[HI]]{{\]}}, [[THREE]], off offset:512
+; GCN: global_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[EXTRA_LO]]:[[EXTRA_HI]]{{\]}}
+define void @shl_base_global_ptr(i32 addrspace(1)* %out, i64 addrspace(1)* %extra.use, [512 x i32] addrspace(1)* %ptr) #0 {
+  %arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(1)* %ptr, i64 0, i64 32
+  %cast = ptrtoint i32 addrspace(1)* %arrayidx0 to i64
+  %shl = shl i64 %cast, 2
+  %castback = inttoptr i64 %shl to i32 addrspace(1)*
+  %val = atomicrmw and i32 addrspace(1)* %castback, i32 3 seq_cst
+  store volatile i64 %cast, i64 addrspace(1)* %extra.use, align 4
+  ret void
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }