The conversion expansions did not properly handle bfloat types.
I'm not certain that these expansions are completely correct;
I don't have any experience with AMDGPU or the ability to run
anything to test it.
Note that it doesn't seem like AMDGPU with GlobalISel can
handle fptrunc of float to bfloat, which is needed for itofp.
I've omitted the GISEL run for the bfloat case.
This fixes#85379.
expand-large-fp-convert cannot handle vector types.
If overly large vector element types survive into
isel, they will likely be scalarized there, but since
isel cannot handle scalar integer types of that size,
it will assert.
Handle vector types in expand-large-fp-convert by
scalarizing them and then expanding the scalar type
operation. For large vectors, this results in a
*massive* code expansion, but it's better than
asserting.
The IR for a double-to-i129 conversion looks like this in one of the
blocks in compiler-rt:
%cmp5.i = icmp ult i16 %3, -129, !dbg !24
But in ExpandLargeFpConvert, it looks like:
%13 = icmp ult i129 %12, 4294967167, !dbg !19
ExpandLargeFpConvert is wrong; the value should have been
signed before negating, but instead we get a very large
unsigned value. Another value in the same pass also has this
issue.
When deciding whether to perform rounding on the significand,
the generated IR was using (width - leading zeros - 1) rather
than (width - leading zeros). This is different from how the
routine in compiler-rt does it:
int sd = srcBits - clzSrcT(a);
int e = sd - 1;
if (sd > dstMantDig) {
This bug means that the following code, when built on -O0:
#include <stdio.h>
_BitInt(233) v_1037 = 0;
int main(void)
{
v_1037 = 18014398509481982wb;
double d = v_1037;
printf("d = %f\n", d);
return 0;
}
prints "d = 9007199254740992.000000", which is incorrect.
The correct result is "d = 18014398509481982.000000".
As stated in
https://discourse.llvm.org/t/rfc-llc-add-expandlargeintfpconvert-pass-for-fp-int-conversion-of-large-bitint/65528,
this implementation is very similar to ExpandLargeDivRem, which expands
‘fptoui .. to’, ‘fptosi .. to’, ‘uitofp .. to’, ‘sitofp .. to’ instructions
with a bitwidth above a threshold into auto-generated functions. This is
useful for targets like x86_64 that cannot lower fp convertions with more
than 128 bits. The expanded nodes are referring from the IR generated by
`compiler-rt/lib/builtins/floattidf.c`, `compiler-rt/lib/builtins/fixdfti.c`,
and etc.
Corner cases:
1. For fp16: as there is no related builtins added in compliler-rt. So I
mainly utilized the fp32 <-> fp16 lib calls to implement.
2. For fp80: as this pass is soft fp emulation and no fp80 instructions can
help in this problem. I recommend users to deprecate this usage. For now, the
implementation uses fp128 as the temporary conversion type and inserts
fptrunc/ext at top/end of the function.
3. For bf16: as clang FE currently doesn't support bf16 algorithm operations
(convert to int, float, +, -, *, ...), this patch doesn't consider bf16 for
now.
4. For unsigned FPToI: since both default hardware behaviors and libgcc are
ignoring "returns 0 for negative input" spec. This pass follows this old way
to ignore unsigned FPToI. See this example:
https://gcc.godbolt.org/z/bnv3jqW1M
The end-to-end tests are uploaded at https://reviews.llvm.org/D138261
Reviewed By: LuoYuanke, mgehre-amd
Differential Revision: https://reviews.llvm.org/D137241
