llvm/test/CodeGen/AArch64/ldexp.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
; RUN: llc -mtriple=aarch64 -mattr=+sve < %s -o - | FileCheck -check-prefixes=SVE,SVELINUX %s
; RUN: llc -mtriple=aarch64 -global-isel -global-isel-abort=2 < %s -o - 2>&1 | FileCheck -check-prefixes=GISEL %s
; RUN: llc -mtriple=aarch64-windows-msvc -mattr=+sve < %s -o - | FileCheck -check-prefixes=SVE,SVEWINDOWS %s
; RUN: llc -mtriple=aarch64-windows-msvc < %s -o - | FileCheck -check-prefixes=WINDOWS %s

; GISEL:       warning: Instruction selection used fallback path for testExpIntrinsic_i16

define double @testExp(double %val, i32 %a) {
; SVE-LABEL: testExp:
; SVE:       // %bb.0: // %entry
; SVE-NEXT:    // kill: def $w0 killed $w0 def $x0
; SVE-NEXT:    sxtw x8, w0
; SVE-NEXT:    ptrue p0.d
; SVE-NEXT:    // kill: def $d0 killed $d0 def $z0
; SVE-NEXT:    fmov d1, x8
; SVE-NEXT:    fscale z0.d, p0/m, z0.d, z1.d
; SVE-NEXT:    // kill: def $d0 killed $d0 killed $z0
; SVE-NEXT:    ret
;
; GISEL-LABEL: testExp:
; GISEL:       // %bb.0: // %entry
; GISEL-NEXT:    b ldexp
;
; WINDOWS-LABEL: testExp:
; WINDOWS:       // %bb.0: // %entry
; WINDOWS-NEXT:    b ldexp
entry:
  %call = tail call fast double @ldexp(double %val, i32 %a)
  ret double %call
}

declare double @ldexp(double, i32) memory(none)

define double @testExpIntrinsic(double %val, i32 %a) {
; SVE-LABEL: testExpIntrinsic:
; SVE:       // %bb.0: // %entry
; SVE-NEXT:    // kill: def $w0 killed $w0 def $x0
; SVE-NEXT:    sxtw x8, w0
; SVE-NEXT:    ptrue p0.d
; SVE-NEXT:    // kill: def $d0 killed $d0 def $z0
; SVE-NEXT:    fmov d1, x8
; SVE-NEXT:    fscale z0.d, p0/m, z0.d, z1.d
; SVE-NEXT:    // kill: def $d0 killed $d0 killed $z0
; SVE-NEXT:    ret
;
; GISEL-LABEL: testExpIntrinsic:
; GISEL:       // %bb.0: // %entry
; GISEL-NEXT:    b ldexp
;
; WINDOWS-LABEL: testExpIntrinsic:
; WINDOWS:       // %bb.0: // %entry
; WINDOWS-NEXT:    b ldexp
entry:
  %call = tail call fast double @llvm.ldexp.f64(double %val, i32 %a)
  ret double %call
}

; A sub-int exponent must be sign-extended to int on the libcall path;
; PromoteIntOp_ExpOp must not crash on it.
define double @testExpIntrinsic_i16(double %val, i16 %a) nounwind {
; SVE-LABEL: testExpIntrinsic_i16:
; SVE:       // %bb.0: // %entry
; SVE-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
; SVE-NEXT:    sxth w0, w0
; SVE-NEXT:    bl ldexp
; SVE-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
; SVE-NEXT:    ret
;
; GISEL-LABEL: testExpIntrinsic_i16:
; GISEL:       // %bb.0: // %entry
; GISEL-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
; GISEL-NEXT:    sxth w0, w0
; GISEL-NEXT:    bl ldexp
; GISEL-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
; GISEL-NEXT:    ret
;
; WINDOWS-LABEL: testExpIntrinsic_i16:
; WINDOWS:       // %bb.0: // %entry
; WINDOWS-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
; WINDOWS-NEXT:    sxth w0, w0
; WINDOWS-NEXT:    bl ldexp
; WINDOWS-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
; WINDOWS-NEXT:    ret
entry:
  %call = tail call fast double @llvm.ldexp.f64.i16(double %val, i16 %a)
  ret double %call
}

define float @testExpf(float %val, i32 %a) {
; SVELINUX-LABEL: testExpf:
; SVELINUX:       // %bb.0: // %entry
; SVELINUX-NEXT:    fmov s1, w0
; SVELINUX-NEXT:    ptrue p0.s
; SVELINUX-NEXT:    // kill: def $s0 killed $s0 def $z0
; SVELINUX-NEXT:    fscale z0.s, p0/m, z0.s, z1.s
; SVELINUX-NEXT:    // kill: def $s0 killed $s0 killed $z0
; SVELINUX-NEXT:    ret
;
; GISEL-LABEL: testExpf:
; GISEL:       // %bb.0: // %entry
; GISEL-NEXT:    b ldexpf
;
; SVEWINDOWS-LABEL: testExpf:
; SVEWINDOWS:       // %bb.0: // %entry
; SVEWINDOWS-NEXT:    b ldexpf
;
; WINDOWS-LABEL: testExpf:
; WINDOWS:       // %bb.0: // %entry
; WINDOWS-NEXT:    b ldexpf
entry:
  %call = tail call fast float @ldexpf(float %val, i32 %a)
  ret float %call
}

define float @testExpfIntrinsic(float %val, i32 %a) {
; SVE-LABEL: testExpfIntrinsic:
; SVE:       // %bb.0: // %entry
; SVE-NEXT:    fmov s1, w0
; SVE-NEXT:    ptrue p0.s
; SVE-NEXT:    // kill: def $s0 killed $s0 def $z0
; SVE-NEXT:    fscale z0.s, p0/m, z0.s, z1.s
; SVE-NEXT:    // kill: def $s0 killed $s0 killed $z0
; SVE-NEXT:    ret
;
; GISEL-LABEL: testExpfIntrinsic:
; GISEL:       // %bb.0: // %entry
; GISEL-NEXT:    b ldexpf
;
; WINDOWS-LABEL: testExpfIntrinsic:
; WINDOWS:       .seh_proc testExpfIntrinsic
; WINDOWS-NEXT:  // %bb.0: // %entry
; WINDOWS-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
; WINDOWS-NEXT:    .seh_save_reg_x x30, 16
; WINDOWS-NEXT:    .seh_endprologue
; WINDOWS-NEXT:    fcvt d0, s0
; WINDOWS-NEXT:    bl ldexp
; WINDOWS-NEXT:    fcvt s0, d0
; WINDOWS-NEXT:    .seh_startepilogue
; WINDOWS-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
; WINDOWS-NEXT:    .seh_save_reg_x x30, 16
; WINDOWS-NEXT:    .seh_endepilogue
; WINDOWS-NEXT:    ret
; WINDOWS-NEXT:    .seh_endfunclet
; WINDOWS-NEXT:    .seh_endproc
entry:
  %call = tail call fast float @llvm.ldexp.f32(float %val, i32 %a)
  ret float %call
}

define <2 x float> @test_ldexp_v2f32_v2i32(<2 x float> %Val, <2 x i32> %Exp) {
; SVE-LABEL: test_ldexp_v2f32_v2i32:
; SVE:       // %bb.0:
; SVE-NEXT:    // kill: def $d1 killed $d1 def $z1
; SVE-NEXT:    mov w8, v1.s[1]
; SVE-NEXT:    // kill: def $d0 killed $d0 def $z0
; SVE-NEXT:    mov s2, v0.s[1]
; SVE-NEXT:    ptrue p0.s
; SVE-NEXT:    fscale z0.s, p0/m, z0.s, z1.s
; SVE-NEXT:    fmov s3, w8
; SVE-NEXT:    fscale z2.s, p0/m, z2.s, z3.s
; SVE-NEXT:    mov v0.s[1], v2.s[0]
; SVE-NEXT:    // kill: def $d0 killed $d0 killed $z0
; SVE-NEXT:    ret
;
; GISEL-LABEL: test_ldexp_v2f32_v2i32:
; GISEL:       // %bb.0:
; GISEL-NEXT:    sub sp, sp, #48
; GISEL-NEXT:    stp d9, d8, [sp, #16] // 16-byte Folded Spill
; GISEL-NEXT:    str x30, [sp, #32] // 8-byte Spill
; GISEL-NEXT:    .cfi_def_cfa_offset 48
; GISEL-NEXT:    .cfi_offset w30, -16
; GISEL-NEXT:    .cfi_offset b8, -24
; GISEL-NEXT:    .cfi_offset b9, -32
; GISEL-NEXT:    // kill: def $d1 killed $d1 def $q1
; GISEL-NEXT:    fmov w0, s1
; GISEL-NEXT:    // kill: def $d0 killed $d0 def $q0
; GISEL-NEXT:    mov s8, v0.s[1]
; GISEL-NEXT:    mov s9, v1.s[1]
; GISEL-NEXT:    // kill: def $s0 killed $s0 killed $q0
; GISEL-NEXT:    bl ldexpf
; GISEL-NEXT:    // kill: def $s0 killed $s0 def $q0
; GISEL-NEXT:    str q0, [sp] // 16-byte Spill
; GISEL-NEXT:    fmov w0, s9
; GISEL-NEXT:    fmov s0, s8
; GISEL-NEXT:    bl ldexpf
; GISEL-NEXT:    ldr q1, [sp] // 16-byte Reload
; GISEL-NEXT:    // kill: def $s0 killed $s0 def $q0
; GISEL-NEXT:    ldp d9, d8, [sp, #16] // 16-byte Folded Reload
; GISEL-NEXT:    ldr x30, [sp, #32] // 8-byte Reload
; GISEL-NEXT:    mov v1.s[1], v0.s[0]
; GISEL-NEXT:    fmov d0, d1
; GISEL-NEXT:    add sp, sp, #48
; GISEL-NEXT:    ret
;
; WINDOWS-LABEL: test_ldexp_v2f32_v2i32:
; WINDOWS:       .seh_proc test_ldexp_v2f32_v2i32
; WINDOWS-NEXT:  // %bb.0:
; WINDOWS-NEXT:    sub sp, sp, #48
; WINDOWS-NEXT:    .seh_stackalloc 48
; WINDOWS-NEXT:    str x30, [sp, #32] // 8-byte Spill
; WINDOWS-NEXT:    .seh_save_reg x30, 32
; WINDOWS-NEXT:    .seh_endprologue
; WINDOWS-NEXT:    // kill: def $d0 killed $d0 def $q0
; WINDOWS-NEXT:    mov s2, v0.s[1]
; WINDOWS-NEXT:    // kill: def $d1 killed $d1 def $q1
; WINDOWS-NEXT:    stp q1, q0, [sp] // 32-byte Folded Spill
; WINDOWS-NEXT:    mov w0, v1.s[1]
; WINDOWS-NEXT:    fcvt d0, s2
; WINDOWS-NEXT:    bl ldexp
; WINDOWS-NEXT:    ldr q1, [sp, #16] // 16-byte Reload
; WINDOWS-NEXT:    fcvt s0, d0
; WINDOWS-NEXT:    fcvt d1, s1
; WINDOWS-NEXT:    str q0, [sp, #16] // 16-byte Spill
; WINDOWS-NEXT:    ldr q0, [sp] // 16-byte Reload
; WINDOWS-NEXT:    fmov w0, s0
; WINDOWS-NEXT:    fmov d0, d1
; WINDOWS-NEXT:    bl ldexp
; WINDOWS-NEXT:    fcvt s0, d0
; WINDOWS-NEXT:    ldr q1, [sp, #16] // 16-byte Reload
; WINDOWS-NEXT:    mov v0.s[1], v1.s[0]
; WINDOWS-NEXT:    // kill: def $d0 killed $d0 killed $q0
; WINDOWS-NEXT:    .seh_startepilogue
; WINDOWS-NEXT:    ldr x30, [sp, #32] // 8-byte Reload
; WINDOWS-NEXT:    .seh_save_reg x30, 32
; WINDOWS-NEXT:    add sp, sp, #48
; WINDOWS-NEXT:    .seh_stackalloc 48
; WINDOWS-NEXT:    .seh_endepilogue
; WINDOWS-NEXT:    ret
; WINDOWS-NEXT:    .seh_endfunclet
; WINDOWS-NEXT:    .seh_endproc
  %result = call <2 x float> @llvm.ldexp.v2f32.v2i32(<2 x float> %Val, <2 x i32> %Exp)
  ret <2 x float> %result
}

declare float @ldexpf(float, i32) memory(none)

define fp128 @testExpl(fp128 %val, i32 %a) {
; SVE-LABEL: testExpl:
; SVE:       // %bb.0: // %entry
; SVE-NEXT:    b ldexpl
;
; GISEL-LABEL: testExpl:
; GISEL:       // %bb.0: // %entry
; GISEL-NEXT:    b ldexpl
;
; WINDOWS-LABEL: testExpl:
; WINDOWS:       // %bb.0: // %entry
; WINDOWS-NEXT:    b ldexpl
entry:
  %call = tail call fast fp128 @ldexpl(fp128 %val, i32 %a)
  ret fp128 %call
}

declare fp128 @ldexpl(fp128, i32) memory(none)

define half @testExpf16(half %val, i32 %a) {
; SVE-LABEL: testExpf16:
; SVE:       // %bb.0: // %entry
; SVE-NEXT:    fcvt s0, h0
; SVE-NEXT:    fmov s1, w0
; SVE-NEXT:    ptrue p0.s
; SVE-NEXT:    fscale z0.s, p0/m, z0.s, z1.s
; SVE-NEXT:    fcvt h0, s0
; SVE-NEXT:    ret
;
; GISEL-LABEL: testExpf16:
; GISEL:       // %bb.0: // %entry
; GISEL-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
; GISEL-NEXT:    .cfi_def_cfa_offset 16
; GISEL-NEXT:    .cfi_offset w30, -16
; GISEL-NEXT:    fcvt s0, h0
; GISEL-NEXT:    bl ldexpf
; GISEL-NEXT:    fcvt h0, s0
; GISEL-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
; GISEL-NEXT:    ret
;
; WINDOWS-LABEL: testExpf16:
; WINDOWS:       .seh_proc testExpf16
; WINDOWS-NEXT:  // %bb.0: // %entry
; WINDOWS-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
; WINDOWS-NEXT:    .seh_save_reg_x x30, 16
; WINDOWS-NEXT:    .seh_endprologue
; WINDOWS-NEXT:    fcvt d0, h0
; WINDOWS-NEXT:    bl ldexp
; WINDOWS-NEXT:    fcvt h0, d0
; WINDOWS-NEXT:    .seh_startepilogue
; WINDOWS-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
; WINDOWS-NEXT:    .seh_save_reg_x x30, 16
; WINDOWS-NEXT:    .seh_endepilogue
; WINDOWS-NEXT:    ret
; WINDOWS-NEXT:    .seh_endfunclet
; WINDOWS-NEXT:    .seh_endproc
entry:
  %0 = tail call fast half @llvm.ldexp.f16.i32(half %val, i32 %a)
  ret half %0
}

define bfloat @testExpbf16(bfloat %val, i32 %a) {
; SVE-LABEL: testExpbf16:
; SVE:       // %bb.0: // %entry
; SVE-NEXT:    // kill: def $h0 killed $h0 def $d0
; SVE-NEXT:    fmov s1, w0
; SVE-NEXT:    ptrue p0.s
; SVE-NEXT:    shll v0.4s, v0.4h, #16
; SVE-NEXT:    fscale z0.s, p0/m, z0.s, z1.s
; SVE-NEXT:    fmov w8, s0
; SVE-NEXT:    fcmp s0, s0
; SVE-NEXT:    orr w9, w8, #0x400000
; SVE-NEXT:    csel w8, w9, w8, vs
; SVE-NEXT:    lsr w8, w8, #16
; SVE-NEXT:    fmov s0, w8
; SVE-NEXT:    // kill: def $h0 killed $h0 killed $s0
; SVE-NEXT:    ret
;
; GISEL-LABEL: testExpbf16:
; GISEL:       // %bb.0: // %entry
; GISEL-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
; GISEL-NEXT:    .cfi_def_cfa_offset 16
; GISEL-NEXT:    .cfi_offset w30, -16
; GISEL-NEXT:    // kill: def $h0 killed $h0 def $d0
; GISEL-NEXT:    shll v0.4s, v0.4h, #16
; GISEL-NEXT:    // kill: def $s0 killed $s0 killed $q0
; GISEL-NEXT:    bl ldexpf
; GISEL-NEXT:    fmov w9, s0
; GISEL-NEXT:    mov w8, #32767 // =0x7fff
; GISEL-NEXT:    ubfx w10, w9, #16, #1
; GISEL-NEXT:    add w8, w9, w8
; GISEL-NEXT:    add w8, w8, w10
; GISEL-NEXT:    lsr w8, w8, #16
; GISEL-NEXT:    fmov s0, w8
; GISEL-NEXT:    // kill: def $h0 killed $h0 killed $s0
; GISEL-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
; GISEL-NEXT:    ret
;
; WINDOWS-LABEL: testExpbf16:
; WINDOWS:       .seh_proc testExpbf16
; WINDOWS-NEXT:  // %bb.0: // %entry
; WINDOWS-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
; WINDOWS-NEXT:    .seh_save_reg_x x30, 16
; WINDOWS-NEXT:    .seh_endprologue
; WINDOWS-NEXT:    // kill: def $h0 killed $h0 def $d0
; WINDOWS-NEXT:    shll v0.4s, v0.4h, #16
; WINDOWS-NEXT:    fcvt d0, s0
; WINDOWS-NEXT:    bl ldexp
; WINDOWS-NEXT:    fcvtxn s0, d0
; WINDOWS-NEXT:    mov w8, #32767 // =0x7fff
; WINDOWS-NEXT:    fmov w9, s0
; WINDOWS-NEXT:    ubfx w10, w9, #16, #1
; WINDOWS-NEXT:    add w8, w9, w8
; WINDOWS-NEXT:    add w8, w10, w8
; WINDOWS-NEXT:    lsr w8, w8, #16
; WINDOWS-NEXT:    fmov s0, w8
; WINDOWS-NEXT:    // kill: def $h0 killed $h0 killed $s0
; WINDOWS-NEXT:    .seh_startepilogue
; WINDOWS-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
; WINDOWS-NEXT:    .seh_save_reg_x x30, 16
; WINDOWS-NEXT:    .seh_endepilogue
; WINDOWS-NEXT:    ret
; WINDOWS-NEXT:    .seh_endfunclet
; WINDOWS-NEXT:    .seh_endproc
entry:
  %0 = tail call fast bfloat @llvm.ldexp.bf16.i32(bfloat %val, i32 %a)
  ret bfloat %0
}

; The v1f16 result is scalarized but the v1i32 exponent is widened, so the
; exponent was never scalarized.
define <1 x half> @test_ldexp_v1f16_v1i32(<1 x half> %val, <1 x i32> %exp) nounwind {
; SVE-LABEL: test_ldexp_v1f16_v1i32:
; SVE:       // %bb.0:
; SVE-NEXT:    fcvt s0, h0
; SVE-NEXT:    ptrue p0.s
; SVE-NEXT:    // kill: def $d1 killed $d1 def $z1
; SVE-NEXT:    fscale z0.s, p0/m, z0.s, z1.s
; SVE-NEXT:    fcvt h0, s0
; SVE-NEXT:    ret
;
; GISEL-LABEL: test_ldexp_v1f16_v1i32:
; GISEL:       // %bb.0:
; GISEL-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
; GISEL-NEXT:    fcvt s0, h0
; GISEL-NEXT:    fmov w0, s1
; GISEL-NEXT:    bl ldexpf
; GISEL-NEXT:    fcvt h0, s0
; GISEL-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
; GISEL-NEXT:    ret
;
; WINDOWS-LABEL: test_ldexp_v1f16_v1i32:
; WINDOWS:       // %bb.0:
; WINDOWS-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
; WINDOWS-NEXT:    fcvt d0, h0
; WINDOWS-NEXT:    // kill: def $d1 killed $d1 def $q1
; WINDOWS-NEXT:    fmov w0, s1
; WINDOWS-NEXT:    bl ldexp
; WINDOWS-NEXT:    fcvt h0, d0
; WINDOWS-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
; WINDOWS-NEXT:    ret
  %result = call <1 x half> @llvm.ldexp.v1f16.v1i32(<1 x half> %val, <1 x i32> %exp)
  ret <1 x half> %result
}