https://llvm.org/bugs/show_bug.cgi?id=28130
Bug ID: 28130
Summary: Missing xxswapd before a stxvd2x
Product: libraries
Version: trunk
Hardware: PC
OS: All
Status: NEW
Severity: normal
Priority: P
Component: Backend: PowerPC
Assignee: unassignedb...@nondot.org
Reporter: timshe...@gmail.com
CC: llvm-bugs@lists.llvm.org
Classification: Unclassified
Created attachment 16542
--> https://llvm.org/bugs/attachment.cgi?id=16542&action=edit
reduced test
The problem is in generated assembly, for the only stxvd2x, there isn't a
preceding xxswapd instruction to recover the "endianness" for a VSX register.
Suppose the attached file is a.ll:
~ % llc --version
LLVM (http://llvm.org/):
LLVM version 3.9.0svn
Optimized build with assertions.
Default target: x86_64-unknown-linux-gnu
Host CPU: haswell
~ % llc -O0 < a.ll
.text
.abiversion 2
.file "<stdin>"
.section .rodata.cst4,"aM",@progbits,4
.p2align 2
.LCPI0_0:
.long 1178406912 # float 12100
.LCPI0_1:
.long 1120403456 # float 100
.LCPI0_5:
.long 1180762112 # float 14400
.section .rodata.cst16,"aM",@progbits,16
.p2align 4
.LCPI0_2:
.quad 4667272928281231360 # double 11000
.quad 4637440978796412928 # double 110
.LCPI0_3:
.quad 4607182418800017408 # double 1
.quad 4666723172467343360 # double 1.0E+4
.LCPI0_4:
.quad 4667822684095119360 # double 12000
.quad 4638144666238189568 # double 120
.text
.globl Test
.p2align 4
.type Test,@function
Test: # @Test
.Lfunc_begin0:
.Lfunc_gep0:
addis 2, 12, .TOC.-.Lfunc_gep0@ha
addi 2, 2, .TOC.-.Lfunc_gep0@l
.Lfunc_lep0:
.localentry Test, .Lfunc_lep0-.Lfunc_gep0
# BB#0: # %bb
lxsdx 0, 0, 3
li 4, 40
lxsdx 1, 3, 4
li 4, 32
lxsdx 2, 3, 4
addis 5, 2, .LCPI0_0@toc@ha
addi 5, 5, .LCPI0_0@toc@l
lxsspx 3, 0, 5
xssubdp 3, 3, 2
xsmuldp 3, 0, 3
xsadddp 2, 2, 3
li 5, 48
lxsdx 3, 3, 5
addis 6, 2, .LCPI0_1@toc@ha
addi 6, 6, .LCPI0_1@toc@l
lxsspx 4, 0, 6
xssubdp 5, 4, 3
xsmuldp 5, 0, 5
xsadddp 3, 3, 5
li 6, 56
lxsdx 5, 3, 6
fmr 6, 1
xxspltd 6, 6, 0
fmr 7, 5
xxspltd 7, 7, 0
xxmrgld 6, 7, 6
addis 7, 2, .LCPI0_2@toc@ha
addi 7, 7, .LCPI0_2@toc@l
lxvd2x 7, 0, 7
xxswapd 7, 7
xvsubdp 7, 7, 6
fmr 8, 0
xxspltd 8, 8, 0
xxspltd 8, 8, 1
xvmuldp 7, 8, 7
xvadddp 6, 6, 7
li 7, 16
lxvd2x 7, 3, 7
xxswapd 7, 7
addis 8, 2, .LCPI0_3@toc@ha
addi 8, 8, .LCPI0_3@toc@l
lxvd2x 9, 0, 8
xxswapd 9, 9
xvsubdp 10, 9, 7
xvmuldp 10, 8, 10
xvadddp 7, 7, 10
xvsubdp 9, 9, 7
xvmuldp 9, 8, 9
xvadddp 7, 7, 9
addis 8, 2, .LCPI0_4@toc@ha
addi 8, 8, .LCPI0_4@toc@l
lxvd2x 9, 0, 8
xxswapd 9, 9
xvsubdp 9, 9, 6
xvmuldp 8, 8, 9
xvadddp 6, 6, 8
addis 8, 2, .LCPI0_5@toc@ha
addi 8, 8, .LCPI0_5@toc@l
lxsspx 1, 0, 8
xssubdp 1, 1, 2
xsmuldp 1, 0, 1
xsadddp 1, 2, 1
stxsdx 1, 3, 4
xssubdp 1, 4, 3
xsmuldp 0, 0, 1
xsadddp 0, 3, 0
stxsdx 0, 3, 5
fmr 0, 6
stxsdx 0, 3, 6
# <== expect `xxswapd 7, 7` here
stxvd2x 7, 3, 7 # <== vsr[7] should be "endianness" swapped.
blr
.long 0
.quad 0
.Lfunc_end0:
.size Test, .Lfunc_end0-.Lfunc_begin0
.section ".note.GNU-stack","",@progbits
Investigation progress:
Disabling the call to DAGCombiner::findBetterNeighborChains() makes the problem
go away. I'm not sure if it's DAGCombiner::findBetterNeighborChains's fault and
I'm still looking.
--
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
