glasm: Catch more register leaks

Add support for null registers. These are used when an instruction has
no usages.

This comes handy when an instruction is only used for its CC value, with
the caveat of having to invalidate all pseudo-instructions before
defining the instruction itself in the register allocator. This commits
changes this.

Workaround a bug on Nvidia's condition codes conditional execution using
branches.
This commit is contained in:
ReinUsesLisp 2021-05-25 02:22:21 -03:00 committed by ameerj
parent 9fbfe7d676
commit ca05a13c62
8 changed files with 114 additions and 41 deletions

View file

@ -51,6 +51,10 @@ void EmitSubgroupGeMask(EmitContext& ctx, IR::Inst& inst) {
static void Shuffle(EmitContext& ctx, IR::Inst& inst, ScalarU32 value, ScalarU32 index,
const IR::Value& clamp, const IR::Value& segmentation_mask,
std::string_view op) {
IR::Inst* const in_bounds{inst.GetAssociatedPseudoOperation(IR::Opcode::GetInBoundsFromOp)};
if (in_bounds) {
in_bounds->Invalidate();
}
std::string mask;
if (clamp.IsImmediate() && segmentation_mask.IsImmediate()) {
mask = fmt::to_string(clamp.U32() | (segmentation_mask.U32() << 8));
@ -61,13 +65,11 @@ static void Shuffle(EmitContext& ctx, IR::Inst& inst, ScalarU32 value, ScalarU32
ScalarU32{ctx.reg_alloc.Consume(clamp)});
}
const Register value_ret{ctx.reg_alloc.Define(inst)};
IR::Inst* const in_bounds{inst.GetAssociatedPseudoOperation(IR::Opcode::GetInBoundsFromOp)};
if (in_bounds) {
const Register bounds_ret{ctx.reg_alloc.Define(*in_bounds)};
ctx.Add("SHF{}.U {},{},{},{};"
"MOV.U {}.x,{}.y;",
op, bounds_ret, value, index, mask, value_ret, bounds_ret);
in_bounds->Invalidate();
} else {
ctx.Add("SHF{}.U {},{},{},{};"
"MOV.U {}.x,{}.y;",