f11a938ad5
With this CL the registers in a register pair get ordered such that the low word register always has a lower register code than the high word register. This should allow easier reasoning about the register allocation, and prevent some register allocation bugs. Background: for many operations in Liftoff, input registers are reused as output registers. With register pairs, input register pairs are reused as output register pairs. Additional reasoning, and sometimes even additional code is needed when the registers of the output register pair are swapped, i.e. when the high word register of the input becomes the low word register of the output. With this CL the additional reasoning is not necessary anymore, as the high word and low word registers would get swapped during register allocation. Additionally this CL fixes the logic of the last_spilled_regs list. This list stored the last spilled registers, but recorded only one of the two registers of a register pair. With this CL, both registers get recorded. This CL does not have a regression test. The regression test was more than 9000 lines long, and quite slow. I was not able to minimize it substantially. The test would be fragile by nature, as it has to create a special register configuration or otherwise does not test anything meaningful. All in all I think it's better not to add the test. R=clemensb@chromium.org Bug: chromium:1074586 Change-Id: I4b2475b0c6537c7ce2e51fee281388cdd85f2953 Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2168875 Reviewed-by: Clemens Backes <clemensb@chromium.org> Commit-Queue: Andreas Haas <ahaas@chromium.org> Cr-Commit-Position: refs/heads/master@{#67473} |
||
---|---|---|
.. | ||
benchmarks | ||
cctest | ||
common | ||
debugger | ||
debugging | ||
fuzzer | ||
inspector | ||
intl | ||
js-perf-test | ||
memory | ||
message | ||
mjsunit | ||
mkgrokdump | ||
mozilla | ||
test262 | ||
torque | ||
unittests | ||
wasm-api-tests | ||
wasm-js | ||
wasm-spec-tests | ||
webkit | ||
BUILD.gn | ||
OWNERS |