Craig Topper

LLVM developer. Compiler engineer at SiFive. Opinions are my own.

[Craig Topper @aurigas81]

@CamelCdr @bekket_mcclane This code assumes the LMUL=2 mask is split into separate registers?

[Craig Topper @aurigas81]

@CamelCdr @pshufb @CamelCdr yes this is actively being worked on.

[Craig Topper @aurigas81]

@pshufb Looks like Clang didn't use RVV because there is no vredmul instruction. So it used fixed vectors so it could generate a fixed number of shuffles to emulate it.

[Craig Topper @aurigas81]

@pshufb @dyaroshev Clang is using vsetvli zero, zero, e64, m2, ta, mu to change the vtype without changing vl. This should be the cheapest form of vsetvli.

[Craig Topper @aurigas81]

@foonathan @CompileExplore -mllvm -trap-unreachable will insert a ud2 in main instead of falling through.

[Craig Topper @aurigas81]

@tom_forsyth Curious how you figured out that’s what the cat wanted the first time.

[Craig Topper @aurigas81]

@damageboy You can pass -mllvm -riscv-v-register-bit-width-lmul=2 to change the LMUL.

[Craig Topper @aurigas81]

@damageboy For better or worse, llvm made a decision a long time ago to convert shuffle/permute intrinsics into generic IR shuffles and allow them to be combined. There is a substantial amount of code in the X86 backend devoted to untangling it back into instructions that exist.

[Craig Topper @aurigas81]

@damageboy @nadavrot Doesn’t look like it would be too hard to fix. Are you going to file a bug?

[Craig Topper @aurigas81]

@lenary @stephentyrone Redacted size_t. You need clearance to know the size.

[Craig Topper @aurigas81]

@johnregehr LLVM does like to truncate the input to SwitchInst to weird integer widths. I wonder how many of these are from that.

[Craig Topper @aurigas81]

@tom_forsyth @rygorous @pervognsen @_monoid I believe it’s modulo 32, not modulo operand size. Except for 64-bit shifts.

[Craig Topper @aurigas81]

@pshufb -Wuninitialized doesn't detect it?

[Craig Topper @aurigas81]

@trav_downs @lemire @richardstartin Is the add keeping clang from seeing that the popcnt is being compared with zero? Hopefully we handle that for a single popcnt?

[Craig Topper @aurigas81]

@rygorous https://t.co/TUVmhX5J0Z I believe we are zeroing for intrinsics already.

[Craig Topper @aurigas81]

@kamidphish That appears to be this gcc bug https://t.co/mTy1CGyxVe

[Craig Topper @aurigas81]

@kamidphish But doesn't seem to be handling NaNs the same way as clang and icc. clang and icc check that the Z flag is set and the P flag is cleared. The Z flag is set for equality or if either input is a NaN. Checking P rules out the latter.

[Craig Topper @aurigas81]

@kamidphish clang isn't thinking about the if statement and the ucomieq as one operation. ucomieq returns an integer, that's the code up to the movzx. Then the result gets compared to 0.

[Craig Topper @aurigas81]

@rocallahan @khuey_ @johnregehr @stephentyrone @jckarter I imagine LZCNT and BSR are even worse?