Wasmtime GitHub notifications bot (Feb 03 2021 at 19:56): Wasmtime GitHub notifications bot (Feb 03 2021 at 19:56):bjorn3 commented on Issue #1119:
Is this still the case with regalloc.rs?
Wasmtime GitHub notifications bot (Feb 03 2021 at 20:09):cfallin commented on Issue #1119:
We've looked at whether to use callee- or caller-saved registers preferentially in the new backends and the current heuristic is to allocate caller-saved registers first, with the additional detail that the clobber modeling (every call instruction
defs all clobbered registers) will naturally push values that are live across calls into callee-saved registers where they're available. So I think we should be able to close this as it's an issue specific to the old regalloc.
Wasmtime GitHub notifications bot (Feb 03 2021 at 20:09):cfallin closed Issue #1119:
To start, some clif:
; guest_func_10 function u0:10(i64 vmctx, i32, i32) -> i32 system_v { sig0 = (i64 vmctx, i32, i32) -> i32 system_v fn0 = colocated u0:10 sig0 ebb0(v0: i64, v1: i32, v2: i32): @09ca v4 = iconst.i32 0 @09d0 v5 = icmp_imm eq v1, 0 @09d0 v6 = bint.i32 v5 @09d1 brnz v6, ebb2(v2) @09d3 jump ebb3(v1, v2) ebb3(v7: i32, v10: i32): @09d7 v8 = iconst.i32 -1 @09d9 v9 = iadd v7, v8 @09de brz v10, ebb5 @09e4 v11 = iconst.i32 -1 @09e6 v12 = iadd v10, v11 @09e7 v13 = call fn0(v0, v7, v12) @09ef brnz v9, ebb3(v9, v13) @09f1 jump ebb2(v13) ebb5: @09f4 v14 = iconst.i32 1 @09fc brnz.i32 v9, ebb3(v9, v14) @09fe jump ebb4 ebb4: @09ff jump ebb2(v14) ebb2(v15: i32): @0a02 v16 = iconst.i32 1 @0a04 v17 = iadd v15, v16 @0a05 jump ebb1(v17) ebb1(v3: i32): @0a05 return v3this had been producing
sym.guest_func_10 (); 0x00003ac0 push rbp 0x00003ac1 mov rbp, rsp 0x00003ac4 sub rsp, 0x10 0x00003ac8 mov qword [rsp], rdi 0x00003ad0 mov dword [local_ch], edx 0x00003ad7 test esi, esi ,=< 0x00003ad9 je 0x3b56 | 0x00003adb mov eax, 1 | 0x00003ae0 mov dword [local_ch], eax ..--> 0x00003ae7 mov eax, esi ::| 0x00003ae9 add eax, -1 ::| 0x00003aec mov dword [local_8h], eax ::| 0x00003af3 test edx, edx ,====< 0x00003af5 je 0x3b31 |::| 0x00003af7 add edx, -1 |::| 0x00003afa mov rax, qword [rsp] |::| 0x00003b02 mov rdi, rax |::| 0x00003b05 call sym.guest_func_10 |::| 0x00003b0a mov ecx, dword [local_8h] |::| 0x00003b11 mov edx, dword [local_8h] |::| 0x00003b18 mov esi, ecx |::| 0x00003b1b mov ecx, edx |::| 0x00003b1e mov edx, eax |::| 0x00003b21 mov eax, ecx |::| 0x00003b24 test eax, eax |`===< 0x00003b26 jne 0x3ae7 | :| 0x00003b28 mov dword [local_ch], edx |,===< 0x00003b2f jmp 0x3b56 `----> 0x00003b31 mov eax, dword [local_ch] |:| 0x00003b38 mov ecx, dword [local_8h] |:| 0x00003b3f mov edx, dword [local_8h] |:| 0x00003b46 mov esi, ecx |:| 0x00003b49 mov ecx, edx |:| 0x00003b4c mov edx, eax |:| 0x00003b4f mov eax, ecx |:| 0x00003b52 test eax, eax |`==< 0x00003b54 jne 0x3ae7 `-`-> 0x00003b56 mov eax, dword [local_ch] 0x00003b5d add eax, 1 0x00003b60 add rsp, 0x10 0x00003b64 pop rbp 0x00003b65 retwith redundant reloads
in
164f91athe redundant reloads are removed (:tada:) but ends up picking (on x86-64) callee-save registers to load into, then producingsym.guest_func_10 (); 0x00003ac0 push rbp 0x00003ac1 mov rbp, rsp 0x00003ac4 push r12 0x00003ac6 push r13 0x00003ac8 push r14 0x00003aca sub rsp, 0x18 0x00003ace mov qword [local_8h], rdi 0x00003ad6 mov dword [local_14h], edx 0x00003add test esi, esi ,=< 0x00003adf je 0x3b48 | 0x00003ae1 mov eax, 1 | 0x00003ae6 mov dword [local_14h], eax ..--> 0x00003aed mov eax, esi ::| 0x00003aef add eax, -1 ::| 0x00003af2 mov dword [local_10h], eax ::| 0x00003af9 test edx, edx ,====< 0x00003afb je 0x3b2f |::| 0x00003afd add edx, -1 |::| 0x00003b00 mov r14, qword [local_8h] |::| 0x00003b08 mov rdi, r14 |::| 0x00003b0b call sym.guest_func_10 |::| 0x00003b10 mov r14d, dword [local_10h] |::| 0x00003b18 mov r13d, r14d |::| 0x00003b1b mov esi, r14d |::| 0x00003b1e mov edx, eax |::| 0x00003b21 test r13d, r13d |`===< 0x00003b24 jne 0x3aed | :| 0x00003b26 mov dword [local_14h], edx |,===< 0x00003b2d jmp 0x3b48 `----> 0x00003b2f mov r14d, dword [local_14h] |:| 0x00003b37 mov r13d, eax |:| 0x00003b3a mov r12d, eax |:| 0x00003b3d mov esi, r13d |:| 0x00003b40 mov edx, r14d |:| 0x00003b43 test r12d, r12d |`==< 0x00003b46 jne 0x3aed `-`-> 0x00003b48 mov r14d, dword [local_14h] 0x00003b50 add r14d, 1 0x00003b54 mov eax, r14d 0x00003b57 add rsp, 0x18 0x00003b5b pop r14 0x00003b5d pop r13 0x00003b5f pop r12 0x00003b61 pop rbp 0x00003b62 retbecause of the callee-save use here we end up net +3 memory accesses from push/pops, even though three redundant loads were removed
In a particularly contorted benchmark (Lucet's
ackermannbenchmark) this ends up dropping runtime by ~30%, though I think in most cases it's probably about the same before and after.ediit: didn't include addresses initially, which didn't read well
Last updated: Jan 09 2026 at 13:15 UTC