[PATCH] D63937: [ARM] MVE: allow soft-float ABI to pass vector types.

[Simon Tatham via Phabricator via llvm-commits]

simon_tatham created this revision.
simon_tatham added reviewers: dmgreen, ostannard.
Herald added subscribers: llvm-commits, hiraditya, kristof.beyls, javed.absar.
Herald added a project: LLVM.

Passing a vector type over the soft-float ABI involves it being split
into four GPRs, so the first thing that has to happen at the start of
the function is to recombine those into a vector register. The ABI
types all vectors as v2f64, so we need to support BUILD_VECTOR for
that type, which I do in this patch by allowing it to be expanded in
terms of INSERT_VECTOR_ELT, and writing an ISel pattern for that in
turn. Similarly, I provide a rule for EXTRACT_VECTOR_ELT so that a
returned vector can be marshalled back into GPRs.

While I'm here, I've also legalized ISD::UNDEF for all vector types,
because I noticed it was being expanded into a BUILD_VECTOR with
explicit zero inputs, which seems like a waste of effort compared to
the optimal handling of 'just do nothing'.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D63937

Files:
  llvm/lib/Target/ARM/ARMISelLowering.cpp
  llvm/lib/Target/ARM/ARMInstrMVE.td
  llvm/test/CodeGen/Thumb2/mve-soft-float-abi.ll


Index: llvm/test/CodeGen/Thumb2/mve-soft-float-abi.ll
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/Thumb2/mve-soft-float-abi.ll
@@ -0,0 +1,18 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=thumbv8.1m -mattr=+mve -o - %s | FileCheck %s
+
+define <16 x i8> @vector_add(<16 x i8> %lhs, <16 x i8> %rhs) {
+; CHECK-LABEL: vector_add:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vmov d1, r2, r3
+; CHECK-NEXT:    vmov d0, r0, r1
+; CHECK-NEXT:    mov r0, sp
+; CHECK-NEXT:    vldrw.u32 q1, [r0]
+; CHECK-NEXT:    vadd.i8 q0, q0, q1
+; CHECK-NEXT:    vmov r0, r1, d0
+; CHECK-NEXT:    vmov r2, r3, d1
+; CHECK-NEXT:    bx lr
+entry:
+  %sum = add <16 x i8> %lhs, %rhs
+  ret <16 x i8> %sum
+}
Index: llvm/lib/Target/ARM/ARMInstrMVE.td
===================================================================
--- llvm/lib/Target/ARM/ARMInstrMVE.td
+++ llvm/lib/Target/ARM/ARMInstrMVE.td
@@ -1709,6 +1709,11 @@
 def MVE_VMOV_to_lane_8     : MVE_VMOV_lane_8 <  "8", 0b0, MVE_VMOV_to_lane>;
 
 let Predicates = [HasMVEInt] in {
+  def : Pat<(extractelt (v2f64 MQPR:$src), imm:$lane),
+            (f64 (EXTRACT_SUBREG MQPR:$src, (DSubReg_f64_reg imm:$lane)))>;
+  def : Pat<(insertelt (v2f64 MQPR:$src1), DPR:$src2, imm:$lane),
+            (INSERT_SUBREG (v2f64 (COPY_TO_REGCLASS MQPR:$src1, MQPR)), DPR:$src2, (DSubReg_f64_reg imm:$lane))>;
+
   def : Pat<(extractelt (v4i32 MQPR:$src), imm:$lane),
             (COPY_TO_REGCLASS
               (i32 (EXTRACT_SUBREG MQPR:$src, (SSubReg_f32_reg imm:$lane))), rGPR)>;
Index: llvm/lib/Target/ARM/ARMISelLowering.cpp
===================================================================
--- llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -265,6 +265,7 @@
     setOperationAction(ISD::BITCAST, VT, Legal);
     setOperationAction(ISD::LOAD, VT, Legal);
     setOperationAction(ISD::STORE, VT, Legal);
+    setOperationAction(ISD::UNDEF, VT, Legal);
 
     if (HasMVEFP) {
       // No native support for these.
@@ -292,6 +293,10 @@
     setOperationAction(ISD::BITCAST, VT, Legal);
     setOperationAction(ISD::LOAD, VT, Legal);
     setOperationAction(ISD::STORE, VT, Legal);
+    setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
+    setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
+    setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
+    setOperationAction(ISD::UNDEF, VT, Legal);
   }
 
   // It is legal to extload from v4i8 to v4i16 or v4i32.
@@ -6674,7 +6679,7 @@
   // Vectors with 32- or 64-bit elements can be built by directly assigning
   // the subregisters.  Lower it to an ARMISD::BUILD_VECTOR so the operands
   // will be legalized.
-  if (ST->hasNEON() && EltSize >= 32) {
+  if (EltSize >= 32) {
     // Do the expansion with floating-point types, since that is what the VFP
     // registers are defined to use, and since i64 is not legal.
     EVT EltVT = EVT::getFloatingPointVT(EltSize);


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D63937.207075.patch
Type: text/x-patch
Size: 3023 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20190628/72a3a617/attachment.bin>