https://bugs.llvm.org/show_bug.cgi?id=49971
Bug ID: 49971
Summary: [X86] 12.0.0 Regression - foldShuffleOfHorizOp doesn't
correctly handle 256-bit X86ISD::VBROADCAST
Product: libraries
Version: trunk
Hardware: PC
OS: Windows NT
Status: NEW
Severity: enhancement
Priority: P
Component: Backend: X86
Assignee: [email protected]
Reporter: [email protected]
CC: [email protected], [email protected],
[email protected], [email protected],
[email protected]
Blocks: 49317
Encountered by an internal test:
define <4 x double> @broadcast_hadd_pd(<4 x double> %0) {
%2 = tail call <4 x double> @llvm.x86.avx.hadd.pd.256(<4 x double> %0, <4 x
double> %0)
%3 = shufflevector <4 x double> %2, <4 x double> undef, <4 x i32> <i32 0, i32
1, i32 0, i32 1>
ret <4 x double> %3
}
declare <4 x double> @llvm.x86.avx.hadd.pd.256(<4 x double>, <4 x double>)
llc -march=znver2
trunk:
broadcast_hadd_pd:
vhaddpd %xmm0, %xmm0, %xmm0
vbroadcastsd %xmm0, %ymm0
retq
llvm12:
broadcast_hadd_pd:
vhaddpd %ymm0, %ymm0, %ymm0 ; WHERE DID THE SPLAT GO?
retq
llvm11:
broadcast_hadd_pd:
vhaddpd %xmm0, %xmm0, %xmm0
vpermpd $68, %ymm0, %ymm0 # ymm0 = ymm0[0,1,0,1]
retq
The issue is that foldShuffleOfHorizOp incorrectly assumes that a 256-bit
broadcast(vhaddpd(x,y)) can be replaced by vhaddpd(x,x).
The safest solution appears to be to drop X86ISD::VBROADCAST handling from
foldShuffleOfHorizOp entirely - canonicalizeShuffleMaskWithHorizOp already
handles it.
In trunk we've already removed all VBROADCAST/MOVDDUP handling from
foldShuffleOfHorizOp (2a0d5da917f1) but in 12.x we still need the MOVDDUP
support in some cases.
Referenced Bugs:
https://bugs.llvm.org/show_bug.cgi?id=49317
[Bug 49317] [meta] 12.0.1 Release Blockers
--
You are receiving this mail because:
You are on the CC list for the bug._______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
