[AArch64][GlobalISel] FNeg constant materialization

[davemgreen]

This is a Global ISel equivalent of #80641, creating fneg(movi) instead of the alternative constant pool load or gpr dup.

[llvmbot]

@llvm/pr-subscribers-backend-aarch64

Author: David Green (davemgreen)

Changes

This is a Global ISel equivalent of #80641, creating fneg(movi) instead of the alternative constant pool load or gpr dup.

---

Full diff: https://github.com/llvm/llvm-project/pull/80643.diff

2 Files Affected:

• (modified) llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp (+49-16)
• (modified) llvm/test/CodeGen/AArch64/neon-mov.ll (+10-10)

diff --git a/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp b/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp
index 2515991fbea11..f96b6f5781479 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp
@@ -5786,24 +5786,57 @@ AArch64InstructionSelector::emitConstantVector(Register Dst, Constant *CV,
 
   if (CV->getSplatValue()) {
     APInt DefBits = APInt::getSplat(DstSize, CV->getUniqueInteger());
-    MachineInstr *NewOp;
-    bool Inv = false;
-    if ((NewOp = tryAdvSIMDModImm64(Dst, DstSize, DefBits, MIRBuilder)) ||
-        (NewOp = tryAdvSIMDModImm32(Dst, DstSize, DefBits, MIRBuilder, Inv)) ||
-        (NewOp =
-             tryAdvSIMDModImm321s(Dst, DstSize, DefBits, MIRBuilder, Inv)) ||
-        (NewOp = tryAdvSIMDModImm16(Dst, DstSize, DefBits, MIRBuilder, Inv)) ||
-        (NewOp = tryAdvSIMDModImm8(Dst, DstSize, DefBits, MIRBuilder)) ||
-        (NewOp = tryAdvSIMDModImmFP(Dst, DstSize, DefBits, MIRBuilder)))
-      return NewOp;
+    auto TryMOVIWithBits = [&](APInt DefBits) -> MachineInstr* {
+      MachineInstr *NewOp;
+      bool Inv = false;
+      if ((NewOp = tryAdvSIMDModImm64(Dst, DstSize, DefBits, MIRBuilder)) ||
+          (NewOp = tryAdvSIMDModImm32(Dst, DstSize, DefBits, MIRBuilder, Inv)) ||
+          (NewOp =
+              tryAdvSIMDModImm321s(Dst, DstSize, DefBits, MIRBuilder, Inv)) ||
+          (NewOp = tryAdvSIMDModImm16(Dst, DstSize, DefBits, MIRBuilder, Inv)) ||
+          (NewOp = tryAdvSIMDModImm8(Dst, DstSize, DefBits, MIRBuilder)) ||
+          (NewOp = tryAdvSIMDModImmFP(Dst, DstSize, DefBits, MIRBuilder)))
+        return NewOp;
+
+      DefBits = ~DefBits;
+      Inv = true;
+      if ((NewOp = tryAdvSIMDModImm32(Dst, DstSize, DefBits, MIRBuilder, Inv)) ||
+          (NewOp =
+              tryAdvSIMDModImm321s(Dst, DstSize, DefBits, MIRBuilder, Inv)) ||
+          (NewOp = tryAdvSIMDModImm16(Dst, DstSize, DefBits, MIRBuilder, Inv)))
+        return NewOp;
+      return nullptr;
+    };
 
-    DefBits = ~DefBits;
-    Inv = true;
-    if ((NewOp = tryAdvSIMDModImm32(Dst, DstSize, DefBits, MIRBuilder, Inv)) ||
-        (NewOp =
-             tryAdvSIMDModImm321s(Dst, DstSize, DefBits, MIRBuilder, Inv)) ||
-        (NewOp = tryAdvSIMDModImm16(Dst, DstSize, DefBits, MIRBuilder, Inv)))
+    if (auto *NewOp = TryMOVIWithBits(DefBits))
       return NewOp;
+
+    // See if a fneg of the constant can be materialized with a MOVI, etc
+    auto TryWithFNeg = [&](APInt DefBits, int NumBits,
+                           unsigned NegOpc) -> MachineInstr * {
+      // FNegate each sub-element of the constant
+      APInt Neg = APInt::getHighBitsSet(NumBits, 1).zext(DstSize);
+      APInt NegBits(DstSize, 0);
+      unsigned NumElts = DstSize / NumBits;
+      for (unsigned i = 0; i < NumElts; i++)
+        NegBits |= Neg << (NumBits * i);
+      NegBits = DefBits ^ NegBits;
+
+      // Try to create the new constants with MOVI, and if so generate a fneg
+      // for it.
+      if (auto *NewOp = TryMOVIWithBits(NegBits)) {
+        Register NewDst = MRI.createVirtualRegister(&AArch64::FPR128RegClass);
+        NewOp->getOperand(0).setReg(NewDst);
+        return MIRBuilder.buildInstr(NegOpc, {Dst}, {NewDst});
+      }
+      return nullptr;
+    };
+    MachineInstr *R;
+    if ((R = TryWithFNeg(DefBits, 32, AArch64::FNEGv4f32)) ||
+        (R = TryWithFNeg(DefBits, 64, AArch64::FNEGv2f64)) ||
+        (STI.hasFullFP16() &&
+         (R = TryWithFNeg(DefBits, 16, AArch64::FNEGv8f16))))
+      return R;
   }
 
   auto *CPLoad = emitLoadFromConstantPool(CV, MIRBuilder);
diff --git a/llvm/test/CodeGen/AArch64/neon-mov.ll b/llvm/test/CodeGen/AArch64/neon-mov.ll
index 219c8b53243e6..e7b0296565196 100644
--- a/llvm/test/CodeGen/AArch64/neon-mov.ll
+++ b/llvm/test/CodeGen/AArch64/neon-mov.ll
@@ -125,14 +125,14 @@ define <4 x i32> @movi4s_fneg() {
 ;
 ; CHECK-NOFP16-GI-LABEL: movi4s_fneg:
 ; CHECK-NOFP16-GI:       // %bb.0:
-; CHECK-NOFP16-GI-NEXT:    adrp x8, .LCPI13_0
-; CHECK-NOFP16-GI-NEXT:    ldr q0, [x8, :lo12:.LCPI13_0]
+; CHECK-NOFP16-GI-NEXT:    movi v0.4s, #240, lsl #8
+; CHECK-NOFP16-GI-NEXT:    fneg v0.4s, v0.4s
 ; CHECK-NOFP16-GI-NEXT:    ret
 ;
 ; CHECK-FP16-GI-LABEL: movi4s_fneg:
 ; CHECK-FP16-GI:       // %bb.0:
-; CHECK-FP16-GI-NEXT:    adrp x8, .LCPI13_0
-; CHECK-FP16-GI-NEXT:    ldr q0, [x8, :lo12:.LCPI13_0]
+; CHECK-FP16-GI-NEXT:    movi v0.4s, #240, lsl #8
+; CHECK-FP16-GI-NEXT:    fneg v0.4s, v0.4s
 ; CHECK-FP16-GI-NEXT:    ret
    ret <4 x i32> <i32 2147545088, i32 2147545088, i32 2147545088, i32 2147545088>
 }
@@ -306,8 +306,8 @@ define <8 x i16> @mvni8h_neg() {
 ;
 ; CHECK-FP16-GI-LABEL: mvni8h_neg:
 ; CHECK-FP16-GI:       // %bb.0:
-; CHECK-FP16-GI-NEXT:    adrp x8, .LCPI32_0
-; CHECK-FP16-GI-NEXT:    ldr q0, [x8, :lo12:.LCPI32_0]
+; CHECK-FP16-GI-NEXT:    movi v0.8h, #240
+; CHECK-FP16-GI-NEXT:    fneg v0.8h, v0.8h
 ; CHECK-FP16-GI-NEXT:    ret
    ret <8 x i16> <i16 33008, i16 33008, i16 33008, i16 33008, i16 33008, i16 33008, i16 33008, i16 33008>
 }
@@ -492,14 +492,14 @@ define <2 x double> @fmov2d_neg0() {
 ;
 ; CHECK-NOFP16-GI-LABEL: fmov2d_neg0:
 ; CHECK-NOFP16-GI:       // %bb.0:
-; CHECK-NOFP16-GI-NEXT:    adrp x8, .LCPI51_0
-; CHECK-NOFP16-GI-NEXT:    ldr q0, [x8, :lo12:.LCPI51_0]
+; CHECK-NOFP16-GI-NEXT:    movi v0.2d, #0000000000000000
+; CHECK-NOFP16-GI-NEXT:    fneg v0.2d, v0.2d
 ; CHECK-NOFP16-GI-NEXT:    ret
 ;
 ; CHECK-FP16-GI-LABEL: fmov2d_neg0:
 ; CHECK-FP16-GI:       // %bb.0:
-; CHECK-FP16-GI-NEXT:    adrp x8, .LCPI51_0
-; CHECK-FP16-GI-NEXT:    ldr q0, [x8, :lo12:.LCPI51_0]
+; CHECK-FP16-GI-NEXT:    movi v0.2d, #0000000000000000
+; CHECK-FP16-GI-NEXT:    fneg v0.2d, v0.2d
 ; CHECK-FP16-GI-NEXT:    ret
 	ret <2 x double> <double -0.0, double -0.0>
 }

[davemgreen]

Thanks. I will wait for #80641 before committing this, to make sure we get the extra test coveraged from running through SDAG.