llvm-mca reporting error on valid x86 assembly.

[hiraditya]

h:
        push    r14 
        push    rbx 
        push    rax 
        mov     r14, qword ptr [rdi]
        cmp     r14, qword ptr [rdi + 8]
        je      .LBB2_3
        mov     rbx, rdi 
.LBB2_2:
        mov     edi, dword ptr [r14]
        call    f(int)@PLT
        add     r14, 4
        cmp     r14, qword ptr [rbx + 8]
        jne     .LBB2_2
.LBB2_3:
        add     rsp, 8
        pop     rbx 
        pop     r14 
        ret

$ ./build/bin/llvm-mca -mtriple x86_64-unknown-linux-gnu -iterations=20 a.s

a.s:5:28: error: unexpected token in argument list
        mov     r14, qword ptr [rdi]
                           ^
a.s:6:28: error: unexpected token in argument list
        cmp     r14, qword ptr [rdi + 8]
                           ^
a.s:8:9: error: unknown use of instruction mnemonic without a size suffix
        mov     rbx, rdi
        ^
a.s:10:28: error: unexpected token in argument list
        mov     edi, dword ptr [r14]
                           ^
a.s:11:19: error: expected register here
        call    f(int)@PLT
                  ^
a.s:12:9: error: unknown use of instruction mnemonic without a size suffix
        add     r14, 4
        ^
a.s:13:28: error: unexpected token in argument list
        cmp     r14, qword ptr [rbx + 8]
                           ^
a.s:16:9: error: unknown use of instruction mnemonic without a size suffix
        add     rsp, 8
        ^
a.s:1:1: error: symbol 'h' is already defined
h:
^
a.s:5:28: error: unexpected token in argument list
        mov     r14, qword ptr [rdi]
                           ^
a.s:6:28: error: unexpected token in argument list
        cmp     r14, qword ptr [rdi + 8]
                           ^
a.s:8:9: error: unknown use of instruction mnemonic without a size suffix
        mov     rbx, rdi
        ^
a.s:9:1: error: symbol '.LBB2_2' is already defined
.LBB2_2:
^
a.s:10:28: error: unexpected token in argument list
        mov     edi, dword ptr [r14]
                           ^
a.s:11:19: error: expected register here
        call    f(int)@PLT
                  ^
a.s:12:9: error: unknown use of instruction mnemonic without a size suffix
        add     r14, 4
        ^
a.s:13:28: error: unexpected token in argument list
        cmp     r14, qword ptr [rbx + 8]
                           ^
a.s:15:1: error: symbol '.LBB2_3' is already defined
.LBB2_3:
^
a.s:16:9: error: unknown use of instruction mnemonic without a size suffix
        add     rsp, 8
        ^

[llvmbot]

@llvm/issue-subscribers-tools-llvm-mca

[llvmbot]

@llvm/issue-subscribers-bug

[topperc]

llvm-mca defaults to AT&T syntax. You need to add .intel_syntax to the top of the file

[hiraditya]

adding .intel_syntax gives another set of errors.

.intel_syntax
h:
        push    r14 
        push    rbx 
        push    rax 
        mov     r14, qword ptr [rdi]
        cmp     r14, qword ptr [rdi + 8]
        je      .LBB2_3
        mov     rbx, rdi 
.LBB2_2:
        mov     edi, dword ptr [r14]
        call    f(int)@PLT
        add     r14, 4
        cmp     r14, qword ptr [rbx + 8]
        jne     .LBB2_2
.LBB2_3:
        add     rsp, 8
        pop     rbx 
        pop     r14 
        ret

$ ./build/bin/llvm-mca --version
LLVM (http://llvm.org/):
  LLVM version 17.0.0git
  Optimized build.
  Default target: x86_64-unknown-linux-gnu
  Host CPU: tigerlake

$ ./build/bin/llvm-mca x.s -mcpu=btver2



x.s:12:18: error: unknown token in expression
        call    f(int)@PLT
                 ^
x.s:2:1: error: symbol 'h' is already defined
h:
^
x.s:10:1: error: symbol '.LBB2_2' is already defined
.LBB2_2:
^
x.s:12:18: error: unknown token in expression
        call    f(int)@PLT
                 ^
x.s:16:1: error: symbol '.LBB2_3' is already defined
.LBB2_3:

[topperc]

I don't see the error: symbol 'h' is already defined or error: symbol '.LBB2_3' is already defined errors.

The error: unknown token in expression is because compiler explorer ran a C++ demangler on the output before you gave it to llvm-mca. Turn off "Demangle identifiers" in the "Output..." drop down

[hiraditya]

nice catch!
with the mangled symbols, the error: unknown token in expression error goes away but the other errors are still there. Seems like llvm-mca can provide some diagnostics to help their users.

_Z1hNSt3__16vectorIiNS_9allocatorIiEEEE:
# %bb.0:
	pushq	%r14
	pushq	%rbx
	pushq	%rax
	movq	(%rdi), %r14
	cmpq	8(%rdi), %r14
	je	.LBB0_3
# %bb.1:
	movq	%rdi, %rbx
	.p2align	4, 0x90
.LBB0_2:                                # =>This Inner Loop Header: Depth=1
	movl	(%r14), %edi
	callq	_Z1fi@PLT
	addq	$4, %r14
	cmpq	8(%rbx), %r14
	jne	.LBB0_2
.LBB0_3:
	addq	$8, %rsp
	popq	%rbx
	popq	%r14
	retq

$ ./build/bin/llvm-mca y1.s -mcpu=btver2

y1.s:1:1: error: symbol '_Z1hNSt3__16vectorIiNS_9allocatorIiEEEE' is already defined
_Z1hNSt3__16vectorIiNS_9allocatorIiEEEE:
^
y1.s:12:1: error: symbol '.LBB0_2' is already defined
.LBB0_2:                                # =>This Inner Loop Header: Depth=1
^
y1.s:18:1: error: symbol '.LBB0_3' is already defined
.LBB0_3:
^