From 156a1b59dfacacacb2c5ac81ee2db4b8e9d46279 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Mon, 24 Feb 2020 15:43:18 -0500
Subject: [PATCH] AMDGPU: Make signext/zeroext behave more sensibly over > i32

Interpret these as extending to the next multiple of 32-bits. This had
no effect with i48 for example, which is really split into {i32, i16},
which should extend the high part.
---
 llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp | 11 ++++++
 llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h   |  3 ++
 llvm/test/CodeGen/AMDGPU/function-returns.ll  | 48 +++++++++++++++++++++++++++
 3 files changed, 62 insertions(+)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 79a0d20..625db96 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -617,6 +617,17 @@ bool AMDGPUTargetLowering::allUsesHaveSourceMods(const SDNode *N,
   return true;
 }
 
+EVT AMDGPUTargetLowering::getTypeForExtReturn(LLVMContext &Context, EVT VT,
+                                              ISD::NodeType ExtendKind) const {
+  assert(!VT.isVector() && "only scalar expected");
+
+  // Round to the next multiple of 32-bits.
+  unsigned Size = VT.getSizeInBits();
+  if (Size <= 32)
+    return MVT::i32;
+  return EVT::getIntegerVT(Context, 32 * ((Size + 31) / 32));
+}
+
 MVT AMDGPUTargetLowering::getVectorIdxTy(const DataLayout &) const {
   return MVT::i32;
 }
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
index ffe5113..a8d5f55 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -178,6 +178,9 @@ public:
 
   bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
 
+  EVT getTypeForExtReturn(LLVMContext &Context, EVT VT,
+                          ISD::NodeType ExtendKind) const override;
+
   MVT getVectorIdxTy(const DataLayout &) const override;
   bool isSelectSupported(SelectSupportKind) const override;
 
diff --git a/llvm/test/CodeGen/AMDGPU/function-returns.ll b/llvm/test/CodeGen/AMDGPU/function-returns.ll
index f40729d..49a6938 100644
--- a/llvm/test/CodeGen/AMDGPU/function-returns.ll
+++ b/llvm/test/CodeGen/AMDGPU/function-returns.ll
@@ -104,6 +104,54 @@ define i48 @i48_func_void() #0 {
   ret i48 %val
 }
 
+; GCN-LABEL: {{^}}i48_zeroext_func_void:
+; GCN: buffer_load_dword v0, off
+; GCN-NEXT: buffer_load_ushort v1, off
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: s_setpc_b64
+define zeroext i48 @i48_zeroext_func_void() #0 {
+  %val = load i48, i48 addrspace(1)* undef, align 8
+  ret i48 %val
+}
+
+; GCN-LABEL: {{^}}i48_signext_func_void:
+; GCN: buffer_load_dword v0, off
+; GCN-NEXT: buffer_load_sshort v1, off
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: s_setpc_b64
+define signext i48 @i48_signext_func_void() #0 {
+  %val = load i48, i48 addrspace(1)* undef, align 8
+  ret i48 %val
+}
+
+; GCN-LABEL: {{^}}i63_func_void:
+; GCN: s_waitcnt
+; GCN-NEXT: s_setpc_b64
+define i63 @i63_func_void(i63 %val) #0 {
+  ret i63 %val
+}
+
+; GCN-LABEL: {{^}}i63_zeroext_func_void:
+; GCN: s_waitcnt
+; GCN-NEXT: v_and_b32_e32 v1, 0x7fffffff, v1
+; GCN-NEXT: s_setpc_b64
+define zeroext i63 @i63_zeroext_func_void(i63 %val) #0 {
+  ret i63 %val
+}
+
+; GCN-LABEL: {{^}}i63_signext_func_void:
+; GCN: s_waitcnt
+; CI-NEXT:	v_lshl_b64 v[0:1], v[0:1], 1
+; CI-NEXT: v_ashr_i64 v[0:1], v[0:1], 1
+
+; GFX89-NEXT:	v_lshlrev_b64 v[0:1], 1, v[0:1]
+; GFX89-NEXT: v_ashrrev_i64 v[0:1], 1, v[0:1]
+
+; GCN-NEXT: s_setpc_b64
+define signext i63 @i63_signext_func_void(i63 %val) #0 {
+  ret i63 %val
+}
+
 ; GCN-LABEL: {{^}}i64_func_void:
 ; GCN: buffer_load_dwordx2 v[0:1], off
 ; GCN-NEXT: s_waitcnt vmcnt(0)
-- 
2.7.4