From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Sat, 2 Feb 2019 23:56:00 +0000 (+0000)
Subject: GlobalISel: Implement widenScalar for G_EXTRACT vector sources
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=757bd101c78fff13b00870638a64c3800a237282;p=llvm

GlobalISel: Implement widenScalar for G_EXTRACT vector sources

Handle the basic element extract case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@352978 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/CodeGen/GlobalISel/LegalizerHelper.cpp b/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
index bd0ad3f8083..f1f7c98eb44 100644
--- a/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
+++ b/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
@@ -942,6 +942,32 @@ LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) {
   switch (MI.getOpcode()) {
   default:
     return UnableToLegalize;
+  case TargetOpcode::G_EXTRACT: {
+    if (TypeIdx != 1)
+      return UnableToLegalize;
+
+    unsigned SrcReg = MI.getOperand(1).getReg();
+    LLT SrcTy = MRI.getType(SrcReg);
+    if (!SrcTy.isVector())
+      return UnableToLegalize;
+
+    unsigned DstReg = MI.getOperand(0).getReg();
+    LLT DstTy = MRI.getType(DstReg);
+    if (DstTy != SrcTy.getElementType())
+      return UnableToLegalize;
+
+    unsigned Offset = MI.getOperand(2).getImm();
+    if (Offset % SrcTy.getScalarSizeInBits() != 0)
+      return UnableToLegalize;
+
+    widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
+
+    MI.getOperand(2).setImm((WideTy.getSizeInBits() / SrcTy.getSizeInBits()) *
+                            Offset);
+    widenScalarDst(MI, WideTy.getScalarType(), 0);
+
+    return Legalized;
+  }
   case TargetOpcode::G_MERGE_VALUES: {
     if (TypeIdx != 1)
       return UnableToLegalize;
diff --git a/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
index 1ce52e9a1a0..509d651eccf 100644
--- a/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
+++ b/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
@@ -439,6 +439,24 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST,
         const LLT &Ty1 = Query.Types[1];
         return (Ty0.getSizeInBits() % 16 == 0) &&
                (Ty1.getSizeInBits() % 16 == 0);
+      })
+    .widenScalarIf(
+      [=](const LegalityQuery &Query) {
+        const LLT &Ty0 = Query.Types[0];
+        const LLT &Ty1 = Query.Types[1];
+        return (Ty1.getScalarSizeInBits() < 16);
+      },
+      // TODO Use generic LegalizeMutation
+      [](const LegalityQuery &Query) {
+        LLT Ty1 = Query.Types[1];
+        unsigned NewEltSizeInBits =
+          std::max(1 << Log2_32_Ceil(Ty1.getScalarSizeInBits()), 16);
+        if (Ty1.isVector()) {
+          return std::make_pair(1, LLT::vector(Ty1.getNumElements(),
+                                               NewEltSizeInBits));
+        }
+
+        return std::make_pair(1, LLT::scalar(NewEltSizeInBits));
       });
 
   // TODO: vectors of pointers
diff --git a/test/CodeGen/AMDGPU/GlobalISel/legalize-extract.mir b/test/CodeGen/AMDGPU/GlobalISel/legalize-extract.mir
index 3aa27eeff6c..75bbcaa6bba 100644
--- a/test/CodeGen/AMDGPU/GlobalISel/legalize-extract.mir
+++ b/test/CodeGen/AMDGPU/GlobalISel/legalize-extract.mir
@@ -103,3 +103,135 @@ body: |
     %1:_(s32) = G_EXTRACT %0, 64
     S_ENDPGM implicit %1
 ...
+
+---
+name: extract_s8_v4s8_offset0
+legalized: true
+
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: extract_s8_v4s8_offset0
+    ; CHECK: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CHECK: [[EXTRACT:%[0-9]+]]:_(s16) = G_EXTRACT [[DEF]](<4 x s16>), 0
+    ; CHECK: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[EXTRACT]](s16)
+    ; CHECK: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(<4 x s8>) = G_IMPLICIT_DEF
+    %1:_(s8) = G_EXTRACT %0, 0
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: extract_s8_v4s8_offset8
+legalized: true
+
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: extract_s8_v4s8_offset8
+    ; CHECK: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CHECK: [[EXTRACT:%[0-9]+]]:_(s16) = G_EXTRACT [[DEF]](<4 x s16>), 16
+    ; CHECK: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[EXTRACT]](s16)
+    ; CHECK: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(<4 x s8>) = G_IMPLICIT_DEF
+    %1:_(s8) = G_EXTRACT %0, 8
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: extract_s8_v4s8_offset16
+legalized: true
+
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: extract_s8_v4s8_offset16
+    ; CHECK: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CHECK: [[EXTRACT:%[0-9]+]]:_(s16) = G_EXTRACT [[DEF]](<4 x s16>), 32
+    ; CHECK: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[EXTRACT]](s16)
+    ; CHECK: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(<4 x s8>) = G_IMPLICIT_DEF
+    %1:_(s8) = G_EXTRACT %0, 16
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: extract_s8_v4s8_offset24
+legalized: true
+
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: extract_s8_v4s8_offset24
+    ; CHECK: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CHECK: [[EXTRACT:%[0-9]+]]:_(s16) = G_EXTRACT [[DEF]](<4 x s16>), 48
+    ; CHECK: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[EXTRACT]](s16)
+    ; CHECK: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(<4 x s8>) = G_IMPLICIT_DEF
+    %1:_(s8) = G_EXTRACT %0, 24
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: extract_s8_v3s8_offset16
+legalized: true
+
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: extract_s8_v3s8_offset16
+    ; CHECK: [[DEF:%[0-9]+]]:_(<3 x s16>) = G_IMPLICIT_DEF
+    ; CHECK: [[EXTRACT:%[0-9]+]]:_(s16) = G_EXTRACT [[DEF]](<3 x s16>), 32
+    ; CHECK: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[EXTRACT]](s16)
+    ; CHECK: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(<3 x s8>) = G_IMPLICIT_DEF
+    %1:_(s8) = G_EXTRACT %0, 16
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: extract_s8_v5s1_offset4
+legalized: true
+
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: extract_s8_v5s1_offset4
+    ; CHECK: [[DEF:%[0-9]+]]:_(<5 x s16>) = G_IMPLICIT_DEF
+    ; CHECK: [[EXTRACT:%[0-9]+]]:_(s16) = G_EXTRACT [[DEF]](<5 x s16>), 80
+    ; CHECK: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[EXTRACT]](s16)
+    ; CHECK: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(<5 x s1>) = G_IMPLICIT_DEF
+    %1:_(s1) = G_EXTRACT %0, 5
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: extract_v2s16_v4s16_offset32
+legalized: true
+
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: extract_v2s16_v4s16_offset32
+    ; CHECK: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CHECK: [[EXTRACT:%[0-9]+]]:_(<2 x s16>) = G_EXTRACT [[DEF]](<4 x s16>), 32
+    ; CHECK: $vgpr0 = COPY [[EXTRACT]](<2 x s16>)
+    %0:_(<4 x s16>) = G_IMPLICIT_DEF
+    %1:_(<2 x s16>) = G_EXTRACT %0, 32
+    $vgpr0 = COPY %1
+...
+
+---
+name: extract_v2s16_v6s16_offset32
+legalized: true
+
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: extract_v2s16_v6s16_offset32
+    ; CHECK: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
+    ; CHECK: [[EXTRACT:%[0-9]+]]:_(<2 x s16>) = G_EXTRACT [[DEF]](<6 x s16>), 32
+    ; CHECK: $vgpr0 = COPY [[EXTRACT]](<2 x s16>)
+    %0:_(<6 x s16>) = G_IMPLICIT_DEF
+    %1:_(<2 x s16>) = G_EXTRACT %0, 32
+    $vgpr0 = COPY %1
+...