[PATCH] D89075: [AArch64][GlobalISel] Regbankselect reductions to use FPR bank for scalars.

[Amara Emerson via Phabricator via llvm-commits]

aemerson created this revision.
aemerson added a reviewer: paquette.
aemerson added a project: LLVM.
Herald added subscribers: danielkiss, hiraditya, kristof.beyls, rovka.
aemerson requested review of this revision.

Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D89075

Files:
  llvm/lib/Target/AArch64/GISel/AArch64RegisterBankInfo.cpp
  llvm/test/CodeGen/AArch64/GlobalISel/regbankselect-reductions.mir


Index: llvm/test/CodeGen/AArch64/GlobalISel/regbankselect-reductions.mir
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/AArch64/GlobalISel/regbankselect-reductions.mir
@@ -0,0 +1,45 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=aarch64-unknown-unknown -verify-machineinstrs -O0 -run-pass=regbankselect -global-isel-abort=1 %s -o - | FileCheck %s
+
+---
+name:            fadd_v2s32
+legalized:       true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    liveins: $d0, $s1
+
+    ; CHECK-LABEL: name: fadd_v2s32
+    ; CHECK: liveins: $d0, $s1
+    ; CHECK: [[COPY:%[0-9]+]]:fpr(<2 x s32>) = COPY $d0
+    ; CHECK: %scalar:fpr(s32) = COPY $s1
+    ; CHECK: [[VECREDUCE_FADD:%[0-9]+]]:fpr(s32) = G_VECREDUCE_FADD %scalar(s32), [[COPY]](<2 x s32>)
+    ; CHECK: $w0 = COPY [[VECREDUCE_FADD]](s32)
+    ; CHECK: RET_ReallyLR implicit $w0
+    %0:_(<2 x s32>) = COPY $d0
+    %scalar:_(s32) = COPY $s1
+    %1:_(s32) = G_VECREDUCE_FADD %scalar(s32), %0(<2 x s32>)
+    $w0 = COPY %1(s32)
+    RET_ReallyLR implicit $w0
+
+...
+---
+name:            add_v4s32
+legalized:       true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    liveins: $q0
+
+    ; CHECK-LABEL: name: add_v4s32
+    ; CHECK: liveins: $q0
+    ; CHECK: [[COPY:%[0-9]+]]:fpr(<4 x s32>) = COPY $q0
+    ; CHECK: [[VECREDUCE_ADD:%[0-9]+]]:fpr(s32) = G_VECREDUCE_ADD [[COPY]](<4 x s32>)
+    ; CHECK: $w0 = COPY [[VECREDUCE_ADD]](s32)
+    ; CHECK: RET_ReallyLR implicit $w0
+    %0:_(<4 x s32>) = COPY $q0
+    %1:_(s32) = G_VECREDUCE_ADD %0(<4 x s32>)
+    $w0 = COPY %1(s32)
+    RET_ReallyLR implicit $w0
+
+...
Index: llvm/lib/Target/AArch64/GISel/AArch64RegisterBankInfo.cpp
===================================================================
--- llvm/lib/Target/AArch64/GISel/AArch64RegisterBankInfo.cpp
+++ llvm/lib/Target/AArch64/GISel/AArch64RegisterBankInfo.cpp
@@ -845,7 +845,7 @@
     }
     break;
   }
-  case TargetOpcode::G_BUILD_VECTOR:
+  case TargetOpcode::G_BUILD_VECTOR: {
     // If the first source operand belongs to a FPR register bank, then make
     // sure that we preserve that.
     if (OpRegBankIdx[1] != PMI_FirstGPR)
@@ -878,6 +878,33 @@
     break;
   }
 
+  case TargetOpcode::G_VECREDUCE_FMAX:
+  case TargetOpcode::G_VECREDUCE_FMIN:
+  case TargetOpcode::G_VECREDUCE_ADD:
+  case TargetOpcode::G_VECREDUCE_MUL:
+  case TargetOpcode::G_VECREDUCE_AND:
+  case TargetOpcode::G_VECREDUCE_OR:
+  case TargetOpcode::G_VECREDUCE_XOR:
+  case TargetOpcode::G_VECREDUCE_SMAX:
+  case TargetOpcode::G_VECREDUCE_SMIN:
+  case TargetOpcode::G_VECREDUCE_UMAX:
+  case TargetOpcode::G_VECREDUCE_UMIN:
+    // Reductions produce a scalar value from a vector, the scalar should be on
+    // FPR bank.
+    OpRegBankIdx[0] = PMI_FirstFPR;
+    OpRegBankIdx[1] = PMI_FirstFPR;
+    break;
+  case TargetOpcode::G_VECREDUCE_SEQ_FADD:
+  case TargetOpcode::G_VECREDUCE_SEQ_FMUL:
+  case TargetOpcode::G_VECREDUCE_FADD:
+  case TargetOpcode::G_VECREDUCE_FMUL:
+    // These reductions also take a scalar accumulator input.
+    // Assign them FPR for now.
+    OpRegBankIdx[0] = PMI_FirstFPR;
+    OpRegBankIdx[1] = PMI_FirstFPR;
+    OpRegBankIdx[2] = PMI_FirstFPR;
+  }
+
   // Finally construct the computed mapping.
   SmallVector<const ValueMapping *, 8> OpdsMapping(NumOperands);
   for (unsigned Idx = 0; Idx < NumOperands; ++Idx) {


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D89075.297056.patch
Type: text/x-patch
Size: 3455 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20201008/b5e6461b/attachment.bin>