[AArch64][SME]: Generate streaming-compatible code for ld2-alloca.

[hassnaa-arm]

Summary

To generate code compatible to streaming mode:

• disable lowering interleaved load to avoid generating invalid NEON intrinsics.

[sdesmalen]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

13968–13969

This also allows the case where the total VecSize == 32 (for e.g. <4 x i8> which is currently not supported by Neon), or whether the number of elements is not a power of 2 (e.g. <6 x i8>. Can you add a test for this case?

13969

The indentation seems weird, did you use clang-format?

llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-ld2-alloca.ll

114

nit: can you add nounwind as one of the parameters, to avoid the .cfi directives in the output?

[hassnaa-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

13968–13969

In the IR, I don't understand why it allocates 16 elements while it only loads 8.
Is there a reason behind that ?

[hassnaa-arm]

Add nounwind to avoid .cfi directives get generated.

[hassnaa-arm]

Add extra test cases

[sdesmalen]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

13968–13969

Based on the code you've added, I expected the tests you've added to use an ld2, do you know why it doesn't?

[hassnaa-arm]

Change the used mask in the test cases, so that interleaved load get used.

[sdesmalen]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

13970

Sorry, I just realise you'll also need to add a check that we can generate a predicate pattern for the number of elements (e.g. vl2, vl3, ...), because e.g. a <9 x i8> has no corresponding predicate pattern. You can use Optional<unsigned> getSVEPredPatternFromNumElements() for this (defined in Utils/AArch64BaseInfo.h).

Can you also add a test for <9 x i8> ?

[hassnaa-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

13970

Sorry, I don't understand why I should check for e.g. a <9 x i8> ? How is that related to the condition in the code ?
and what do you mean by "You can use Optional<unsigned> getSVEPredPatternFromNumElements() for this " ?
Do you mean that in addition to adding the check, I will add a code change also ?

[sdesmalen]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

13970

What I meant was that if you do a shuffle mask like this:

%load = load <6 x i8>, ptr %alloc
%strided.vec = shufflevector <6 x i8> %load, <6 x i8> poison, <3 x i32> <i32 1, i32 3, i32 5>

You get the following LD2 instruction:

ptrue p0.b, vl3
ld2b { z0.b, z1.b }, p0/z, [x8]

Which uses vl3 for the predicate, which means enable 3 lanes. But there is no vl9, so if you'd end up with NumElements == 9, then you can't code-generate the interleaved access using LD2. To ask if there is a ptrue predicate for NumElements, you can use getSVEPredPatternFromNumElements.

(I meant <9 x i8> as the result type of the shuffle by the way)

[hassnaa-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

13970

you mean if the result type of the shuffle is <9 x i8> there will be a problem, and you are asking me to add a test cases for that and fix its problem, correct ?

[sdesmalen]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

13970

Correct.

[hassnaa-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

13970

but why did you choose vl9 specifically ? what about other vl ? e.g. vl10.

[sdesmalen]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

13970

Because the available predicate patterns are vl1, vl2, vl3, ..., vl8, vl16, ... So vl9 is the first non-power-of-2 vector length that can't be represented.

[hassnaa-arm]

Skip using interleaved_load for vl9, because vl9 predicate is not available.

[sdesmalen]

LGTM with nit addressed

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

13962

nit: you can just do !getSVEPredPatternFromNumElements(NumElements))

[hassnaa-arm]

Add check for hasSVE() to avoid affecting NEON interleaved-access tests

[FreddyYe]

Failed Tests (1):

LLVM :: CodeGen/AArch64/sve-streaming-mode-fixed-length-ld2-alloca.ll

Testing Time: 345.12s

Skipped          :    38
Unsupported      :  1357
Passed           : 88906
Expectedly Failed:   195
Failed           :     1

https://lab.llvm.org/buildbot/#/builders/85/builds/12665

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

bool AArch64TargetLowering::isLegalInterleavedAccessType(
 
   unsigned VecSize = DL.getTypeSizeInBits(VecTy);
   unsigned ElSize = DL.getTypeSizeInBits(VecTy->getElementType());
   unsigned NumElements = cast<FixedVectorType>(VecTy)->getNumElements();
 
   UseScalable = false;
 
   // Ensure the number of vector elements is greater than 1.
   if (NumElements < 2)
     return false;
 
   // Ensure the element type is legal.
   if (ElSize != 8 && ElSize != 16 && ElSize != 32 && ElSize != 64)
     return false;
 
-  if (Subtarget->useSVEForFixedLengthVectors() &&
+  if (Subtarget->forceStreamingCompatibleSVE() ||
+      (Subtarget->useSVEForFixedLengthVectors() &&
        (VecSize % Subtarget->getMinSVEVectorSizeInBits() == 0 ||
         (VecSize < Subtarget->getMinSVEVectorSizeInBits() &&
-        isPowerOf2_32(NumElements) && VecSize > 128))) {
+         isPowerOf2_32(NumElements) && VecSize > 128)))) {
     UseScalable = true;
     return true;
   }
 
   // Ensure the total vector size is 64 or a multiple of 128. Types larger than
   // 128 will be split into multiple interleaved accesses.
   return VecSize == 64 || VecSize % 128 == 0;
 }

llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-ld2-alloca.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -force-streaming-compatible-sve < %s | FileCheck %s
 
 target triple = "aarch64-unknown-linux-gnu"
 
 declare void @def(ptr)
 
 define void @alloc_v4i8(ptr %st_ptr) #0 {
 ; CHECK-LABEL: alloc_v4i8:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    sub sp, sp, #32
 ; CHECK-NEXT:    stp x30, x19, [sp, #16] // 16-byte Folded Spill
 ; CHECK-NEXT:    mov x19, x0
 ; CHECK-NEXT:    add x0, sp, #12
 ; CHECK-NEXT:    bl def
-; CHECK-NEXT:    ldr s0, [sp, #12]
-; CHECK-NEXT:    ptrue p0.h, vl4
-; CHECK-NEXT:    uunpklo z0.h, z0.b
-; CHECK-NEXT:    fmov w8, s0
-; CHECK-NEXT:    mov z1.h, z0.h[3]
-; CHECK-NEXT:    mov z2.h, z0.h[1]
-; CHECK-NEXT:    mov z0.h, z0.h[2]
-; CHECK-NEXT:    fmov w9, s1
-; CHECK-NEXT:    fmov w10, s2
-; CHECK-NEXT:    strh w8, [sp]
+; CHECK-NEXT:    add x8, sp, #12
+; CHECK-NEXT:    ptrue p0.b, vl2
+; CHECK-NEXT:    ld2b { z0.b, z1.b }, p0/z, [x8]
+; CHECK-NEXT:    ptrue p0.s, vl2
+; CHECK-NEXT:    mov z2.b, z0.b[1]
 ; CHECK-NEXT:    fmov w8, s0
-; CHECK-NEXT:    strh w9, [sp, #6]
-; CHECK-NEXT:    strh w10, [sp, #4]
-; CHECK-NEXT:    strh w8, [sp, #2]
+; CHECK-NEXT:    fmov w9, s2
+; CHECK-NEXT:    stp w8, w9, [sp]
 ; CHECK-NEXT:    ldr d0, [sp]
-; CHECK-NEXT:    st1b { z0.h }, p0, [x19]
+; CHECK-NEXT:    st1b { z0.s }, p0, [x19]
 ; CHECK-NEXT:    ldp x30, x19, [sp, #16] // 16-byte Folded Reload
 ; CHECK-NEXT:    add sp, sp, #32
 ; CHECK-NEXT:    ret
   %alloc = alloca [4 x i8]
   call void @def(ptr %alloc)
   %load = load <4 x i8>, ptr %alloc
-  %strided.vec = shufflevector <4 x i8> %load, <4 x i8> poison, <4 x i32> <i32 0, i32 2, i32 1, i32 3>
-  store <4 x i8> %strided.vec, ptr %st_ptr
+  %strided.vec = shufflevector <4 x i8> %load, <4 x i8> poison, <2 x i32> <i32 0, i32 2>
+  store <2 x i8> %strided.vec, ptr %st_ptr
   ret void
 }
 
 define void @alloc_v6i8(ptr %st_ptr) #0 {
 ; CHECK-LABEL: alloc_v6i8:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    sub sp, sp, #32
-; CHECK-NEXT:    stp x30, x19, [sp, #16] // 16-byte Folded Spill
+; CHECK-NEXT:    sub sp, sp, #48
+; CHECK-NEXT:    stp x30, x19, [sp, #32] // 16-byte Folded Spill
 ; CHECK-NEXT:    mov x19, x0
-; CHECK-NEXT:    add x0, sp, #8
+; CHECK-NEXT:    add x0, sp, #24
 ; CHECK-NEXT:    bl def
-; CHECK-NEXT:    ldr d0, [sp, #8]
-; CHECK-NEXT:    mov z1.b, z0.b[4]
-; CHECK-NEXT:    mov z2.b, z0.b[5]
-; CHECK-NEXT:    fmov w8, s0
-; CHECK-NEXT:    fmov w9, s1
-; CHECK-NEXT:    fmov w10, s2
-; CHECK-NEXT:    mov z3.b, z0.b[3]
-; CHECK-NEXT:    mov z4.b, z0.b[1]
-; CHECK-NEXT:    mov z0.b, z0.b[2]
-; CHECK-NEXT:    strb w8, [sp]
-; CHECK-NEXT:    fmov w8, s3
-; CHECK-NEXT:    strb w9, [sp, #5]
-; CHECK-NEXT:    fmov w9, s4
-; CHECK-NEXT:    strb w10, [sp, #4]
-; CHECK-NEXT:    fmov w10, s0
-; CHECK-NEXT:    strb w8, [sp, #3]
-; CHECK-NEXT:    strb w9, [sp, #2]
-; CHECK-NEXT:    strb w10, [sp, #1]
-; CHECK-NEXT:    ldr d0, [sp]
-; CHECK-NEXT:    mov z1.h, z0.h[2]
+; CHECK-NEXT:    add x8, sp, #24
+; CHECK-NEXT:    ptrue p0.b, vl3
+; CHECK-NEXT:    ld2b { z0.b, z1.b }, p0/z, [x8]
+; CHECK-NEXT:    ptrue p0.h, vl4
+; CHECK-NEXT:    fmov w8, s1
+; CHECK-NEXT:    mov z2.b, z1.b[3]
+; CHECK-NEXT:    mov z3.b, z1.b[2]
+; CHECK-NEXT:    mov z0.b, z1.b[1]
+; CHECK-NEXT:    fmov w9, s2
+; CHECK-NEXT:    fmov w10, s3
+; CHECK-NEXT:    strh w8, [sp, #8]
 ; CHECK-NEXT:    fmov w8, s0
-; CHECK-NEXT:    fmov w9, s1
-; CHECK-NEXT:    str w8, [x19]
-; CHECK-NEXT:    strh w9, [x19, #4]
-; CHECK-NEXT:    ldp x30, x19, [sp, #16] // 16-byte Folded Reload
-; CHECK-NEXT:    add sp, sp, #32
+; CHECK-NEXT:    strh w9, [sp, #14]
+; CHECK-NEXT:    strh w10, [sp, #12]
+; CHECK-NEXT:    strh w8, [sp, #10]
+; CHECK-NEXT:    add x8, sp, #20
+; CHECK-NEXT:    ldr d0, [sp, #8]
+; CHECK-NEXT:    st1b { z0.h }, p0, [x8]
+; CHECK-NEXT:    ldrh w8, [sp, #20]
+; CHECK-NEXT:    strb w10, [x19, #2]
+; CHECK-NEXT:    strh w8, [x19]
+; CHECK-NEXT:    ldp x30, x19, [sp, #32] // 16-byte Folded Reload
+; CHECK-NEXT:    add sp, sp, #48
 ; CHECK-NEXT:    ret
   %alloc = alloca [6 x i8]
   call void @def(ptr %alloc)
   %load = load <6 x i8>, ptr %alloc
-  %strided.vec = shufflevector <6 x i8> %load, <6 x i8> poison, <6 x i32> <i32 0, i32 2, i32 1, i32 3, i32 5, i32 4>
-  store <6 x i8> %strided.vec, ptr %st_ptr
+  %strided.vec = shufflevector <6 x i8> %load, <6 x i8> poison, <3 x i32> <i32 1, i32 3, i32 5>
+  store <3 x i8> %strided.vec, ptr %st_ptr
   ret void
 }
 
 define void @alloc_v8f64(ptr %st_ptr) #0 {
 ; CHECK-LABEL: alloc_v8f64:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    sub sp, sp, #96
-; CHECK-NEXT:    stp x20, x19, [sp, #80] // 16-byte Folded Spill
+; CHECK-NEXT:    str x29, [sp, #-32]! // 8-byte Folded Spill
+; CHECK-NEXT:    stp x30, x19, [sp, #16] // 16-byte Folded Spill
+; CHECK-NEXT:    addvl sp, sp, #-1
+; CHECK-NEXT:    sub sp, sp, #64
 ; CHECK-NEXT:    mov x19, x0
 ; CHECK-NEXT:    mov x0, sp
-; CHECK-NEXT:    str x30, [sp, #64] // 8-byte Folded Spill
-; CHECK-NEXT:    mov x20, sp
 ; CHECK-NEXT:    bl def
-; CHECK-NEXT:    ld2 { v0.2d, v1.2d }, [x20], #32
-; CHECK-NEXT:    ldr x30, [sp, #64] // 8-byte Folded Reload
-; CHECK-NEXT:    ld2 { v2.2d, v3.2d }, [x20]
+; CHECK-NEXT:    cntd x8
+; CHECK-NEXT:    ptrue p0.d, vl4
+; CHECK-NEXT:    sub x8, x8, #2
+; CHECK-NEXT:    ld2d { z0.d, z1.d }, p0/z, [sp]
+; CHECK-NEXT:    mov w9, #2
+; CHECK-NEXT:    cmp x8, #2
+; CHECK-NEXT:    csel x8, x8, x9, lo
+; CHECK-NEXT:    add x10, sp, #64
+; CHECK-NEXT:    lsl x8, x8, #3
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    add x9, sp, #64
+; CHECK-NEXT:    st1d { z0.d }, p0, [x10]
+; CHECK-NEXT:    ldr q2, [x9, x8]
 ; CHECK-NEXT:    stp q0, q2, [x19]
-; CHECK-NEXT:    ldp x20, x19, [sp, #80] // 16-byte Folded Reload
-; CHECK-NEXT:    add sp, sp, #96
+; CHECK-NEXT:    addvl sp, sp, #1
+; CHECK-NEXT:    add sp, sp, #64
+; CHECK-NEXT:    ldp x30, x19, [sp, #16] // 16-byte Folded Reload
+; CHECK-NEXT:    ldr x29, [sp], #32 // 8-byte Folded Reload
 ; CHECK-NEXT:    ret
   %alloc = alloca [8 x double]
   call void @def(ptr %alloc)
   %load = load <8 x double>, ptr %alloc
   %strided.vec = shufflevector <8 x double> %load, <8 x double> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
   store <4 x double> %strided.vec, ptr %st_ptr
   ret void
 }
 
 attributes #0 = { "target-features"="+sve" nounwind}