[AArch64] Don't materialize 0 with "fmov h0, .." when FullFP16 is not supported.

[SjoerdMeijer]

Summary

We were generating:

fmov h0, wzr

instructions when FullFP16 is not enabled.

I've not added any tests, because the problem was visible in:

test/CodeGen/AArch64/arm64-zero-cycle-zeroing.ll

which I had to change: I don't think Cyclone has FullFP16 enabled
by default, so it shouldn't be using this v8.2a instruction.

[olista01]

LGTM, but please wait a day to let other time zones comment.

[evandro]

Other than the comments below, I second @olista01, LGTM.

test/CodeGen/AArch64/arm64-zero-cycle-zeroing.ll

7

Perhaps adding a test with just -mattr=fullfp16 is in order.

17

Since this line is identical to #23, it could be prefixed with ALL and #23 deleted.

[SjoerdMeijer]

Thanks for the speedy reviews! Feedback addressed, so
this is the diff I will commit my tomorrow morning. I've also
removed these rdar tags, please shout if there are any objections.

lib/Target/AArch64/AArch64ISelLowering.cpp

bool AArch64TargetLowering::isOffsetFoldingLegal(
   DEBUG(dbgs() << "AArch64 doesn't support folding offsets into global "
         "addresses\n");
   return false;
 }
 
 bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
   // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
   // FIXME: We should be able to handle f128 as well with a clever lowering.
-  if (Imm.isPosZero() && (VT == MVT::f16 || VT == MVT::f64 || VT == MVT::f32)) {
+  if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32 ||
+                          (VT == MVT::f16 && Subtarget->hasFullFP16()))) {
     DEBUG(dbgs() << "Legal fp imm: materialize 0 using the zero register\n");
     return true;
   }
 
   StringRef FPType;
   bool IsLegal = false;
   SmallString<128> ImmStrVal;
   Imm.toString(ImmStrVal);

lib/Target/AArch64/AArch64InstrInfo.td

 // Unscaled integer to floating point conversion instruction.
 //===----------------------------------------------------------------------===//
 
 defm FMOV : UnscaledConversion<"fmov">;
 
 // Add pseudo ops for FMOV 0 so we can mark them as isReMaterializable
 let isReMaterializable = 1, isCodeGenOnly = 1, isAsCheapAsAMove = 1 in {
 def FMOVH0 : Pseudo<(outs FPR16:$Rd), (ins), [(set f16:$Rd, (fpimm0))]>,
-    Sched<[WriteF]>;
+    Sched<[WriteF]>, Requires<[HasFullFP16]>;
 def FMOVS0 : Pseudo<(outs FPR32:$Rd), (ins), [(set f32:$Rd, (fpimm0))]>,
     Sched<[WriteF]>;
 def FMOVD0 : Pseudo<(outs FPR64:$Rd), (ins), [(set f64:$Rd, (fpimm0))]>,
     Sched<[WriteF]>;
 }
 // Similarly add aliases
 def : InstAlias<"fmov $Rd, #0.0", (FMOVWHr FPR16:$Rd, WZR), 0>,
     Requires<[HasFullFP16]>;

test/CodeGen/AArch64/arm64-zero-cycle-zeroing.ll

 ; RUN: llc -mtriple=arm64-apple-ios   -mcpu=cyclone   < %s | FileCheck %s -check-prefixes=ALL,CYCLONE
+; RUN: llc -mtriple=arm64-apple-ios   -mcpu=cyclone -mattr=+fullfp16 < %s | FileCheck %s -check-prefixes=CYCLONE-FULLFP16
 ; RUN: llc -mtriple=aarch64-gnu-linux -mcpu=exynos-m1 < %s | FileCheck %s -check-prefixes=ALL,OTHERS
 ; RUN: llc -mtriple=aarch64-gnu-linux -mcpu=exynos-m3 < %s | FileCheck %s -check-prefixes=ALL,OTHERS
 ; RUN: llc -mtriple=aarch64-gnu-linux -mcpu=kryo      < %s | FileCheck %s -check-prefixes=ALL,OTHERS
 ; RUN: llc -mtriple=aarch64-gnu-linux -mcpu=falkor    < %s | FileCheck %s -check-prefixes=ALL,OTHERS
 
-; rdar://11481771
-; rdar://13713797
-
 declare void @bar(half, float, double, <2 x double>)
 declare void @bari(i32, i32)
 declare void @barl(i64, i64)
 declare void @barf(float, float)
 
 define void @t1() nounwind ssp {
 entry:
 ; ALL-LABEL: t1:
 ; ALL-NOT: fmov
-; CYCLONE: fmov h0, wzr
+; ALL:     ldr h0,{{.*}}
 ; CYCLONE: fmov s1, wzr
 ; CYCLONE: fmov d2, xzr
 ; CYCLONE: movi.16b v3, #0
-; OTHERS: movi v{{[0-3]+}}.2d, #0000000000000000
+; CYCLONE-FULLFP16: fmov h0, wzr
+; CYCLONE-FULLFP16: fmov s1, wzr
+; CYCLONE-FULLFP16: fmov d2, xzr
+; CYCLONE-FULLFP16: movi.16b v3, #0
 ; OTHERS: movi v{{[0-3]+}}.2d, #0000000000000000
 ; OTHERS: movi v{{[0-3]+}}.2d, #0000000000000000
 ; OTHERS: movi v{{[0-3]+}}.2d, #0000000000000000
   tail call void @bar(half 0.000000e+00, float 0.000000e+00, double 0.000000e+00, <2 x double> <double 0.000000e+00, double 0.000000e+00>) nounwind
   ret void
 }
 
 define void @t2() nounwind ssp {
; ALL: mov x{{[0-3]+}}, #0
   ret void
 }
 
 define void @t4() nounwind ssp {
 ; ALL-LABEL: t4:
 ; ALL-NOT: fmov
 ; CYCLONE: fmov s{{[0-3]+}}, wzr
 ; CYCLONE: fmov s{{[0-3]+}}, wzr
+; CYCLONE-FULLFP16: fmov s{{[0-3]+}}, wzr
+; CYCLONE-FULLFP16: fmov s{{[0-3]+}}, wzr
 ; OTHERS: movi v{{[0-3]+}}.2d, #0000000000000000
 ; OTHERS: movi v{{[0-3]+}}.2d, #0000000000000000
   tail call void @barf(float 0.000000e+00, float 0.000000e+00) nounwind
   ret void
 }
 
 ; We used to produce spills+reloads for a Q register with zero cycle zeroing
 ; enabled.