[AArch64][GISel] and+or+shl => bfi

[jroelofs]

Summary

This fixes a GISEL vs SDAG regression that showed up at -Os in 256.bzip2

In _getAndMoveToFrontDecode:

gisel:

and w9, w0, #0xff
orr w9, w9, w8, lsl #8

sdag:

bfi w0, w8, #8, #24

[aemerson]

llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp

2164

Aesthetic opinion but no big deal: it's nicer to use:

if (!Ty.isScalar())
  return false;
if (Size != 32 && Size != 64)
  return false;

and bring the unsigned Size = ... up to here.

2171

Can we avoid re-matching on Dst here? We already know that MI == a G_OR, so we can just match the individual operands.

[jroelofs]

llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp

2164

ooh, I like that better too.

2171

I considered that, but we're getting one more thing from the m_GOR() here: the fact that it matches both orders of its operands via that commutative check in BinaryOp_match::match(). Is there a clean way to write that inline, or should I add another matcher for it?

[paquette]

llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp

2158

Why is this a doxygen comment?

2190

I think emitInstr handles the constraining for you already:

MachineInstr *AArch64InstructionSelector::emitInstr(
...
  constrainSelectedInstRegOperands(*MI, TII, TRI, RBI);
  return &*MI;
}

llvm/test/CodeGen/AArch64/GlobalISel/select-bitfield-insert.ll

2

Is it possible to use a MIR testcase instead?

[jroelofs]

llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp

2158

Copy-pasta from an older version of the patch when it was on a helper function. I'll fix.

2190

👍

llvm/test/CodeGen/AArch64/GlobalISel/select-bitfield-insert.ll

2

IIUC, a MIR test case wouldn't show that we're doing as well as (and in some cases better than) SDAG, but I'm happy to convert it if you think that doesn't matter so much.

[paquette]

llvm/test/CodeGen/AArch64/GlobalISel/select-bitfield-insert.ll

2

Ah, okay, I see. I think it's fine either way.

[aemerson]

LGTM with minor nits addressed, thanks.

llvm/test/CodeGen/AArch64/GlobalISel/select-bitfield-insert.ll

2

I'm ok with either kind of test, the .ll tests do have an advantage in that its easier to see what the semantics of the generated code are.

Should probably add -global-isel-abort=1 to the test though.

llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp

     auto Pred =
         static_cast<CmpInst::Predicate>(Cmp->getOperand(1).getPredicate());
     emitIntegerCompare(Cmp->getOperand(2), Cmp->getOperand(3),
                        Cmp->getOperand(1), MIB);
     emitCSetForICMP(I.getOperand(0).getReg(), Pred, MIB, X);
     I.eraseFromParent();
     return true;
   }
+  case TargetOpcode::G_OR: {
+    // Look for operations that take the lower `Width=Size-ShiftImm` bits of
+    // `ShiftSrc` and insert them into the upper `Width` bits of `MaskSrc` via
+    // shifting and masking that we can replace with a BFI (encoded as a BFM).
+    Register Dst = I.getOperand(0).getReg();
+    LLT Ty = MRI.getType(Dst);
+
+    if (!Ty.isScalar())
+      return false;
+
+    unsigned Size = Ty.getSizeInBits();
+    if (Size != 32 && Size != 64)
+      return false;
+
+    Register ShiftSrc;
+    int64_t ShiftImm;
+    Register MaskSrc;
+    int64_t MaskImm;
+    if (!mi_match(
+            Dst, MRI,
+            m_GOr(m_OneNonDBGUse(m_GShl(m_Reg(ShiftSrc), m_ICst(ShiftImm))),
+                  m_OneNonDBGUse(m_GAnd(m_Reg(MaskSrc), m_ICst(MaskImm))))))
+      return false;
+
+    if (ShiftImm > Size || ((1ULL << ShiftImm) - 1ULL) != uint64_t(MaskImm))
+      return false;
+
+    int64_t Immr = Size - ShiftImm;
+    int64_t Imms = Size - ShiftImm - 1;
+    unsigned Opc = Size == 32 ? AArch64::BFMWri : AArch64::BFMXri;
+    emitInstr(Opc, {Dst}, {MaskSrc, ShiftSrc, Immr, Imms}, MIB);
+    I.eraseFromParent();
+    return true;
+  }
   default:
     return false;
   }
 }
 
 bool AArch64InstructionSelector::select(MachineInstr &I) {
   assert(I.getParent() && "Instruction should be in a basic block!");
   assert(I.getParent()->getParent() && "Instruction should be in a function!");

llvm/test/CodeGen/AArch64/GlobalISel/select-bitfield-insert.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -o - -verify-machineinstrs -global-isel=1 -global-isel-abort=1 | FileCheck %s --check-prefixes=CHECK,GISEL
+; RUN: llc < %s -o - -verify-machineinstrs -global-isel=0 | FileCheck %s --check-prefixes=CHECK,SDAG
+
+target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
+target triple = "arm64-apple-ios14.5.0"
+
+define i32 @bfi_w_31(i32 %in1, i32 %in2) {
+; CHECK-LABEL: bfi_w_31:
+; CHECK:       ; %bb.0: ; %bb
+; CHECK-NEXT:    bfi w1, w0, #31, #1
+; CHECK-NEXT:    mov w0, w1
+; CHECK-NEXT:    ret
+bb:
+  %tmp3 = shl i32 %in1, 31
+  %tmp4 = and i32 %in2, 2147483647
+  %out = or i32 %tmp3, %tmp4
+  ret i32 %out
+}
+
+define i32 @bfi_w_8(i32 %in1, i32 %in2) {
+; CHECK-LABEL: bfi_w_8:
+; CHECK:       ; %bb.0: ; %bb
+; CHECK-NEXT:    bfi w1, w0, #8, #24
+; CHECK-NEXT:    mov w0, w1
+; CHECK-NEXT:    ret
+bb:
+  %tmp3 = shl i32 %in1, 8
+  %tmp4 = and i32 %in2, 255
+  %out = or i32 %tmp3, %tmp4
+  ret i32 %out
+}
+
+define i32 @bfi_w_1(i32 %in1, i32 %in2) {
+; CHECK-LABEL: bfi_w_1:
+; CHECK:       ; %bb.0: ; %bb
+; CHECK-NEXT:    bfi w1, w0, #1, #31
+; CHECK-NEXT:    mov w0, w1
+; CHECK-NEXT:    ret
+bb:
+  %tmp3 = shl i32 %in1, 1
+  %tmp4 = and i32 %in2, 1
+  %out = or i32 %tmp3, %tmp4
+  ret i32 %out
+}
+
+define i64 @bfi_x_63(i64 %in1, i64 %in2) {
+; CHECK-LABEL: bfi_x_63:
+; CHECK:       ; %bb.0: ; %bb
+; CHECK-NEXT:    bfi x1, x0, #63, #1
+; CHECK-NEXT:    mov x0, x1
+; CHECK-NEXT:    ret
+bb:
+  %tmp3 = shl i64 %in1, 63
+  %tmp4 = and i64 %in2, 9223372036854775807
+  %out = or i64 %tmp3, %tmp4
+  ret i64 %out
+}
+
+define i64 @bfi_x_31(i64 %in1, i64 %in2) {
+; CHECK-LABEL: bfi_x_31:
+; CHECK:       ; %bb.0: ; %bb
+; CHECK-NEXT:    bfi x1, x0, #31, #33
+; CHECK-NEXT:    mov x0, x1
+; CHECK-NEXT:    ret
+bb:
+  %tmp3 = shl i64 %in1, 31
+  %tmp4 = and i64 %in2, 2147483647
+  %out = or i64 %tmp3, %tmp4
+  ret i64 %out
+}
+
+define i64 @bfi_x_8(i64 %in1, i64 %in2) {
+; CHECK-LABEL: bfi_x_8:
+; CHECK:       ; %bb.0: ; %bb
+; CHECK-NEXT:    bfi x1, x0, #8, #56
+; CHECK-NEXT:    mov x0, x1
+; CHECK-NEXT:    ret
+bb:
+  %tmp3 = shl i64 %in1, 8
+  %tmp4 = and i64 %in2, 255
+  %out = or i64 %tmp3, %tmp4
+  ret i64 %out
+}
+
+define i64 @bfi_x_1(i64 %in1, i64 %in2) {
+; CHECK-LABEL: bfi_x_1:
+; CHECK:       ; %bb.0: ; %bb
+; CHECK-NEXT:    bfi x1, x0, #1, #63
+; CHECK-NEXT:    mov x0, x1
+; CHECK-NEXT:    ret
+bb:
+  %tmp3 = shl i64 %in1, 1
+  %tmp4 = and i64 %in2, 1
+  %out = or i64 %tmp3, %tmp4
+  ret i64 %out
+}
+
+define i64 @bfi_x_1_swapped(i64 %in1, i64 %in2) {
+; CHECK-LABEL: bfi_x_1_swapped:
+; CHECK:       ; %bb.0: ; %bb
+; CHECK-NEXT:    bfi x1, x0, #1, #63
+; CHECK-NEXT:    mov x0, x1
+; CHECK-NEXT:    ret
+bb:
+  %tmp3 = shl i64 %in1, 1
+  %tmp4 = and i64 %in2, 1
+  %out = or i64 %tmp4, %tmp3
+  ret i64 %out
+}
+
+define i64 @extra_use1(i64 %in1, i64 %in2, i64* %p) {
+; GISEL-LABEL: extra_use1:
+; GISEL:       ; %bb.0: ; %bb
+; GISEL-NEXT:    lsl x8, x0, #1
+; GISEL-NEXT:    and x9, x1, #0x1
+; GISEL-NEXT:    orr x0, x8, x9
+; GISEL-NEXT:    str x8, [x2]
+; GISEL-NEXT:    ret
+;
+; SDAG-LABEL: extra_use1:
+; SDAG:       ; %bb.0: ; %bb
+; SDAG-NEXT:    bfi x1, x0, #1, #63
+; SDAG-NEXT:    lsl x8, x0, #1
+; SDAG-NEXT:    mov x0, x1
+; SDAG-NEXT:    str x8, [x2]
+; SDAG-NEXT:    ret
+bb:
+  %tmp3 = shl i64 %in1, 1
+  %tmp4 = and i64 %in2, 1
+  %out = or i64 %tmp3, %tmp4
+  store i64 %tmp3, i64* %p
+  ret i64 %out
+}
+
+define i64 @extra_use2(i64 %in1, i64 %in2, i64* %p) {
+; GISEL-LABEL: extra_use2:
+; GISEL:       ; %bb.0: ; %bb
+; GISEL-NEXT:    and x8, x1, #0x1
+; GISEL-NEXT:    orr x0, x8, x0, lsl #1
+; GISEL-NEXT:    str x8, [x2]
+; GISEL-NEXT:    ret
+;
+; SDAG-LABEL: extra_use2:
+; SDAG:       ; %bb.0: ; %bb
+; SDAG-NEXT:    and x8, x1, #0x1
+; SDAG-NEXT:    bfi x1, x0, #1, #63
+; SDAG-NEXT:    mov x0, x1
+; SDAG-NEXT:    str x8, [x2]
+; SDAG-NEXT:    ret
+bb:
+  %tmp3 = shl i64 %in1, 1
+  %tmp4 = and i64 %in2, 1
+  %out = or i64 %tmp3, %tmp4
+  store i64 %tmp4, i64* %p
+  ret i64 %out
+}