nv50/ir: swap the least-ref'd source into src1 when both const/imm

The whole point of inlining sources is to reduce loads. We can end up in
a situation where one value is used a lot of times, and one value is
used only once per instruction. The once-per-instruction one is the one
that should get inlined, but with the previous algorithm, it was given
no preference.

This flips things around to preferring putting less-referenced values
into src1 which increases the likelihood of them being inlined.

While we're at it, adjust the heuristic to not treat 0 as an immediate,
as well as (effectively) check for situations where LIMMs can't be
loaded. All this yields improvements on nvc0:

total instructions in shared programs : 6261157 -> 6255985 (-0.08%)
total gprs used in shared programs    : 945082 -> 943417 (-0.18%)
total local used in shared programs   : 30372 -> 30288 (-0.28%)
total bytes used in shared programs   : 50089256 -> 50047880 (-0.08%)

                local        gpr       inst      bytes
    helped          21         822        3332        3332
      hurt           0         278         565         565

And more importantly avoids generating really bad code with SSBOs, where
we end up checking a lot of different values (usually immediates) against
the length.

On nv50 we get comparable results, and even improve packing (bytes went
down more than instructions):

total instructions in shared programs : 6346564 -> 6341277 (-0.08%)
total gprs used in shared programs    : 728719 -> 725131 (-0.49%)
total local used in shared programs   : 3552 -> 3552 (0.00%)
total bytes used in shared programs   : 43995688 -> 43932928 (-0.14%)

                local        gpr       inst      bytes
    helped           0        1380        3252        3774
      hurt           0         287        1710        1365

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
This commit is contained in:
Ilia Mirkin 2016-01-17 22:28:19 -05:00
parent af686e7de3
commit a31819cff8
1 changed files with 15 additions and 10 deletions

View File

@ -171,7 +171,10 @@ LoadPropagation::isImmdLoad(Instruction *ld)
if (!ld || (ld->op != OP_MOV) ||
((typeSizeof(ld->dType) != 4) && (typeSizeof(ld->dType) != 8)))
return false;
return ld->src(0).getFile() == FILE_IMMEDIATE;
// A 0 can be replaced with a register, so it doesn't count as an immediate.
ImmediateValue val;
return ld->src(0).getImmediate(val) && !val.isInteger(0);
}
bool
@ -187,7 +190,8 @@ LoadPropagation::isAttribOrSharedLoad(Instruction *ld)
void
LoadPropagation::checkSwapSrc01(Instruction *insn)
{
if (!prog->getTarget()->getOpInfo(insn).commutative)
const Target *targ = prog->getTarget();
if (!targ->getOpInfo(insn).commutative)
if (insn->op != OP_SET && insn->op != OP_SLCT)
return;
if (insn->src(1).getFile() != FILE_GPR)
@ -196,14 +200,15 @@ LoadPropagation::checkSwapSrc01(Instruction *insn)
Instruction *i0 = insn->getSrc(0)->getInsn();
Instruction *i1 = insn->getSrc(1)->getInsn();
if (isCSpaceLoad(i0)) {
if (!isCSpaceLoad(i1))
insn->swapSources(0, 1);
else
return;
} else
if (isImmdLoad(i0)) {
if (!isCSpaceLoad(i1) && !isImmdLoad(i1))
// Swap sources to inline the less frequently used source. That way,
// optimistically, it will eventually be able to remove the instruction.
int i0refs = insn->getSrc(0)->refCount();
int i1refs = insn->getSrc(1)->refCount();
if ((isCSpaceLoad(i0) || isImmdLoad(i0)) && targ->insnCanLoad(insn, 1, i0)) {
if ((!isImmdLoad(i1) && !isCSpaceLoad(i1)) ||
!targ->insnCanLoad(insn, 1, i1) ||
i0refs < i1refs)
insn->swapSources(0, 1);
else
return;