[AMDGPU][GlobalISel] Handle 64byte EltSIze in getRegSplitParts

Reviewers: arsenm

Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D74518

Author: kerbowa

llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp

@@ -1772,27 +1772,42 @@ ArrayRef<int16_t> SIRegisterInfo::getRegSplitParts(const TargetRegisterClass *RC
     }
   }
 
-  assert(EltSize == 32 && "unhandled elt size");
+  if (EltSize == 32) {
+    static const int16_t Sub0_31_256[] = {
+      AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7,
+      AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15,
+      AMDGPU::sub16_sub17_sub18_sub19_sub20_sub21_sub22_sub23,
+      AMDGPU::sub24_sub25_sub26_sub27_sub28_sub29_sub30_sub31
+    };
 
-  static const int16_t Sub0_31_256[] = {
-    AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7,
-    AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15,
-    AMDGPU::sub16_sub17_sub18_sub19_sub20_sub21_sub22_sub23,
-    AMDGPU::sub24_sub25_sub26_sub27_sub28_sub29_sub30_sub31
-  };
+    static const int16_t Sub0_15_256[] = {
+      AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7,
+      AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15
+    };
+
+    switch (AMDGPU::getRegBitWidth(*RC->MC)) {
+    case 256:
+      return {};
+    case 512:
+      return makeArrayRef(Sub0_15_256);
+    case 1024:
+      return makeArrayRef(Sub0_31_256);
+    default:
+      llvm_unreachable("unhandled register size");
+    }
+  }
 
-  static const int16_t Sub0_15_256[] = {
-    AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7,
-    AMDGPU::sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15
+  assert(EltSize == 64 && "unhandled elt size");
+  static const int16_t Sub0_31_512[] = {
+    AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7_sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15,
+    AMDGPU::sub16_sub17_sub18_sub19_sub20_sub21_sub22_sub23_sub24_sub25_sub26_sub27_sub28_sub29_sub30_sub31
   };
 
   switch (AMDGPU::getRegBitWidth(*RC->MC)) {
-  case 256:
-    return {};
   case 512:
-    return makeArrayRef(Sub0_15_256);
+    return {};
   case 1024:
-    return makeArrayRef(Sub0_31_256);
+    return makeArrayRef(Sub0_31_512);
   default:
     llvm_unreachable("unhandled register size");
   }

llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-concat-vectors.mir

@@ -586,6 +586,26 @@ body: |
     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %8
 ...
 
+---
+name: test_concat_vectors_s_v32s32_s_v16s32_s_v16s32
+legalized:       true
+regBankSelected: true
+
+body: |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
+
+    ; GCN-LABEL: name: test_concat_vectors_s_v32s32_s_v16s32_s_v16s32
+    ; GCN: [[COPY:%[0-9]+]]:sreg_512 = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN: [[COPY1:%[0-9]+]]:sreg_512 = COPY $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
+    ; GCN: [[REG_SEQUENCE:%[0-9]+]]:sreg_1024 = REG_SEQUENCE [[COPY]], %subreg.sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7_sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15, [[COPY1]], %subreg.sub16_sub17_sub18_sub19_sub20_sub21_sub22_sub23_sub24_sub25_sub26_sub27_sub28_sub29_sub30_sub31
+    ; GCN: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = COPY [[REG_SEQUENCE]]
+    %0:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
+    %1:sgpr(<16 x s32>) = COPY $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
+    %2:sgpr(<32 x s32>) = G_CONCAT_VECTORS %0, %1
+    $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = COPY %2
+...
+
 ---
 name: test_concat_vectors_s_v4s64_s_v2s64_s_v2s64
 legalized:       true