AMDGPU/GlobalISel: Legalize more concat_vectors

llvm-svn: 365488

diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
index e65e56d..49e8b36 100644 (file)
--- a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
@@ -82,6 +82,21 @@ static LegalityPredicate numElementsNotEven(unsigned TypeIdx) {
   };
 }
 
+// Any combination of 32 or 64-bit elements up to 512 bits, and multiples of
+// v2s16.
+static LegalityPredicate isRegisterType(unsigned TypeIdx) {
+  return [=](const LegalityQuery &Query) {
+    const LLT Ty = Query.Types[TypeIdx];
+    if (Ty.isVector()) {
+      const int EltSize = Ty.getElementType().getSizeInBits();
+      return EltSize == 32 || EltSize == 64 ||
+            (EltSize == 16 && Ty.getNumElements() % 2 == 0);
+    }
+
+    return Ty.getSizeInBits() % 32 == 0 && Ty.getSizeInBits() <= 512;
+  };
+}
+
 AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
                                          const GCNTargetMachine &TM)
   :  ST(ST_) {
@@ -102,7 +117,6 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
 
   const LLT V2S16 = LLT::vector(2, 16);
   const LLT V4S16 = LLT::vector(4, 16);
-  const LLT V8S16 = LLT::vector(8, 16);
 
   const LLT V2S32 = LLT::vector(2, 32);
   const LLT V3S32 = LLT::vector(3, 32);
@@ -647,19 +661,8 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
                Query.Types[0].getScalarSizeInBits() == 64;
       });
 
-  // TODO: Support any combination of s16, s32, s64, pointer vectors.
   getActionDefinitionsBuilder(G_CONCAT_VECTORS)
-    .legalFor({{V4S32, V2S32},
-               {V8S32, V2S32},
-               {V8S32, V4S32},
-               {V4S64, V2S64},
-               {V4S16, V2S16},
-               {V8S16, V2S16},
-               {V8S16, V4S16},
-               {LLT::vector(4, LocalPtr), LLT::vector(2, LocalPtr)},
-               {LLT::vector(4, PrivatePtr), LLT::vector(2, PrivatePtr)}})
-    // FIXME: Should restrict maximum size, but there seems to be a missing predicate.
-    .legalIf(typeInSet(1, {V2S32, V4S32, V8S32,V2S16, V4S16, V8S16, LLT::vector(16, 16), V2S64}));
+    .legalIf(isRegisterType(0));
 
   // Merge/Unmerge
   for (unsigned Op : {G_MERGE_VALUES, G_UNMERGE_VALUES}) {

diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-concat-vectors.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-concat-vectors.mir
index 4b3b206..123123a 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-concat-vectors.mir
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-concat-vectors.mir
@@ -639,10 +639,10 @@ body: |
     liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 
     ; GCN-LABEL: name: test_concat_vectors_s_v8s64_s_v4s64_s_v4s64
-    ; GCN: [[COPY:%[0-9]+]]:sgpr(<4 x s64>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
-    ; GCN: [[COPY1:%[0-9]+]]:sgpr(<4 x s64>) = COPY $sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN: [[CONCAT_VECTORS:%[0-9]+]]:sgpr(<8 x s64>) = G_CONCAT_VECTORS [[COPY]](<4 x s64>), [[COPY1]](<4 x s64>)
-    ; GCN: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = COPY [[CONCAT_VECTORS]](<8 x s64>)
+    ; GCN: [[COPY:%[0-9]+]]:sreg_256 = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
+    ; GCN: [[COPY1:%[0-9]+]]:sreg_256 = COPY $sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN: [[REG_SEQUENCE:%[0-9]+]]:sreg_512 = REG_SEQUENCE [[COPY]], %subreg.sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7, [[COPY1]], %subreg.sub8_sub9_sub10_sub11_sub12_sub13_sub14_sub15
+    ; GCN: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = COPY [[REG_SEQUENCE]]
     %0:sgpr(<4 x s64>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7,
     %1:sgpr(<4 x s64>) = COPY $sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
     %4:sgpr(<8 x s64>) = G_CONCAT_VECTORS %0, %1
@@ -683,10 +683,10 @@ body: |
     liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4_sgpr5_sgpr6_sgpr7
 
     ; GCN-LABEL: name: test_concat_vectors_s_v4p1_s_v2p1_s_v2p1
-    ; GCN: [[COPY:%[0-9]+]]:sgpr(<2 x p1>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
-    ; GCN: [[COPY1:%[0-9]+]]:sgpr(<2 x p1>) = COPY $sgpr4_sgpr5_sgpr6_sgpr7
-    ; GCN: [[CONCAT_VECTORS:%[0-9]+]]:sgpr(<4 x p1>) = G_CONCAT_VECTORS [[COPY]](<2 x p1>), [[COPY1]](<2 x p1>)
-    ; GCN: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7 = COPY [[CONCAT_VECTORS]](<4 x p1>)
+    ; GCN: [[COPY:%[0-9]+]]:sreg_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    ; GCN: [[COPY1:%[0-9]+]]:sreg_128 = COPY $sgpr4_sgpr5_sgpr6_sgpr7
+    ; GCN: [[REG_SEQUENCE:%[0-9]+]]:sreg_256 = REG_SEQUENCE [[COPY]], %subreg.sub0_sub1_sub2_sub3, [[COPY1]], %subreg.sub4_sub5_sub6_sub7
+    ; GCN: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7 = COPY [[REG_SEQUENCE]]
     %0:sgpr(<2 x p1>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
     %1:sgpr(<2 x p1>) = COPY $sgpr4_sgpr5_sgpr6_sgpr7
     %3:sgpr(<4 x p1>) = G_CONCAT_VECTORS %0, %1
@@ -723,12 +723,12 @@ body: |
     liveins: $sgpr0_sgpr1, $sgpr2_sgpr3, $sgpr4_sgpr5, $sgpr6_sgpr7
 
     ; GCN-LABEL: name: test_concat_vectors_s_v8p3_s_v2p3_s_v2p3_v2p3_s_v2p3
-    ; GCN: [[COPY:%[0-9]+]]:sgpr(<2 x p3>) = COPY $sgpr0_sgpr1
-    ; GCN: [[COPY1:%[0-9]+]]:sgpr(<2 x p3>) = COPY $sgpr2_sgpr3
-    ; GCN: [[COPY2:%[0-9]+]]:sgpr(<2 x p3>) = COPY $sgpr4_sgpr5
-    ; GCN: [[COPY3:%[0-9]+]]:sgpr(<2 x p3>) = COPY $sgpr6_sgpr7
-    ; GCN: [[CONCAT_VECTORS:%[0-9]+]]:sgpr(<8 x p3>) = G_CONCAT_VECTORS [[COPY]](<2 x p3>), [[COPY1]](<2 x p3>), [[COPY2]](<2 x p3>), [[COPY3]](<2 x p3>)
-    ; GCN: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7 = COPY [[CONCAT_VECTORS]](<8 x p3>)
+    ; GCN: [[COPY:%[0-9]+]]:sreg_64_xexec = COPY $sgpr0_sgpr1
+    ; GCN: [[COPY1:%[0-9]+]]:sreg_64_xexec = COPY $sgpr2_sgpr3
+    ; GCN: [[COPY2:%[0-9]+]]:sreg_64_xexec = COPY $sgpr4_sgpr5
+    ; GCN: [[COPY3:%[0-9]+]]:sreg_64_xexec = COPY $sgpr6_sgpr7
+    ; GCN: [[REG_SEQUENCE:%[0-9]+]]:sreg_256 = REG_SEQUENCE [[COPY]], %subreg.sub0_sub1, [[COPY1]], %subreg.sub2_sub3, [[COPY2]], %subreg.sub4_sub5, [[COPY3]], %subreg.sub6_sub7
+    ; GCN: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7 = COPY [[REG_SEQUENCE]]
     %0:sgpr(<2 x p3>) = COPY $sgpr0_sgpr1
     %1:sgpr(<2 x p3>) = COPY $sgpr2_sgpr3
     %2:sgpr(<2 x p3>) = COPY $sgpr4_sgpr5

diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-concat-vectors.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-concat-vectors.mir
index c3e1bca..f5c924b 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-concat-vectors.mir
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-concat-vectors.mir
@@ -127,3 +127,88 @@ body: |
     %2:_(<4 x s64>) = G_CONCAT_VECTORS %0, %1
     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %2
 ...
+
+---
+name: concat_vectors_v2p1_v2p1
+
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4_vgpr5_vgpr6_vgpr7
+    ; CHECK-LABEL: name: concat_vectors_v2p1_v2p1
+    ; CHECK: [[COPY:%[0-9]+]]:_(<2 x p1>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    ; CHECK: [[COPY1:%[0-9]+]]:_(<2 x p1>) = COPY $vgpr4_vgpr5_vgpr6_vgpr7
+    ; CHECK: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x p1>) = G_CONCAT_VECTORS [[COPY]](<2 x p1>), [[COPY1]](<2 x p1>)
+    ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<4 x p1>)
+    %0:_(<2 x p1>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    %1:_(<2 x p1>) = COPY $vgpr4_vgpr5_vgpr6_vgpr7
+    %2:_(<4 x p1>) = G_CONCAT_VECTORS %0, %1
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %2
+...
+
+---
+name: concat_vectors_v2p0_v2p0
+
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4_vgpr5_vgpr6_vgpr7
+    ; CHECK-LABEL: name: concat_vectors_v2p0_v2p0
+    ; CHECK: [[COPY:%[0-9]+]]:_(<2 x p0>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    ; CHECK: [[COPY1:%[0-9]+]]:_(<2 x p0>) = COPY $vgpr4_vgpr5_vgpr6_vgpr7
+    ; CHECK: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x p0>) = G_CONCAT_VECTORS [[COPY]](<2 x p0>), [[COPY1]](<2 x p0>)
+    ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<4 x p0>)
+    %0:_(<2 x p0>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    %1:_(<2 x p0>) = COPY $vgpr4_vgpr5_vgpr6_vgpr7
+    %2:_(<4 x p0>) = G_CONCAT_VECTORS %0, %1
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %2
+...
+
+---
+name: concat_vectors_v2p3_v2p3
+
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1, $vgpr2_vgpr3
+    ; CHECK-LABEL: name: concat_vectors_v2p3_v2p3
+    ; CHECK: [[COPY:%[0-9]+]]:_(<2 x p3>) = COPY $vgpr0_vgpr1
+    ; CHECK: [[COPY1:%[0-9]+]]:_(<2 x p3>) = COPY $vgpr2_vgpr3
+    ; CHECK: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x p3>) = G_CONCAT_VECTORS [[COPY]](<2 x p3>), [[COPY1]](<2 x p3>)
+    ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<4 x p3>)
+    %0:_(<2 x p3>) = COPY $vgpr0_vgpr1
+    %1:_(<2 x p3>) = COPY $vgpr2_vgpr3
+    %2:_(<4 x p3>) = G_CONCAT_VECTORS %0, %1
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %2
+...
+
+---
+name: concat_vectors_v2p5_v2p5
+
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1, $vgpr2_vgpr3
+    ; CHECK-LABEL: name: concat_vectors_v2p5_v2p5
+    ; CHECK: [[COPY:%[0-9]+]]:_(<2 x p5>) = COPY $vgpr0_vgpr1
+    ; CHECK: [[COPY1:%[0-9]+]]:_(<2 x p5>) = COPY $vgpr2_vgpr3
+    ; CHECK: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x p5>) = G_CONCAT_VECTORS [[COPY]](<2 x p5>), [[COPY1]](<2 x p5>)
+    ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<4 x p5>)
+    %0:_(<2 x p5>) = COPY $vgpr0_vgpr1
+    %1:_(<2 x p5>) = COPY $vgpr2_vgpr3
+    %2:_(<4 x p5>) = G_CONCAT_VECTORS %0, %1
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %2
+...
+
+---
+name: concat_vectors_v2p999_v2p999
+
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4_vgpr5_vgpr6_vgpr7
+    ; CHECK-LABEL: name: concat_vectors_v2p999_v2p999
+    ; CHECK: [[COPY:%[0-9]+]]:_(<2 x p999>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    ; CHECK: [[COPY1:%[0-9]+]]:_(<2 x p999>) = COPY $vgpr4_vgpr5_vgpr6_vgpr7
+    ; CHECK: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x p999>) = G_CONCAT_VECTORS [[COPY]](<2 x p999>), [[COPY1]](<2 x p999>)
+    ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<4 x p999>)
+    %0:_(<2 x p999>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    %1:_(<2 x p999>) = COPY $vgpr4_vgpr5_vgpr6_vgpr7
+    %2:_(<4 x p999>) = G_CONCAT_VECTORS %0, %1
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %2
+...