https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127148
--- Comment #1 from Jeevitha <jeevitha at gcc dot gnu.org> ---
vec_extractm:
Purpose: Collect the high-order bits of each element of the input vector.
Result value: The contents of the high-order bit of each element of a are
concatenated and placed
in the rightmost bits of r, with other bits of r set to 0.
Consider for char:
For a vector unsigned char, there are 16 elements, so vec_extractm collects the
low-order bit (bit[0]) from each of the 16 bytes.
The vec_extractm intrinsic is mapped to the corresponding instruction for char
elements:
vextractbm RT,VRB
The instruction operates as follows:
do i = 0 to 15
GPR[RT].bit[48+i] ← VSR[VRB+32].byte[i].bit[0]
end
GPR[RT].bit[0:47] ← 0
For this testcase:
vbc_bi_src[0] = 0xFF;
vbc_bi_src[1] = 0xFF;
vbc_bi_src[2] = 0x0;
vbc_bi_src[3] = 0x0;
vbc_bi_src[4] = 0x0;
vbc_bi_src[5] = 0x0;
vbc_bi_src[6] = 0xFF;
vbc_bi_src[7] = 0xFF;
vbc_bi_src[8] = 0xFF;
vbc_bi_src[9] = 0xFF;
vbc_bi_src[10] = 0xFF;
vbc_bi_src[11] = 0xFF;
vbc_bi_src[12] = 0xFF;
vbc_bi_src[13] = 0x0;
vbc_bi_src[14] = 0xFF;
vbc_bi_src[15] = 0xFF;
expected_result_wi = 0b1101111111000011;
result_wi = vec_extractm (vbc_bi_src);
----------------------------------------
Instruction generated: vextractbm r2,v0
So vector vbc_bi_src is expressed like:
vbc_bi_src -> v0
LE:
{uint128 = 0xffff00ffffffffffffff00000000ffff, v4_float = {0xffff, 0xffff0000,
0xffffffff, 0xffff00ff},
v4_int32 = {0xffff, 0xffff0000, 0xffffffff, 0xffff00ff}, v8_int16 = {0xffff,
0x0, 0x0, 0xffff, 0xffff, 0xffff, 0xff,
0xffff}, v16_int8 = {0xff, 0xff, 0x0, 0x0, 0x0, 0x0, 0xff, 0xff, 0xff,
0xff, 0xff, 0xff, 0xff, 0x0, 0xff, 0xff}}
BE:
{uint128 = 0xffff00000000ffffffffffffff00ffff, v4_float = {0xffff0000, 0xffff,
0xffffffff, 0xff00ffff}, v4_int32 = {
0xffff0000, 0xffff, 0xffffffff, 0xff00ffff}, v8_int16 = {0xffff, 0x0, 0x0,
0xffff, 0xffff, 0xffff, 0xff00, 0xffff},
v16_int8 = {0xff, 0xff, 0x0, 0x0, 0x0, 0x0, 0xff, 0xff, 0xff, 0xff, 0xff,
0xff, 0xff, 0x0, 0xff, 0xff}}
The difference is in how the 128-bit vector is laid out in the hex
representation.
On LE:
uint128 = 0xffff00ffffffffffffff00000000ffff
The low-order bit of each byte is:
1 1 0 1 1 1 1 1 1 1 0 0 0 0 1 1
On BE:
uint128 = 0xffff00000000ffffffffffffff00ffff
The low-order bit of each byte is:
1 1 0 0 0 0 1 1 1 1 1 1 1 0 1 1
The testcase should therefore account for the target endianness when checking
the expected result.
