DAGCombiner: Turn truncate of a bitcasted vector to an extract

[arsenm]

Summary

On AMDGPU where operations i64 operations are often bitcasted to v2i32
and back, this pattern shows up regularly where it breaks some
expected combines on i64, such as load width reducing.

This fixes some test failures in a future commit when i64 loads
are changed to promote.

[• tstellarAMD]

LGTM.

[resistor]

LGTM.

[arsenm]

r262397

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if (NumDefs == 1) {
         SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
         AddToWorklist(NV.getNode());
         Opnds.push_back(NV);
       }
       return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
     }
   }
 
+  // Fold truncate of a bitcast of a vector to an extract of the low vector
+  // element.
+  //
+  // e.g. trunc (i64 (bitcast v2i32:x)) -> extract_vector_elt v2i32:x, 0
+  if (N0.getOpcode() == ISD::BITCAST && !VT.isVector()) {
+    SDValue VecSrc = N0.getOperand(0);
+    EVT SrcVT = VecSrc.getValueType();
+    if (SrcVT.isVector() && SrcVT.getScalarType() == VT) {
+      SDLoc SL(N);
+
+      EVT IdxVT = TLI.getVectorIdxTy(DAG.getDataLayout());
+      return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, VT,
+                         VecSrc, DAG.getConstant(0, SL, IdxVT));
+    }
+  }
+
   // Simplify the operands using demanded-bits information.
   if (!VT.isVector() &&
       SimplifyDemandedBits(SDValue(N, 0)))
     return SDValue(N, 0);
 
   return SDValue();
 }
 

test/CodeGen/AMDGPU/half.ll

define void @global_extload_v2f16_to_v2f64(<2 x double> addrspace(1)* %out, <2 x half> addrspace(1)* %in) #0 {
   %cvt = fpext <2 x half> %val to <2 x double>
   store <2 x double> %cvt, <2 x double> addrspace(1)* %out
   ret void
 }
 
 ; GCN-LABEL: {{^}}global_extload_v3f16_to_v3f64:
 
 ; GCN: buffer_load_dwordx2 [[LOAD:v\[[0-9]+:[0-9]+\]]]
-; GCN: v_lshrrev_b32_e32 {{v[0-9]+}}, 16, {{v[0-9]+}}
-
 ; GCN: v_cvt_f32_f16_e32
 ; GCN: v_cvt_f32_f16_e32
+; GCN-DAG: v_lshrrev_b32_e32 {{v[0-9]+}}, 16, {{v[0-9]+}}
 ; GCN: v_cvt_f32_f16_e32
-; GCN-NOT: v_cvt_f32_f16_e32
+; GCN-NOT: v_cvt_f32_f16
 
 ; GCN: v_cvt_f64_f32_e32
 ; GCN: v_cvt_f64_f32_e32
 ; GCN: v_cvt_f64_f32_e32
 ; GCN-NOT: v_cvt_f64_f32_e32
 
 ; GCN-DAG: buffer_store_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
 ; GCN-DAG: buffer_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 offset:16

test/CodeGen/AMDGPU/trunc-bitcast-vector.ll

+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck %s
+; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck %s
+
+; CHECK-LABEL: {{^}}trunc_i64_bitcast_v2i32:
+; CHECK: buffer_load_dword v
+; CHECK: buffer_store_dword v
+define void @trunc_i64_bitcast_v2i32(i32 addrspace(1)* %out, <2 x i32> addrspace(1)* %in) {
+  %ld = load <2 x i32>, <2 x i32> addrspace(1)* %in
+  %bc = bitcast <2 x i32> %ld to i64
+  %trunc = trunc i64 %bc to i32
+  store i32 %trunc, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: {{^}}trunc_i96_bitcast_v3i32:
+; CHECK: buffer_load_dword v
+; CHECK: buffer_store_dword v
+define void @trunc_i96_bitcast_v3i32(i32 addrspace(1)* %out, <3 x i32> addrspace(1)* %in) {
+  %ld = load <3 x i32>, <3 x i32> addrspace(1)* %in
+  %bc = bitcast <3 x i32> %ld to i96
+  %trunc = trunc i96 %bc to i32
+  store i32 %trunc, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: {{^}}trunc_i128_bitcast_v4i32:
+; CHECK: buffer_load_dword v
+; CHECK: buffer_store_dword v
+define void @trunc_i128_bitcast_v4i32(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
+  %ld = load <4 x i32>, <4 x i32> addrspace(1)* %in
+  %bc = bitcast <4 x i32> %ld to i128
+  %trunc = trunc i128 %bc to i32
+  store i32 %trunc, i32 addrspace(1)* %out
+  ret void
+}
+
+; Don't want load width reduced in this case.
+; CHECK-LABEL: {{^}}trunc_i16_bitcast_v2i16:
+; CHECK: buffer_load_dword [[VAL:v[0-9]+]]
+; CHECK: buffer_store_short [[VAL]]
+define void @trunc_i16_bitcast_v2i16(i16 addrspace(1)* %out, <2 x i16> addrspace(1)* %in) {
+  %ld = load <2 x i16>, <2 x i16> addrspace(1)* %in
+  %bc = bitcast <2 x i16> %ld to i32
+  %trunc = trunc i32 %bc to i16
+  store i16 %trunc, i16 addrspace(1)* %out
+  ret void
+}
+
+; FIXME: Don't want load width reduced here.
+; CHECK-LABEL: {{^}}trunc_i16_bitcast_v4i16:
+; CHECK: buffer_load_ushort [[VAL:v[0-9]+]]
+; CHECK: buffer_store_short [[VAL]]
+define void @trunc_i16_bitcast_v4i16(i16 addrspace(1)* %out, <4 x i16> addrspace(1)* %in) {
+  %ld = load <4 x i16>, <4 x i16> addrspace(1)* %in
+  %bc = bitcast <4 x i16> %ld to i64
+  %trunc = trunc i64 %bc to i16
+  store i16 %trunc, i16 addrspace(1)* %out
+  ret void
+}
+
+; FIXME: Don't want load width reduced in this case.
+; CHECK-LABEL: {{^}}trunc_i8_bitcast_v2i8:
+; CHECK: buffer_load_ubyte [[VAL:v[0-9]+]]
+; CHECK: buffer_store_byte [[VAL]]
+define void @trunc_i8_bitcast_v2i8(i8 addrspace(1)* %out, <2 x i8> addrspace(1)* %in) {
+  %ld = load <2 x i8>, <2 x i8> addrspace(1)* %in
+  %bc = bitcast <2 x i8> %ld to i16
+  %trunc = trunc i16 %bc to i8
+  store i8 %trunc, i8 addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: {{^}}trunc_i32_bitcast_v4i8:
+; CHECK: buffer_load_dword [[VAL:v[0-9]+]]
+; CHECK: buffer_store_byte [[VAL]]
+define void @trunc_i32_bitcast_v4i8(i8 addrspace(1)* %out, <4 x i8> addrspace(1)* %in) {
+  %ld = load <4 x i8>, <4 x i8> addrspace(1)* %in
+  %bc = bitcast <4 x i8> %ld to i32
+  %trunc = trunc i32 %bc to i8
+  store i8 %trunc, i8 addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: {{^}}trunc_i24_bitcast_v3i8:
+; CHECK: buffer_load_dword [[VAL:v[0-9]+]]
+; CHECK: buffer_store_byte [[VAL]]
+define void @trunc_i24_bitcast_v3i8(i8 addrspace(1)* %out, <3 x i8> addrspace(1)* %in) {
+  %ld = load <3 x i8>, <3 x i8> addrspace(1)* %in
+  %bc = bitcast <3 x i8> %ld to i24
+  %trunc = trunc i24 %bc to i8
+  store i8 %trunc, i8 addrspace(1)* %out
+  ret void
+}