summary refs log tree commit diff stats
path: root/results/scraper/fex/3805
diff options
context:
space:
mode:
Diffstat (limited to 'results/scraper/fex/3805')
-rw-r--r--results/scraper/fex/38052
1 files changed, 2 insertions, 0 deletions
diff --git a/results/scraper/fex/3805 b/results/scraper/fex/3805
new file mode 100644
index 000000000..f65141530
--- /dev/null
+++ b/results/scraper/fex/3805
@@ -0,0 +1,2 @@
+AVX128: Prescale gathers with 64-bit index and scale factor of 2 or 4
+We know that overflow behaviour with 64-bit indexes just throws away the top bits. Instead of falling back to the ASIMD implementation when the scale factor is 2 or 4, just prescale the values in the vector registers and then pass scale to the IR operation as 1. This will let it continue to use the faster SVE codepath while just eating two more instructions compared to 1 or 8 scaling.
\ No newline at end of file