; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=x86_64-linux-gnu                 			                                  -global-isel < %s -o - | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSE_FAST
; RUN: llc -mtriple=x86_64-linux-gnu                                  -regbankselect-greedy -global-isel < %s -o - | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSE_GREEDY
; RUN: llc -mtriple=x86_64-linux-gnu -mattr=+avx                                            -global-isel < %s -o - | FileCheck %s --check-prefix=ALL --check-prefix=ALL_AVX --check-prefix=ALL_AVX_FAST   --check-prefix=AVX_FAST
; RUN: llc -mtriple=x86_64-linux-gnu -mattr=+avx                      -regbankselect-greedy -global-isel < %s -o - | FileCheck %s --check-prefix=ALL --check-prefix=ALL_AVX --check-prefix=ALL_AVX_GREEDY --check-prefix=AVX_GREEDY
; RUN: llc -mtriple=x86_64-linux-gnu -mattr=+avx512f 	                                      -global-isel < %s -o - | FileCheck %s --check-prefix=ALL --check-prefix=ALL_AVX --check-prefix=ALL_AVX_FAST   --check-prefix=AVX512F_FAST
; RUN: llc -mtriple=x86_64-linux-gnu -mattr=+avx512f                  -regbankselect-greedy -global-isel < %s -o - | FileCheck %s --check-prefix=ALL --check-prefix=ALL_AVX --check-prefix=ALL_AVX_GREEDY --check-prefix=AVX512F_GREEDY
; RUN: llc -mtriple=x86_64-linux-gnu -mattr=+avx512f -mattr=+avx512vl			                  -global-isel < %s -o - | FileCheck %s --check-prefix=ALL --check-prefix=ALL_AVX --check-prefix=ALL_AVX_FAST   --check-prefix=AVX512VL_FAST
; RUN: llc -mtriple=x86_64-linux-gnu -mattr=+avx512f -mattr=+avx512vl -regbankselect-greedy -global-isel < %s -o - | FileCheck %s --check-prefix=ALL --check-prefix=ALL_AVX --check-prefix=ALL_AVX_GREEDY --check-prefix=AVX512VL_GREEDY


define i8 @test_load_i8(i8 * %p1) {
; ALL-LABEL: test_load_i8:
; ALL:       # BB#0:
; ALL-NEXT:    movb (%rdi), %al
; ALL-NEXT:    retq
  %r = load i8, i8* %p1
  ret i8 %r
}

define i16 @test_load_i16(i16 * %p1) {
; ALL-LABEL: test_load_i16:
; ALL:       # BB#0:
; ALL-NEXT:    movzwl (%rdi), %eax
; ALL-NEXT:    retq
  %r = load i16, i16* %p1
  ret i16 %r
}

define i32 @test_load_i32(i32 * %p1) {
; ALL-LABEL: test_load_i32:
; ALL:       # BB#0:
; ALL-NEXT:    movl (%rdi), %eax
; ALL-NEXT:    retq
  %r = load i32, i32* %p1
  ret i32 %r
}

define i64 @test_load_i64(i64 * %p1) {
; ALL-LABEL: test_load_i64:
; ALL:       # BB#0:
; ALL-NEXT:    movq (%rdi), %rax
; ALL-NEXT:    retq
  %r = load i64, i64* %p1
  ret i64 %r
}

define float @test_load_float(float * %p1) {
; SSE-LABEL: test_load_float:
; SSE:       # BB#0:
; SSE-NEXT:    movl (%rdi), %eax
; SSE-NEXT:    movd %eax, %xmm0
; SSE-NEXT:    retq
;
; ALL_AVX-LABEL: test_load_float:
; ALL_AVX:       # BB#0:
; ALL_AVX-NEXT:    movl (%rdi), %eax
; ALL_AVX-NEXT:    vmovd %eax, %xmm0
; ALL_AVX-NEXT:    retq
  %r = load float, float* %p1
  ret float %r
}

define double @test_load_double(double * %p1) {
; SSE-LABEL: test_load_double:
; SSE:       # BB#0:
; SSE-NEXT:    movq (%rdi), %rax
; SSE-NEXT:    movd %rax, %xmm0
; SSE-NEXT:    retq
;
; ALL_AVX-LABEL: test_load_double:
; ALL_AVX:       # BB#0:
; ALL_AVX-NEXT:    movq (%rdi), %rax
; ALL_AVX-NEXT:    vmovq %rax, %xmm0
; ALL_AVX-NEXT:    retq
  %r = load double, double* %p1
  ret double %r
}

define <4 x i32> @test_load_v4i32_noalign(<4 x i32> * %p1) {
; SSE-LABEL: test_load_v4i32_noalign:
; SSE:       # BB#0:
; SSE-NEXT:    movups (%rdi), %xmm0
; SSE-NEXT:    retq
;
; ALL_AVX-LABEL: test_load_v4i32_noalign:
; ALL_AVX:       # BB#0:
; ALL_AVX-NEXT:    vmovups (%rdi), %xmm0
; ALL_AVX-NEXT:    retq
  %r = load <4 x i32>, <4 x i32>* %p1, align 1
  ret <4 x i32> %r
}

define <4 x i32> @test_load_v4i32_align(<4 x i32> * %p1) {
; SSE-LABEL: test_load_v4i32_align:
; SSE:       # BB#0:
; SSE-NEXT:    movaps (%rdi), %xmm0
; SSE-NEXT:    retq
;
; ALL_AVX-LABEL: test_load_v4i32_align:
; ALL_AVX:       # BB#0:
; ALL_AVX-NEXT:    vmovaps (%rdi), %xmm0
; ALL_AVX-NEXT:    retq
  %r = load <4 x i32>, <4 x i32>* %p1, align 16
  ret <4 x i32> %r
}

define i32 * @test_store_i32(i32 %val, i32 * %p1) {
; ALL-LABEL: test_store_i32:
; ALL:       # BB#0:
; ALL-NEXT:    movl %edi, (%rsi)
; ALL-NEXT:    movq %rsi, %rax
; ALL-NEXT:    retq
  store i32 %val, i32* %p1
  ret i32 * %p1;
}

define i64 * @test_store_i64(i64 %val, i64 * %p1) {
; ALL-LABEL: test_store_i64:
; ALL:       # BB#0:
; ALL-NEXT:    movq %rdi, (%rsi)
; ALL-NEXT:    movq %rsi, %rax
; ALL-NEXT:    retq
  store i64 %val, i64* %p1
  ret i64 * %p1;
}

define float * @test_store_float(float %val, float * %p1) {
;
; SSE_FAST-LABEL: test_store_float:
; SSE_FAST:       # BB#0:
; SSE_FAST-NEXT:    movd %xmm0, %eax
; SSE_FAST-NEXT:    movl %eax, (%rdi)
; SSE_FAST-NEXT:    movq %rdi, %rax
; SSE_FAST-NEXT:    retq
;
; SSE_GREEDY-LABEL: test_store_float:
; SSE_GREEDY:       # BB#0:
; SSE_GREEDY-NEXT:    movss %xmm0, (%rdi)
; SSE_GREEDY-NEXT:    movq %rdi, %rax
; SSE_GREEDY-NEXT:    retq
;
; ALL_AVX_FAST-LABEL: test_store_float:
; ALL_AVX_FAST:       # BB#0:
; ALL_AVX_FAST-NEXT:    vmovd %xmm0, %eax
; ALL_AVX_FAST-NEXT:    movl %eax, (%rdi)
; ALL_AVX_FAST-NEXT:    movq %rdi, %rax
; ALL_AVX_FAST-NEXT:    retq
;
; ALL_AVX_GREEDY-LABEL: test_store_float:
; ALL_AVX_GREEDY:       # BB#0:
; ALL_AVX_GREEDY-NEXT:    vmovss %xmm0, (%rdi)
; ALL_AVX_GREEDY-NEXT:    movq %rdi, %rax
; ALL_AVX_GREEDY-NEXT:    retq
  store float %val, float* %p1
  ret float * %p1;
}

define double * @test_store_double(double %val, double * %p1) {
;
; SSE_FAST-LABEL: test_store_double:
; SSE_FAST:       # BB#0:
; SSE_FAST-NEXT:    movd %xmm0, %rax
; SSE_FAST-NEXT:    movq %rax, (%rdi)
; SSE_FAST-NEXT:    movq %rdi, %rax
; SSE_FAST-NEXT:    retq
;
; SSE_GREEDY-LABEL: test_store_double:
; SSE_GREEDY:       # BB#0:
; SSE_GREEDY-NEXT:    movsd %xmm0, (%rdi)
; SSE_GREEDY-NEXT:    movq %rdi, %rax
; SSE_GREEDY-NEXT:    retq
;
; ALL_AVX_FAST-LABEL: test_store_double:
; ALL_AVX_FAST:       # BB#0:
; ALL_AVX_FAST-NEXT:    vmovq %xmm0, %rax
; ALL_AVX_FAST-NEXT:    movq %rax, (%rdi)
; ALL_AVX_FAST-NEXT:    movq %rdi, %rax
; ALL_AVX_FAST-NEXT:    retq
;
; ALL_AVX_GREEDY-LABEL: test_store_double:
; ALL_AVX_GREEDY:       # BB#0:
; ALL_AVX_GREEDY-NEXT:    vmovsd %xmm0, (%rdi)
; ALL_AVX_GREEDY-NEXT:    movq %rdi, %rax
; ALL_AVX_GREEDY-NEXT:    retq
  store double %val, double* %p1
  ret double * %p1;
}