CellSPU assert C != 0 && "LowerEXTRACT_VECTOR_ELT expecting constant SDNode"

[Quuxplusone]

Bugzilla Link	PR3056
Status	RESOLVED FIXED
Importance	P normal
Reported by	Brian Watt (bwatt@us.ibm.com)
Reported on	2008-11-12 09:59:13 -0800
Last modified on	2010-02-22 12:56:15 -0800
Version	trunk
Hardware	PC Linux
CC	llvm-bugs@lists.llvm.org, scooter.phd@gmail.com
Fixed by commit(s)
Attachments
Blocks
Blocked by
See also

For C-language code:

typedef float float4 __attribute__((ext_vector_type(4)));

float check(int i) {
    float4 a = {1.0f,2.0f,3.0f,4.0f};
    return a[i];
}

Which produces LLVM IR (after optimization):

; ModuleID = 'test-float4'
target datalayout = "E-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-
f32:32:32-f64:64:64-v128:128:128"
target triple = "ppc-unknown-linux-gnu"

define float @check(i32 %i) nounwind readnone {
entry:
    %0 = extractelement <4 x float> < float 1.000000e+00, float 2.000000e+00,
float 3.000000e+00, float 4.000000e+00 >, i32 %i        ; <float> [#uses=1]
    ret float %0
}

Produces the following assertion failure for CellSPU:

llc: SPUISelLowering.cpp:2044: llvm::SDValue
LowerEXTRACT_VECTOR_ELT(llvm::SDValue, llvm::SelectionDAG&): Assertion `C != 0
&& "LowerEXTRACT_VECTOR_ELT expecting constant SDNode"' failed.

[Quuxplusone]

Only constant element extraction is currently implemented in the code.

[Quuxplusone]

Fixed by revision 59891.

[Quuxplusone]

Now producing IR as follows:

; ModuleID = 'test-float4'
target datalayout = "E-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-
f32:32:32-f64:64:64-v128:128:128"
target triple = "ppc-unknown-linux-gnu"

define float @check(i32 %i) nounwind readnone {
entry:
    %vecext = extractelement <4 x float> < float 1.000000e+00, float 2.000000e+00,
float 3.000000e+00, float 4.000000e+00 >, i32 %i        ; <float> [#uses=1]
    ret float %vecext
}

However, the resulting Cell SPU assembler code now produces:

    .text

    .section    .rodata.cst16,"aM",@progbits,16
    .align  4
.LCPI1_0:
    .long   1065353216  # float 1.000000e+00
    .long   1073741824  # float 2.000000e+00
    .long   1077936128  # float 3.000000e+00
    .long   1082130432  # float 4.000000e+00
    .text
    .align  3
    .global check
    .type   check, @function
check:
    lqa $4, .LCPI1_0
    shli    $3, $3, 2
    ila $5, 66051
    shlqbyi $3, $4, $3
    shufb   $3, $3, $3, $5
.Llabel1:
    bi  $lr
    .size   check,.-check

Which when compiled with spu-as, I produces an assembler warning

spu-as -g -o test-float4.o test-float4.s # test-float4.o
test-float4.s: Assembler messages:
test-float4.s:19: Warning: Treating '$3' as a symbol.

It appears to be warning that "shlqbyi $3, $4, $3" should have an immediate
value and not a register in the last operand.

As a result, I am reopening this bug. Sorry.

[Quuxplusone]

Typo in instruction mnemonic. Fixed.