[PATCH] D49257: AMDGPU: Fix assert in truncate combine with vectors

Matt Arsenault via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Thu Jul 12 11:03:37 PDT 2018


arsenm created this revision.
arsenm added reviewers: msearles, rampitec.
Herald added subscribers: t-tye, tpr, dstuttard, yaxunl, nhaehnle, wdng, kzhuravl.

https://reviews.llvm.org/D49257

Files:
  lib/Target/AMDGPU/AMDGPUISelLowering.cpp
  test/CodeGen/AMDGPU/trunc-combine.ll


Index: test/CodeGen/AMDGPU/trunc-combine.ll
===================================================================
--- test/CodeGen/AMDGPU/trunc-combine.ll
+++ test/CodeGen/AMDGPU/trunc-combine.ll
@@ -53,3 +53,30 @@
   %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
+}
Index: lib/Target/AMDGPU/AMDGPUISelLowering.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -3055,7 +3055,7 @@
   // 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));


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D49257.155224.patch
Type: text/x-patch
Size: 2289 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180712/95a10647/attachment.bin>


More information about the llvm-commits mailing list