incorrect instruction selected for materializing 4-way add of 0x01010101

[zygoloid]

Bugzilla Link	34271
Resolution	FIXED
Resolved on	Aug 22, 2017 09:17
Version	trunk
OS	All
Blocks	#33196
CC	@chandlerc,@topperc,@zmodem,@RKSimon,@rotateright

Extended Description

We miscompile ffmpeg's mjpegenc_common.c due to a bad instruction being selected to materialize a vector constant. Testcase:

target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

define <4 x i32> @​f(<4 x i32> %a) {
%v = add nuw nsw <4 x i32> %a, <i32 16843009, i32 16843009, i32 16843009, i32 16843009>
ret <4 x i32> %v
}

This compiles to:

    pcmpeqd %xmm1, %xmm1
    psubd   %xmm1, %xmm0

... which computes %a + {1, 1, 1, 1}, not %a + {0x01010101, 0x01010101, 0x01010101, 0x01010101}.

[topperc]

Here's a similar case that results in an assertion failure.

./bin/llc test.ll -o - -mattr=avx512vl,avx512vbmi

define <16 x i16> @​foo(<16 x i32> %i) {
%x3 = icmp ult <16 x i32> %i, <i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009>
%x5 = select <16 x i1> %x3, <16 x i32> %i, <16 x i32> <i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009, i32 16843009>
%x6 = trunc <16 x i32> %x5 to <16 x i16>
ret <16 x i16> %x6
}

These both boil down to the X86 backend not understanding that ISD::isConstantSplatVector returns an APInt with the smallest splat size. Both these cases return an 8-bit APInt containing a value of 1. We then interpret that as being a 1 in the original scalar size.

[topperc]

Patch up for review https://reviews.llvm.org/D36996

[chandlerc]

CC-ing Hans as this may be in the 5.0 release as well...

[topperc]

Fixed in r311429.

Both test cases fail on 5.0.

[zmodem]

CC-ing Hans as this may be in the 5.0 release as well...

Thanks! Merged in r311462.