On 6/3/25 5:24 PM, Vineet Gupta wrote:


So as of tip of tree, even w/o the patch we no longer see the  pattern "beq
<reg>, 0, .L12 "
They are replaced with reg anyways, so the test won't even show issue, even w/o
the fix.

     beq    a1,a5,.L12
     beq    a6,t0,.L12
     beq    a7,t3,.L12

I think a better proxy of improvement would be to grep for vsetvls with null
output VL.

orig
     vsetvli    a5,a1,e8,mf4,ta,ma
     vsetvli    t0,a6,e8,mf4,ta,ma
     vsetvli    zero,a5,e32,m1,tu,ma
     vsetvli    zero,t0,e32,m1,ta,ma
     vsetvli    t3,a7,e8,mf4,ta,ma
     vsetvli    t4,t1,e8,mf4,ta,ma
     vsetvli    zero,t0,e32,m1,tu,ma
     vsetvli    zero,t3,e32,m1,tu,ma
     vsetvli    zero,t4,e32,m1,tu,ma

patch
     vsetvli    a5,a1,e32,m1,tu,ma
     vsetvli    t1,t0,e32,m1,tu,ma
     vsetvli    a2,a1,e32,m1,tu,ma
     vsetvli    t4,t3,e32,m1,tu,ma
Yea, we should have probably realized this test would end up being temperamental. Not sure if there's any knobs we can turn to stabilize it over time either.

I'll leave it up to you and Edwin. As far as I'm concerned the patch is ACK'd for the trunk once the test is adjusted.

jeff

Reply via email to