xref: /llvm-project/llvm/test/CodeGen/X86/vshift-6.ll (revision 53fecef1ec5cb502900f82b07036b28cb0fb9f0f)
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=i686-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X86
3; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X64
4
5; This test makes sure that the compiler does not crash with an
6; assertion failure when trying to fold a vector shift left
7; by immediate count if the type of the input vector is different
8; to the result type.
9;
10; This happens for example when lowering a shift left of a MVT::v16i8 vector.
11; This is custom lowered into the following sequence:
12;     count << 5
13;     A =  VSHLI(MVT::v8i16, r & (char16)15, 4)
14;     B = BITCAST MVT::v16i8, A
15;     VSELECT(r, B, count);
16;     count += count
17;     C = VSHLI(MVT::v8i16, r & (char16)63, 2)
18;     D = BITCAST MVT::v16i8, C
19;     r = VSELECT(r, C, count);
20;     count += count
21;     VSELECT(r, r+r, count);
22;     count = count << 5;
23;
24; Where 'r' is a vector of type MVT::v16i8, and
25; 'count' is the vector shift count.
26
27define <16 x i8> @do_not_crash(ptr, ptr, ptr, i32, i64, i8) {
28; X86-LABEL: do_not_crash:
29; X86:       # %bb.0: # %entry
30; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
31; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
32; X86-NEXT:    movb %al, (%ecx)
33; X86-NEXT:    movd %eax, %xmm1
34; X86-NEXT:    psllq $56, %xmm1
35; X86-NEXT:    pcmpeqd %xmm3, %xmm3
36; X86-NEXT:    psllw $5, %xmm1
37; X86-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
38; X86-NEXT:    pxor %xmm2, %xmm2
39; X86-NEXT:    pxor %xmm0, %xmm0
40; X86-NEXT:    pcmpgtb %xmm1, %xmm0
41; X86-NEXT:    pxor %xmm0, %xmm3
42; X86-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
43; X86-NEXT:    por %xmm3, %xmm0
44; X86-NEXT:    paddb %xmm1, %xmm1
45; X86-NEXT:    pxor %xmm3, %xmm3
46; X86-NEXT:    pcmpgtb %xmm1, %xmm3
47; X86-NEXT:    movdqa %xmm3, %xmm4
48; X86-NEXT:    pandn %xmm0, %xmm4
49; X86-NEXT:    psllw $2, %xmm0
50; X86-NEXT:    pand %xmm3, %xmm0
51; X86-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
52; X86-NEXT:    por %xmm4, %xmm0
53; X86-NEXT:    paddb %xmm1, %xmm1
54; X86-NEXT:    pcmpgtb %xmm1, %xmm2
55; X86-NEXT:    movdqa %xmm2, %xmm1
56; X86-NEXT:    pandn %xmm0, %xmm1
57; X86-NEXT:    paddb %xmm0, %xmm0
58; X86-NEXT:    pand %xmm2, %xmm0
59; X86-NEXT:    por %xmm1, %xmm0
60; X86-NEXT:    retl
61;
62; X64-LABEL: do_not_crash:
63; X64:       # %bb.0: # %entry
64; X64-NEXT:    movb %r9b, (%rdi)
65; X64-NEXT:    movd %r9d, %xmm1
66; X64-NEXT:    psllq $56, %xmm1
67; X64-NEXT:    pcmpeqd %xmm2, %xmm2
68; X64-NEXT:    psllw $5, %xmm1
69; X64-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
70; X64-NEXT:    pxor %xmm3, %xmm3
71; X64-NEXT:    pxor %xmm0, %xmm0
72; X64-NEXT:    pcmpgtb %xmm1, %xmm0
73; X64-NEXT:    pxor %xmm0, %xmm2
74; X64-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
75; X64-NEXT:    por %xmm2, %xmm0
76; X64-NEXT:    paddb %xmm1, %xmm1
77; X64-NEXT:    pxor %xmm2, %xmm2
78; X64-NEXT:    pcmpgtb %xmm1, %xmm2
79; X64-NEXT:    movdqa %xmm2, %xmm4
80; X64-NEXT:    pandn %xmm0, %xmm4
81; X64-NEXT:    psllw $2, %xmm0
82; X64-NEXT:    pand %xmm2, %xmm0
83; X64-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
84; X64-NEXT:    por %xmm4, %xmm0
85; X64-NEXT:    paddb %xmm1, %xmm1
86; X64-NEXT:    pcmpgtb %xmm1, %xmm3
87; X64-NEXT:    movdqa %xmm3, %xmm1
88; X64-NEXT:    pandn %xmm0, %xmm1
89; X64-NEXT:    paddb %xmm0, %xmm0
90; X64-NEXT:    pand %xmm3, %xmm0
91; X64-NEXT:    por %xmm1, %xmm0
92; X64-NEXT:    retq
93entry:
94  store i8 %5, ptr %0
95  %L5 = load i8, ptr %0
96  %I8 = insertelement <16 x i8> <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>, i8 %L5, i32 7
97  %B51 = shl <16 x i8> <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>, %I8
98  ret <16 x i8> %B51
99}
100