[AArch64] Customer lowering of CTPOP to SIMD should check for NEON availability

[weimingz]

Summary

When passing -mgeneral-regs-only, compiler crashes as it tries to lower CTPOP to "cnt v0.8b..." unless -mno-implicit-float presents. But users may not aware the -mno-implicit-float flag and ICE is not good.

[t.p.northover]

Hi,

The change itself looks reasonable, but I think the tests could be improved:

test/CodeGen/AArch64/arm64-popcnt.ll

12

Can we check that what actually is produced is sensible? Purely negative tests are discouraged when alternatives are possible.

[weimingz]

Add more check in unit test.

[t.p.northover]

Hi Weiming,

Are you sure you updated the patch? It's showing up the same for me.

Cheers.

Tim.

[mcrosier]

In D6299#10, @t.p.northover wrote:

Are you sure you updated the patch? It's showing up the same for me.

Same here.

[weimingz]

Sorry. I uploaded the wrong file in patch set 2.

[t.p.northover]

Thanks Weiming. I think this looks OK.

Tim.

[weimingz]

Hi Tim,

Nice. Thanks for reviewing.
I will commit the change

Weiming

[mcrosier]

Committed in r222292.

lib/Target/AArch64/AArch64ISelLowering.cpp

           AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
     return SDValue();
 
+  if (!Subtarget->hasNEON())
+    return SDValue();
+
   // While there is no integer popcount instruction, it can
   // be more efficiently lowered to the following sequence that uses
   // AdvSIMD registers/instructions as long as the copies to/from

test/CodeGen/AArch64/arm64-popcnt.ll

 ; RUN: llc < %s -march=arm64 -aarch64-neon-syntax=apple | FileCheck %s
+; RUN: llc < %s -march=aarch64 -mattr -neon -aarch64-neon-syntax=apple | FileCheck -check-prefix=CHECK-NONEON %s
 
 define i32 @cnt32_advsimd(i32 %x) nounwind readnone {
   %cnt = tail call i32 @llvm.ctpop.i32(i32 %x)

 ; CHECK: uaddlv.8b  h0, v0
 ; CHECK: fmov w0, s0
 ; CHECK: ret
+; CHECK-NONEON-LABEL: cnt32_advsimd
+; CHECK-NONEON-NOT: 8b
+; CHECK-NONEON: and w{{[0-9]+}}, w{{[0-9]+}}, #0x55555555
+; CHECK-NONEON: and w{{[0-9]+}}, w{{[0-9]+}}, #0x33333333
+; CHECK-NONEON: and w{{[0-9]+}}, w{{[0-9]+}}, #0xf0f0f0f
+; CHECK-NONEON: mul
+
 }
 
 define i64 @cnt64_advsimd(i64 %x) nounwind readnone {

 ; CHECK: uaddlv.8b  h0, v0
 ; CHECK: fmov w0, s0
 ; CHECK: ret
+; CHECK-NONEON-LABEL: cnt64_advsimd
+; CHECK-NONEON-NOT: 8b
+; CHECK-NONEON: and x{{[0-9]+}}, x{{[0-9]+}}, #0x5555555555555555
+; CHECK-NONEON: and x{{[0-9]+}}, x{{[0-9]+}}, #0x3333333333333333
+; CHECK-NONEON: and x{{[0-9]+}}, x{{[0-9]+}}, #0xf0f0f0f0f0f0f0f
+; CHECK-NONEON: mul
 }
 
 ; Do not use AdvSIMD when -mno-implicit-float is specified.