6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2024-07-21 18:29:45 +00:00
Code Issues Projects Releases Wiki Activity
99d1c88279
llvm-6502/test/CodeGen/AArch64/arm64-umaxv.ll
Ahmed Bougacha 4a3cd42601 [AArch64] Avoid going through GPRs for across-vector instructions. 
This adds new node types for each intrinsic.
For instance, for addv, we have AArch64ISD::UADDV, such that:
  (v4i32 (uaddv ...))
is the same as
  (v4i32 (scalar_to_vector (i32 (int_aarch64_neon_uaddv ...))))
that is,
  (v4i32 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)),
           (i32 (int_aarch64_neon_uaddv ...)), ssub)

In a combine, we transform all such across-vector-lanes intrinsics to:

  (i32 (extract_vector_elt (uaddv ...), 0))

This has one big advantage: by making the extract_element explicit, we
enable the existing patterns for lane-aware instructions to fire.
This lets us avoid needlessly going through the GPRs.  Consider:

    uint32x4_t test_mul(uint32x4_t a, uint32x4_t b) {
        return vmulq_n_u32(a, vaddvq_u32(b));
    }

We now generate:
    addv.4s  s1, v1
    mul.4s   v0, v0, v1[0]
instead of the previous:
    addv.4s  s1, v1
    fmov     w8, s1
    dup.4s   v1, w8
    mul.4s   v0, v1, v0

rdar://20044838


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231840 91177308-0d34-0410-b5e6-96231b3b80d8
2015-03-10 20:45:38 +00:00

165 lines
5.2 KiB
LLVM
Raw Blame History

; RUN: llc < %s -march=arm64 -aarch64-neon-syntax=apple -asm-verbose=false | FileCheck %s
define i32 @vmax_u8x8(<8 x i8> %a) nounwind ssp {
; CHECK-LABEL: vmax_u8x8:
; CHECK: umaxv.8b b[[REG:[0-9]+]], v0
; CHECK: fmov [[REG2:w[0-9]+]], s[[REG]]
; CHECK-NOT: and
; CHECK: cbz [[REG2]],
entry:
%vmaxv.i = tail call i32 @llvm.aarch64.neon.umaxv.i32.v8i8(<8 x i8> %a) nounwind
%tmp = trunc i32 %vmaxv.i to i8
%tobool = icmp eq i8 %tmp, 0
br i1 %tobool, label %return, label %if.then
if.then:
%call1 = tail call i32 bitcast (i32 (...)* @bar to i32 ()*)() nounwind
br label %return
return:
%retval.0 = phi i32 [ %call1, %if.then ], [ 0, %entry ]
ret i32 %retval.0
}
declare i32 @bar(...)
define i32 @vmax_u4x16(<4 x i16> %a) nounwind ssp {
; CHECK-LABEL: vmax_u4x16:
; CHECK: umaxv.4h h[[REG:[0-9]+]], v0
; CHECK: fmov [[REG2:w[0-9]+]], s[[REG]]
; CHECK-NOT: and
; CHECK: cbz [[REG2]],
entry:
%vmaxv.i = tail call i32 @llvm.aarch64.neon.umaxv.i32.v4i16(<4 x i16> %a) nounwind
%tmp = trunc i32 %vmaxv.i to i16
%tobool = icmp eq i16 %tmp, 0
br i1 %tobool, label %return, label %if.then
if.then:
%call1 = tail call i32 bitcast (i32 (...)* @bar to i32 ()*)() nounwind
br label %return
return:
%retval.0 = phi i32 [ %call1, %if.then ], [ 0, %entry ]
ret i32 %retval.0
}
define i32 @vmax_u8x16(<8 x i16> %a) nounwind ssp {
; CHECK-LABEL: vmax_u8x16:
; CHECK: umaxv.8h h[[REG:[0-9]+]], v0
; CHECK: fmov [[REG2:w[0-9]+]], s[[REG]]
; CHECK-NOT: and
; CHECK: cbz [[REG2]],
entry:
%vmaxv.i = tail call i32 @llvm.aarch64.neon.umaxv.i32.v8i16(<8 x i16> %a) nounwind
%tmp = trunc i32 %vmaxv.i to i16
%tobool = icmp eq i16 %tmp, 0
br i1 %tobool, label %return, label %if.then
if.then:
%call1 = tail call i32 bitcast (i32 (...)* @bar to i32 ()*)() nounwind
br label %return
return:
%retval.0 = phi i32 [ %call1, %if.then ], [ 0, %entry ]
ret i32 %retval.0
}
define i32 @vmax_u16x8(<16 x i8> %a) nounwind ssp {
; CHECK-LABEL: vmax_u16x8:
; CHECK: umaxv.16b b[[REG:[0-9]+]], v0
; CHECK: fmov [[REG2:w[0-9]+]], s[[REG]]
; CHECK-NOT: and
; CHECK: cbz [[REG2]],
entry:
%vmaxv.i = tail call i32 @llvm.aarch64.neon.umaxv.i32.v16i8(<16 x i8> %a) nounwind
%tmp = trunc i32 %vmaxv.i to i8
%tobool = icmp eq i8 %tmp, 0
br i1 %tobool, label %return, label %if.then
if.then:
%call1 = tail call i32 bitcast (i32 (...)* @bar to i32 ()*)() nounwind
br label %return
return:
%retval.0 = phi i32 [ %call1, %if.then ], [ 0, %entry ]
ret i32 %retval.0
}
define <8 x i8> @test_vmaxv_u8_used_by_laneop(<8 x i8> %a1, <8 x i8> %a2) {
; CHECK-LABEL: test_vmaxv_u8_used_by_laneop:
; CHECK: umaxv.8b b[[REGNUM:[0-9]+]], v1
; CHECK-NEXT: ins.b v0[3], v[[REGNUM]][0]
; CHECK-NEXT: ret
entry:
%0 = tail call i32 @llvm.aarch64.neon.umaxv.i32.v8i8(<8 x i8> %a2)
%1 = trunc i32 %0 to i8
%2 = insertelement <8 x i8> %a1, i8 %1, i32 3
ret <8 x i8> %2
}
define <4 x i16> @test_vmaxv_u16_used_by_laneop(<4 x i16> %a1, <4 x i16> %a2) {
; CHECK-LABEL: test_vmaxv_u16_used_by_laneop:
; CHECK: umaxv.4h h[[REGNUM:[0-9]+]], v1
; CHECK-NEXT: ins.h v0[3], v[[REGNUM]][0]
; CHECK-NEXT: ret
entry:
%0 = tail call i32 @llvm.aarch64.neon.umaxv.i32.v4i16(<4 x i16> %a2)
%1 = trunc i32 %0 to i16
%2 = insertelement <4 x i16> %a1, i16 %1, i32 3
ret <4 x i16> %2
}
define <2 x i32> @test_vmaxv_u32_used_by_laneop(<2 x i32> %a1, <2 x i32> %a2) {
; CHECK-LABEL: test_vmaxv_u32_used_by_laneop:
; CHECK: umaxp.2s v[[REGNUM:[0-9]+]], v1, v1
; CHECK-NEXT: ins.s v0[1], v[[REGNUM]][0]
; CHECK-NEXT: ret
entry:
%0 = tail call i32 @llvm.aarch64.neon.umaxv.i32.v2i32(<2 x i32> %a2)
%1 = insertelement <2 x i32> %a1, i32 %0, i32 1
ret <2 x i32> %1
}
define <16 x i8> @test_vmaxvq_u8_used_by_laneop(<16 x i8> %a1, <16 x i8> %a2) {
; CHECK-LABEL: test_vmaxvq_u8_used_by_laneop:
; CHECK: umaxv.16b b[[REGNUM:[0-9]+]], v1
; CHECK-NEXT: ins.b v0[3], v[[REGNUM]][0]
; CHECK-NEXT: ret
entry:
%0 = tail call i32 @llvm.aarch64.neon.umaxv.i32.v16i8(<16 x i8> %a2)
%1 = trunc i32 %0 to i8
%2 = insertelement <16 x i8> %a1, i8 %1, i32 3
ret <16 x i8> %2
}
define <8 x i16> @test_vmaxvq_u16_used_by_laneop(<8 x i16> %a1, <8 x i16> %a2) {
; CHECK-LABEL: test_vmaxvq_u16_used_by_laneop:
; CHECK: umaxv.8h h[[REGNUM:[0-9]+]], v1
; CHECK-NEXT: ins.h v0[3], v[[REGNUM]][0]
; CHECK-NEXT: ret
entry:
%0 = tail call i32 @llvm.aarch64.neon.umaxv.i32.v8i16(<8 x i16> %a2)
%1 = trunc i32 %0 to i16
%2 = insertelement <8 x i16> %a1, i16 %1, i32 3
ret <8 x i16> %2
}
define <4 x i32> @test_vmaxvq_u32_used_by_laneop(<4 x i32> %a1, <4 x i32> %a2) {
; CHECK-LABEL: test_vmaxvq_u32_used_by_laneop:
; CHECK: umaxv.4s s[[REGNUM:[0-9]+]], v1
; CHECK-NEXT: ins.s v0[3], v[[REGNUM]][0]
; CHECK-NEXT: ret
entry:
%0 = tail call i32 @llvm.aarch64.neon.umaxv.i32.v4i32(<4 x i32> %a2)
%1 = insertelement <4 x i32> %a1, i32 %0, i32 3
ret <4 x i32> %1
}
declare i32 @llvm.aarch64.neon.umaxv.i32.v16i8(<16 x i8>) nounwind readnone
declare i32 @llvm.aarch64.neon.umaxv.i32.v8i16(<8 x i16>) nounwind readnone
declare i32 @llvm.aarch64.neon.umaxv.i32.v4i16(<4 x i16>) nounwind readnone
declare i32 @llvm.aarch64.neon.umaxv.i32.v8i8(<8 x i8>) nounwind readnone
declare i32 @llvm.aarch64.neon.umaxv.i32.v2i32(<2 x i32>) nounwind readnone
declare i32 @llvm.aarch64.neon.umaxv.i32.v4i32(<4 x i32>) nounwind readnone
Reference in New Issue View Git Blame Copy Permalink