Kmeakin added the bug label to Issue #9041.
Kmeakin added the cranelift label to Issue #9041.
Kmeakin opened issue #9041:
Adding this optimization to
codegen/src/opts/icmp.isle
;; sge(x, c) == sgt(x, c-1), for c != SMIN. (rule (simplify (sge (fits_in_64 (ty_int bty)) x (iconst cty (u64_from_imm64 c)))) (if-let $false (u64_eq c (ty_smin cty))) (sgt bty x (iconst cty (imm64 (u64_sub c 1)))))
Triggers a verification error when optimizing
filetests/filetests/egraph/icmp-parameterized.clif
:FAIL filetests/filetests/egraph/icmp-parameterized.clif: optimize Caused by: function %icmp_sgt_umax(i32) -> i8 fast { block0(v0: i32): v5 = iconst.i32 -1 ; ^~~~~~~~~~~~~~~~~~ ; error: inst5 (v5 = iconst.i32 -1): constant immediate is out of bounds v10 = icmp sgt v0, v5 ; v5 = -1 return v10 } ; 1 verifier error detected (see above). Compilation aborted.
I suspect somewhere the immediate is being sign extended when it should be zero extended
bjorn3 commented on issue #9041:
-1
in clif ir text files is-1i64
, which is the same asu64::MAX
. Cranelift used to ignore the upper half of the immediate, but was changed to expect the upper half to be zeroed. We didn't yet change imm64 to uimm64 yet though, so immediates are still represented as signed integers, despite unsigned integers effectively being expected by the backend.
Kmeakin commented on issue #9041:
Kmeakin edited a comment on issue #9041:
Fixed by https://github.com/bytecodealliance/wasmtime/commit/0683b84b40207d8ec0c9044ef418fbd1f23a62f9
Kmeakin edited a comment on issue #9041:
~Fixed by https://github.com/bytecodealliance/wasmtime/commit/0683b84b40207d8ec0c9044ef418fbd1f23a62f9~
Last updated: Nov 22 2024 at 17:03 UTC