[LLVMbugs] [Bug 23289] New: Trivial values not rematerialized

[bugzilla-daemon at llvm.org]

https://llvm.org/bugs/show_bug.cgi?id=23289

            Bug ID: 23289
           Summary: Trivial values not rematerialized
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: normal
          Priority: P
         Component: Register Allocator
          Assignee: unassignedbugs at nondot.org
          Reporter: james.molloy at arm.com
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

The following code produces really suboptimal code seemingly on all targets.

#define N 16
extern void g(int);
int f(int *p, int n) {
    for (int i = 0; i < n; ++i) {
        for (int k = 0; k < n; ++k) {
#pragma clang loop unroll(full)
            for (int j = 0; j < N; ++j)
                g((j+i)*k);
        }
    }
}

It produces a bunch of spills: (main loop block):

.LBB0_2:                                @ %.preheader.lr.ph.us                 
                                                                               
                            [31/1853]
                                        @ =>This Loop Header: Depth=1
                                        @     Child Loop BB0_3 Depth 2
        mov     r5, r4
        mov     r9, #0
        add     r0, r5, #11
        add     r6, r5, #4
        add     r7, r5, #3
        add     r8, r5, #2
        add     r4, r5, #1
        str     r0, [sp, #24]           @ 4-byte Spill
        add     r0, r5, #10
        str     r0, [sp, #20]           @ 4-byte Spill
        add     r0, r5, #9
        str     r0, [sp, #16]           @ 4-byte Spill
        add     r0, r5, #8
        str     r0, [sp, #12]           @ 4-byte Spill
        add     r0, r5, #7
        str     r0, [sp, #8]            @ 4-byte Spill
        add     r0, r5, #6
        str     r0, [sp, #4]            @ 4-byte Spill
        add     r0, r5, #5
        str     r0, [sp]                @ 4-byte Spill
.LBB0_3:                                @ %.preheader.us
                                        @   Parent Loop BB0_2 Depth=1
                                        @ =>  This Inner Loop Header: Depth=2
        mul     r0, r5, r9
        bl      g
        mul     r0, r4, r9
        bl      g
        mul     r0, r8, r9
        bl      g
        mul     r0, r7, r9
        bl      g
        mul     r0, r6, r9
        bl      g
        ldr     r0, [sp]                @ 4-byte Reload
        mul     r0, r0, r9
        bl      g
        ldr     r0, [sp, #4]            @ 4-byte Reload
        mul     r0, r0, r9
        bl      g
        ldr     r0, [sp, #8]            @ 4-byte Reload
... snip, more like this ...

X86 also shows massive spilling. The instructions outside the loop are
trivially rematerializable, but for some reason seem not to be considered.

Because this affects all targets, I suspect this is a bug in the core RA/Spill
placement algorithm.

-- 
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/20150419/e8f17efb/attachment.html>