[AArch64][GlobalISel] Add basic legalizer rules for supported add/fadd reductions.

[aemerson]

Summary

NEON is pretty limited in it's reduction support. As a first step add some basic rules for the legal types we can select.

[aemerson]

Remove v2s16 FADD case because we can't actually support that with a single instruction. Exactly how we're going to decide what's "legal" and what will need to be legalized first isn't totally clear yet. Using a criteria of a single selected instruction seems a good start.

[aemerson]

Rebase and simplify.

[paquette]

LGTM

llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp

AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
 
   getActionDefinitionsBuilder(G_DYN_STACKALLOC).lower();
 
   getActionDefinitionsBuilder({G_MEMCPY, G_MEMMOVE, G_MEMSET}).libcall();
 
   getActionDefinitionsBuilder(G_ABS).lowerIf(
       [=](const LegalityQuery &Query) { return Query.Types[0].isScalar(); });
 
+  getActionDefinitionsBuilder(G_VECREDUCE_FADD)
+      // We only have FADDP to do reduction-like operations. Lower the rest.
+      .legalFor({{s32, v2s32}, {s64, v2s64}})
+      .lower();
+
+  getActionDefinitionsBuilder(G_VECREDUCE_ADD)
+      .legalFor({{s8, v16s8}, {s16, v8s16}, {s32, v4s32}, {s64, v2s64}})
+      .lower();
+
   computeTables();
   verify(*ST.getInstrInfo());
 }
 
 bool AArch64LegalizerInfo::legalizeCustom(LegalizerHelper &Helper,
                                           MachineInstr &MI) const {
   MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
   MachineRegisterInfo &MRI = *MIRBuilder.getMRI();

llvm/test/CodeGen/AArch64/GlobalISel/legalize-reduce-add.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -O0 -march=aarch64 -run-pass=legalizer -global-isel-abort=1 %s -o - | FileCheck %s
+---
+name:            add_v16s8
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: add_v16s8
+    ; CHECK: liveins: $x0
+    ; CHECK: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK: [[LOAD:%[0-9]+]]:_(<16 x s8>) = G_LOAD [[COPY]](p0) :: (load 16)
+    ; CHECK: [[VECREDUCE_ADD:%[0-9]+]]:_(s8) = G_VECREDUCE_ADD [[LOAD]](<16 x s8>)
+    ; CHECK: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[VECREDUCE_ADD]](s8)
+    ; CHECK: $w0 = COPY [[ANYEXT]](s32)
+    ; CHECK: RET_ReallyLR implicit $w0
+    %0:_(p0) = COPY $x0
+    %1:_(<16 x s8>) = G_LOAD %0(p0) :: (load 16)
+    %2:_(s8) = G_VECREDUCE_ADD %1(<16 x s8>)
+    %3:_(s32) = G_ANYEXT %2(s8)
+    $w0 = COPY %3(s32)
+    RET_ReallyLR implicit $w0
+
+...
+---
+name:            add_v8s16
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: add_v8s16
+    ; CHECK: liveins: $x0
+    ; CHECK: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK: [[LOAD:%[0-9]+]]:_(<8 x s16>) = G_LOAD [[COPY]](p0) :: (load 16)
+    ; CHECK: [[VECREDUCE_ADD:%[0-9]+]]:_(s16) = G_VECREDUCE_ADD [[LOAD]](<8 x s16>)
+    ; CHECK: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[VECREDUCE_ADD]](s16)
+    ; CHECK: $w0 = COPY [[ANYEXT]](s32)
+    ; CHECK: RET_ReallyLR implicit $w0
+    %0:_(p0) = COPY $x0
+    %1:_(<8 x s16>) = G_LOAD %0(p0) :: (load 16)
+    %2:_(s16) = G_VECREDUCE_ADD %1(<8 x s16>)
+    %3:_(s32) = G_ANYEXT %2(s16)
+    $w0 = COPY %3(s32)
+    RET_ReallyLR implicit $w0
+
+...
+---
+name:            add_v4s32
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: add_v4s32
+    ; CHECK: liveins: $x0
+    ; CHECK: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p0) :: (load 16)
+    ; CHECK: [[VECREDUCE_ADD:%[0-9]+]]:_(s32) = G_VECREDUCE_ADD [[LOAD]](<4 x s32>)
+    ; CHECK: $w0 = COPY [[VECREDUCE_ADD]](s32)
+    ; CHECK: RET_ReallyLR implicit $w0
+    %0:_(p0) = COPY $x0
+    %1:_(<4 x s32>) = G_LOAD %0(p0) :: (load 16)
+    %2:_(s32) = G_VECREDUCE_ADD %1(<4 x s32>)
+    $w0 = COPY %2(s32)
+    RET_ReallyLR implicit $w0
+
+...
+---
+name:            add_v2s64
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: add_v2s64
+    ; CHECK: liveins: $x0
+    ; CHECK: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p0) :: (load 16)
+    ; CHECK: [[VECREDUCE_ADD:%[0-9]+]]:_(s64) = G_VECREDUCE_ADD [[LOAD]](<2 x s64>)
+    ; CHECK: $x0 = COPY [[VECREDUCE_ADD]](s64)
+    ; CHECK: RET_ReallyLR implicit $x0
+    %0:_(p0) = COPY $x0
+    %1:_(<2 x s64>) = G_LOAD %0(p0) :: (load 16)
+    %2:_(s64) = G_VECREDUCE_ADD %1(<2 x s64>)
+    $x0 = COPY %2(s64)
+    RET_ReallyLR implicit $x0
+
+...

llvm/test/CodeGen/AArch64/GlobalISel/legalize-reduce-fadd.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -O0 -march=aarch64 -run-pass=legalizer -global-isel-abort=1 %s -o - | FileCheck %s
+
+---
+name:            fadd_v2s32
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    liveins: $d0
+
+    ; CHECK-LABEL: name: fadd_v2s32
+    ; CHECK: liveins: $d0
+    ; CHECK: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $d0
+    ; CHECK: [[VECREDUCE_FADD:%[0-9]+]]:_(s32) = G_VECREDUCE_FADD [[COPY]](<2 x s32>)
+    ; CHECK: $w0 = COPY [[VECREDUCE_FADD]](s32)
+    ; CHECK: RET_ReallyLR implicit $w0
+    %0:_(<2 x s32>) = COPY $d0
+    %1:_(s32) = G_VECREDUCE_FADD %0(<2 x s32>)
+    $w0 = COPY %1(s32)
+    RET_ReallyLR implicit $w0
+
+...
+---
+name:            fadd_v2s64
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    liveins: $q0
+
+    ; CHECK-LABEL: name: fadd_v2s64
+    ; CHECK: liveins: $q0
+    ; CHECK: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $q0
+    ; CHECK: [[VECREDUCE_FADD:%[0-9]+]]:_(s64) = G_VECREDUCE_FADD [[COPY]](<2 x s64>)
+    ; CHECK: $x0 = COPY [[VECREDUCE_FADD]](s64)
+    ; CHECK: RET_ReallyLR implicit $x0
+    %0:_(<2 x s64>) = COPY $q0
+    %2:_(s64) = G_VECREDUCE_FADD %0(<2 x s64>)
+    $x0 = COPY %2(s64)
+    RET_ReallyLR implicit $x0
+
+...

llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir

 # DEBUG-NEXT: .. imm index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: G_VECREDUCE_SEQ_FADD (opcode {{[0-9]+}}): 3 type indices, 0 imm indices
 # DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: G_VECREDUCE_SEQ_FMUL (opcode {{[0-9]+}}): 3 type indices, 0 imm indices
 # DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: G_VECREDUCE_FADD (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
-# DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
-# DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
+# DEBUG-NEXT: .. type index coverage check SKIPPED: user-defined predicate detected
+# DEBUG-NEXT: .. imm index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: G_VECREDUCE_FMUL (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
 # DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: G_VECREDUCE_FMAX (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
 # DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: G_VECREDUCE_FMIN (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
 # DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: G_VECREDUCE_ADD (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
-# DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
-# DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
+# DEBUG-NEXT: .. type index coverage check SKIPPED: user-defined predicate detected
+# DEBUG-NEXT: .. imm index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: G_VECREDUCE_MUL (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
 # DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: G_VECREDUCE_AND (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
 # DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: G_VECREDUCE_OR (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
 # DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined