replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

AMDGPU: Fix assert in truncate combine with vectors 

The piece above probably has the same problem, but I need
to try to come up with a test for it.

llvm-svn: 336935
This commit is contained in:
Matt Arsenault 2018-07-12 19:40:16 +00:00
parent f03f32d478
commit 4dca0a9904
2 changed files with 28 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
View File

@ -3021,7 +3021,7 @@ SDValue AMDGPUTargetLowering::performTruncateCombine(
// Equivalent of above for accessing the high element of a vector as an
// integer operation.
// trunc (srl (bitcast (build_vector x, y))), 16 -> trunc (bitcast y)
if (Src.getOpcode() == ISD::SRL) {
if (Src.getOpcode() == ISD::SRL && !VT.isVector()) {
if (auto K = isConstOrConstSplat(Src.getOperand(1))) {
if (2 * K->getZExtValue() == Src.getValueType().getScalarSizeInBits()) {
SDValue BV = stripBitcast(Src.getOperand(0));

27
llvm/test/CodeGen/AMDGPU/trunc-combine.ll
View File

@ -53,3 +53,30 @@ define i16 @trunc_bitcast_v2f32_to_i16(<2 x float> %bar) {
%add = add i16 %trunc, 4
ret i16 %add
}
; GCN-LABEL: {{^}}truncate_high_elt_extract_vector:
; GCN: s_load_dword s
; GCN: s_load_dword s
; GCN: s_sext_i32_i16
; GCN: s_sext_i32_i16
; GCN: v_mul_i32_i24
; GCN: v_lshrrev_b32_e32
define amdgpu_kernel void @truncate_high_elt_extract_vector(<2 x i16> addrspace(1)* nocapture readonly %arg, <2 x i16> addrspace(1)* nocapture readonly %arg1, <2 x i16> addrspace(1)* nocapture %arg2) local_unnamed_addr {
bb:
%tmp = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %arg, i64 undef
%tmp3 = load <2 x i16>, <2 x i16> addrspace(1)* %tmp, align 4
%tmp4 = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %arg1, i64 undef
%tmp5 = load <2 x i16>, <2 x i16> addrspace(1)* %tmp4, align 4
%tmp6 = sext <2 x i16> %tmp3 to <2 x i32>
%tmp7 = sext <2 x i16> %tmp5 to <2 x i32>
%tmp8 = extractelement <2 x i32> %tmp6, i64 0
%tmp9 = extractelement <2 x i32> %tmp7, i64 0
%tmp10 = mul nsw i32 %tmp9, %tmp8
%tmp11 = insertelement <2 x i32> undef, i32 %tmp10, i32 0
%tmp12 = insertelement <2 x i32> %tmp11, i32 undef, i32 1
%tmp13 = lshr <2 x i32> %tmp12, <i32 16, i32 16>
%tmp14 = trunc <2 x i32> %tmp13 to <2 x i16>
%tmp15 = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %arg2, i64 undef
store <2 x i16> %tmp14, <2 x i16> addrspace(1)* %tmp15, align 4
ret void
}