[llvm-bugs] [Bug 36698] New: X87 FPU inline assembly input operands loaded after the assembly on x86_32

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=36698

            Bug ID: 36698
           Summary: X87 FPU inline assembly input operands loaded after
                    the assembly on x86_32
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: martin at martin.st
                CC: llvm-bugs at lists.llvm.org

If building x87 FPU inline assembly for x86_32 with clang, LLVM can reorder the
loads of input arguments to after the actual assembly snippet.

An example function:

void asmfunc(double in, double* outptr) {
  double out;
  asm("nop":"=t"(out):"0"(in));
  if (outptr)
    *outptr = out;
}

When built with e.g. "clang -S -O2 x87.c -target i686-linux-gnu -o -", this
produces:

asmfunc:                                # @asmfunc
# %bb.0:                                # %entry
        subl    $12, %esp
        movl    24(%esp), %eax
        testl   %eax, %eax
        je      .LBB0_2
# %bb.1:                                # %if.then
        movsd   16(%esp), %xmm0         # xmm0 = mem[0],zero
        #APP
        nop
        #NO_APP
        movsd   %xmm0, (%esp)
        fldl    (%esp)                  # <---- Input operand to the #APP
snippet above is loaded too late
        fstpl   (%eax)
.LBB0_2:                                # %if.end
        addl    $12, %esp
        retl


If the resulting LLVM IR is lowered with -print-after-all, I get the following
steps:

# *** IR Dump After Post-RA pseudo instruction expansion pass ***:

...
  LD_F64m $esp, 1, $noreg, 0, $noreg, implicit-def dead $fpsw;
mem:LD8[FixedStack0](align=4)  
  INLINEASM &nop [attdialect], $0:[regdef], implicit-def $st0, $1:[reguse
tiedto:$0], killed $st0(tied-def 3), $2:[clobber], implicit-def dead
early-clobber $eflags, !3  
  ST_FP64m killed renamable $eax, 1, $noreg, 0, $noreg, implicit-def dead
$fpsw; mem:ST8[%outptr](align=4)(tbaa=!5)
...

# *** IR Dump After Post RA top-down list latency scheduler ***:

...
  INLINEASM &nop [attdialect], $0:[regdef], implicit-def $st0, $1:[reguse
tiedto
:$0], $st0(tied-def 3), $2:[clobber], implicit-def dead early-clobber $eflags,
!
3
  MOVSDmr $esp, 1, $noreg, 0, $noreg, killed renamable $xmm0;
mem:ST8[FixedStack
0](align=4)
  LD_F64m $esp, 1, $noreg, 0, $noreg, implicit-def dead $fpsw;
mem:LD8[FixedStac
k0](align=4)
  ST_FP64m killed renamable $eax, 1, $noreg, 0, $noreg, implicit-def dead
$fpsw;
 mem:ST8[%outptr](align=4)(tbaa=!5)
...


If the inline assembly has the "volatile" attribute added, the resulting code
turns out ok.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20180312/52535b0d/attachment-0001.html>