File: unpack-half.ll

package info (click to toggle)
llvm-toolchain-15 1%3A15.0.6-4
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 1,554,644 kB
  • sloc: cpp: 5,922,452; ansic: 1,012,136; asm: 674,362; python: 191,568; objc: 73,855; f90: 42,327; lisp: 31,913; pascal: 11,973; javascript: 10,144; sh: 9,421; perl: 7,447; ml: 5,527; awk: 3,523; makefile: 2,520; xml: 885; cs: 573; fortran: 567
file content (26 lines) | stat: -rw-r--r-- 955 bytes parent folder | download | duplicates (14) 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
 
; RUN: llc -march=amdgcn -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck %s
; RUN: llc -march=amdgcn -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck %s

; On gfx6 and gfx7, this test shows a bug in SelectionDAG where scalarizing the
; extension of a vector of f16 generates an illegal node that errors later.

; CHECK-LABEL: {{^}}main:
; CHECK: v_cvt_f32_f16

define amdgpu_gs void @main(i32 inreg %arg) local_unnamed_addr #0 {
.entry:
  %tmp = load volatile float, float addrspace(1)* undef
  %tmp1 = bitcast float %tmp to i32
  %im0.i = lshr i32 %tmp1, 16
  %tmp2 = insertelement <2 x i32> undef, i32 %im0.i, i32 1
  %tmp3 = trunc <2 x i32> %tmp2 to <2 x i16>
  %tmp4 = bitcast <2 x i16> %tmp3 to <2 x half>
  %tmp5 = fpext <2 x half> %tmp4 to <2 x float>
  %bc = bitcast <2 x float> %tmp5 to <2 x i32>
  %tmp6 = extractelement <2 x i32> %bc, i32 1
  store volatile i32 %tmp6, i32 addrspace(1)* undef
  ret void
}

attributes #0 = { nounwind }