i965/fs: Fix compute_to_mrf() to coalesce VGRFs initialized by multiple single-GRF writes.

Which requires using a bitset instead of a boolean flag to keep track
of the GRFs we've seen a generating instruction for already.  The
search loop continues until all instructions initializing the value of
the source VGRF have been found, or it is determined that coalescing
is not possible.

Fixes a few piglit test cases on Gen4-6 which were regressed by
6956015aa5 due to the different (yet
perfectly valid) ordering in which copy instructions are emitted now
by the simd lowering pass, which had the side effect of causing this
optimization pass to start corrupting the program in cases where a
VGRF-to-MRF copy instruction would be eliminated but only the last
instruction writing to the source VGRF region would be rewritten to
point to the target MRF.

Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
(cherry picked from commit 4fe4f6e8a7)
This commit is contained in:
Francisco Jerez 2016-05-27 16:03:34 -07:00 committed by Emil Velikov
parent 910fa7a824
commit 44029d4237

View file

@ -2767,6 +2767,20 @@ fs_visitor::opt_redundant_discard_jumps()
return progress;
}
/**
* Compute a bitmask with GRF granularity with a bit set for each GRF starting
* from \p r which overlaps the region starting at \p r and spanning \p n GRF
* units.
*/
static inline unsigned
mask_relative_to(const fs_reg &r, const fs_reg &s, unsigned n)
{
const int rel_offset = (reg_offset(s) - reg_offset(r)) / REG_SIZE;
assert(reg_space(r) == reg_space(s) &&
rel_offset >= 0 && rel_offset < int(8 * sizeof(unsigned)));
return ((1 << n) - 1) << rel_offset;
}
bool
fs_visitor::compute_to_mrf()
{
@ -2798,10 +2812,12 @@ fs_visitor::compute_to_mrf()
if (this->virtual_grf_end[inst->src[0].nr] > ip)
continue;
/* Found a move of a GRF to a MRF. Let's see if we can go
* rewrite the thing that made this GRF to write into the MRF.
/* Found a move of a GRF to a MRF. Let's see if we can go rewrite the
* things that computed the value of all GRFs of the source region. The
* regs_left bitset keeps track of the registers we haven't yet found a
* generating instruction for.
*/
bool found = false;
unsigned regs_left = (1 << inst->regs_read(0)) - 1;
foreach_inst_in_block_reverse_starting_from(fs_inst, scan_inst, inst) {
if (regions_overlap(scan_inst->dst, scan_inst->regs_written * REG_SIZE,
@ -2840,10 +2856,11 @@ fs_visitor::compute_to_mrf()
}
}
if (scan_inst->dst.reg_offset == inst->src[0].reg_offset)
found = true;
break;
/* Clear the bits for any registers this instruction overwrites. */
regs_left &= ~mask_relative_to(
inst->src[0], scan_inst->dst, scan_inst->regs_written);
if (!regs_left)
break;
}
/* We don't handle control flow here. Most computation of
@ -2886,14 +2903,21 @@ fs_visitor::compute_to_mrf()
}
}
if (!found)
if (regs_left)
continue;
/* Found all generating instructions of our MRF's source value.
/* Found all generating instructions of our MRF's source value, so it
* should be safe to rewrite them to point to the MRF directly.
*/
regs_left = (1 << inst->regs_read(0)) - 1;
foreach_inst_in_block_reverse_starting_from(fs_inst, scan_inst, inst) {
if (regions_overlap(scan_inst->dst, scan_inst->regs_written * REG_SIZE,
inst->src[0], inst->regs_read(0) * REG_SIZE)) {
/* Clear the bits for any registers this instruction overwrites. */
regs_left &= ~mask_relative_to(
inst->src[0], scan_inst->dst, scan_inst->regs_written);
const unsigned rel_offset = (reg_offset(scan_inst->dst) -
reg_offset(inst->src[0])) / REG_SIZE;
@ -2920,10 +2944,12 @@ fs_visitor::compute_to_mrf()
scan_inst->dst.file = MRF;
scan_inst->dst.reg_offset = 0;
scan_inst->saturate |= inst->saturate;
break;
if (!regs_left)
break;
}
}
assert(!regs_left);
inst->remove(block);
progress = true;
}