• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s --check-prefix=EG --check-prefix=FUNC
2; RUN: llc < %s -march=r600 -mcpu=cayman | FileCheck %s --check-prefix=EG --check-prefix=FUNC
3; RUN: llc < %s -march=amdgcn -mcpu=SI -verify-machineinstrs | FileCheck %s --check-prefix=SI --check-prefix=FUNC
4; RUN: llc < %s -march=amdgcn -mcpu=tonga -verify-machineinstrs | FileCheck %s --check-prefix=SI --check-prefix=FUNC
5
6; FUNC-LABEL: {{^}}u32_mad24:
7; EG: MULADD_UINT24
8; SI: v_mad_u32_u24
9
10define void @u32_mad24(i32 addrspace(1)* %out, i32 %a, i32 %b, i32 %c) {
11entry:
12  %0 = shl i32 %a, 8
13  %a_24 = lshr i32 %0, 8
14  %1 = shl i32 %b, 8
15  %b_24 = lshr i32 %1, 8
16  %2 = mul i32 %a_24, %b_24
17  %3 = add i32 %2, %c
18  store i32 %3, i32 addrspace(1)* %out
19  ret void
20}
21
22; FUNC-LABEL: {{^}}i16_mad24:
23; The order of A and B does not matter.
24; EG: MULADD_UINT24 {{[* ]*}}T{{[0-9]}}.[[MAD_CHAN:[XYZW]]]
25; The result must be sign-extended
26; EG: BFE_INT {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[MAD_CHAN]], 0.0, literal.x
27; EG: 16
28; SI: v_mad_u32_u24 [[MAD:v[0-9]]], {{[sv][0-9], [sv][0-9]}}
29; SI: v_bfe_i32 v{{[0-9]}}, [[MAD]], 0, 16
30
31define void @i16_mad24(i32 addrspace(1)* %out, i16 %a, i16 %b, i16 %c) {
32entry:
33  %0 = mul i16 %a, %b
34  %1 = add i16 %0, %c
35  %2 = sext i16 %1 to i32
36  store i32 %2, i32 addrspace(1)* %out
37  ret void
38}
39
40; FUNC-LABEL: {{^}}i8_mad24:
41; EG: MULADD_UINT24 {{[* ]*}}T{{[0-9]}}.[[MAD_CHAN:[XYZW]]]
42; The result must be sign-extended
43; EG: BFE_INT {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[MAD_CHAN]], 0.0, literal.x
44; EG: 8
45; SI: v_mad_u32_u24 [[MUL:v[0-9]]], {{[sv][0-9], [sv][0-9]}}
46; SI: v_bfe_i32 v{{[0-9]}}, [[MUL]], 0, 8
47
48define void @i8_mad24(i32 addrspace(1)* %out, i8 %a, i8 %b, i8 %c) {
49entry:
50  %0 = mul i8 %a, %b
51  %1 = add i8 %0, %c
52  %2 = sext i8 %1 to i32
53  store i32 %2, i32 addrspace(1)* %out
54  ret void
55}
56
57; This tests for a bug where the mad_u24 pattern matcher would call
58; SimplifyDemandedBits on the first operand of the mul instruction
59; assuming that the pattern would be matched to a 24-bit mad.  This
60; led to some instructions being incorrectly erased when the entire
61; 24-bit mad pattern wasn't being matched.
62
63; Check that the select instruction is not deleted.
64; FUNC-LABEL: {{^}}i24_i32_i32_mad:
65; EG: CNDE_INT
66; SI: v_cndmask
67define void @i24_i32_i32_mad(i32 addrspace(1)* %out, i32 %a, i32 %b, i32 %c, i32 %d) {
68entry:
69  %0 = ashr i32 %a, 8
70  %1 = icmp ne i32 %c, 0
71  %2 = select i1 %1, i32 %0, i32 34
72  %3 = mul i32 %2, %c
73  %4 = add i32 %3, %d
74  store i32 %4, i32 addrspace(1)* %out
75  ret void
76}
77