[LLVMbugs] [Bug 3473] New: need load/store merging in dag combine

[bugzilla-daemon at cs.uiuc.edu]

http://llvm.org/bugs/show_bug.cgi?id=3473

           Summary: need load/store merging in dag combine
           Product: libraries
           Version: trunk
          Platform: PC
        OS/Version: All
            Status: NEW
          Keywords: code-quality
          Severity: normal
          Priority: P2
         Component: Common Code Generator Code
        AssignedTo: unassignedbugs at nondot.org
        ReportedBy: clattner at apple.com
                CC: llvmbugs at cs.uiuc.edu


This example:

struct s { short a, b, c, d; };
struct s foo(struct s x) {
  return x;
}

compiles into:

$ llvm-gcc t.c -S -o - -O3 -fomit-frame-pointer -m32
foo:
        movzwl  6(%esp), %ecx
        shll    $16, %ecx
        movzwl  4(%esp), %eax
        orl     %ecx, %eax
        movzwl  10(%esp), %ecx
        shll    $16, %ecx
        movzwl  8(%esp), %edx
        orl     %ecx, %edx
        ret

This is because the struct is passed with by-val, but then reassembled into an
i64 for returning.

        %struct.s = type { i16, i16, i16, i16 }
define i64 @foo(%struct.s* nocapture byval align 4 %x) nounwind readonly {
entry:
        %0 = getelementptr %struct.s* %x, i32 0, i32 0          ; <i16*>
[#uses=1]
        %1 = load i16* %0, align 2              ; <i16> [#uses=1]
        %2 = getelementptr %struct.s* %x, i32 0, i32 1          ; <i16*>
[#uses=1]
        %3 = load i16* %2, align 2              ; <i16> [#uses=1]
        %4 = getelementptr %struct.s* %x, i32 0, i32 2          ; <i16*>
[#uses=1]
        %5 = load i16* %4, align 2              ; <i16> [#uses=1]
        %6 = getelementptr %struct.s* %x, i32 0, i32 3          ; <i16*>
[#uses=1]
        %7 = load i16* %6, align 2              ; <i16> [#uses=1]
        %8 = zext i16 %1 to i64         ; <i64> [#uses=1]
        %9 = zext i16 %3 to i64         ; <i64> [#uses=1]
        %10 = shl i64 %9, 16            ; <i64> [#uses=1]
        %11 = zext i16 %5 to i64                ; <i64> [#uses=1]
        %12 = shl i64 %11, 32           ; <i64> [#uses=1]
        %13 = zext i16 %7 to i64                ; <i64> [#uses=1]
        %14 = shl i64 %13, 48           ; <i64> [#uses=1]
        %15 = or i64 %10, %8            ; <i64> [#uses=1]
        %16 = or i64 %15, %12           ; <i64> [#uses=1]
        %17 = or i64 %16, %14           ; <i64> [#uses=1]
        ret i64 %17
}

The code generator should merge the small loads (and similar cases with stores)
to avoid this packing logic.  gcc compiles this to:

_foo:
        movl    4(%esp), %eax
        movl    8(%esp), %edx
        ret


-- 
Configure bugmail: http://llvm.org/bugs/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.