38200e94f1
Always try to fold freeze(op(....)) -> op(freeze(),freeze(),freeze(),...). This patch proposes we drop the opt-in limit for opcodes that are allowed to push a freeze through the op to freeze all its operands, through the tree towards the roots. I'm struggling to find a strong reason for this limit apart from the DAG freeze handling being immature for so long - as we've improved coverage in canCreateUndefOrPoison/isGuaranteedNotToBeUndefOrPoison it looks like the regressions are not as severe. Hopefully this will help some of the regression issues in #143102 etc.
33 lines
1.3 KiB
LLVM
33 lines
1.3 KiB
LLVM
; REQUIRES: asserts
|
|
; RUN: llc -o /dev/null %s -debug-only=legalize-types 2>&1 | FileCheck %s
|
|
|
|
; This test check that when v4f64 gets broken down to two v2f64 it maintains
|
|
; the "nnan" flags.
|
|
|
|
; CHECK: Legalizing node: [[VFOUR:t.*]]: v4f64 = BUILD_VECTOR
|
|
; CHECK-NEXT: Analyzing result type: v4f64
|
|
; CHECK-NEXT: Split node result: [[VFOUR]]: v4f64 = BUILD_VECTOR
|
|
|
|
; CHECK: Legalizing node: [[VTWOA:t.*]]: v2f64 = BUILD_VECTOR
|
|
; CHECK: Legally typed node: [[VTWOA]]: v2f64 = BUILD_VECTOR
|
|
; CHECK: Legalizing node: [[VTWOB:t.*]]: v2f64 = BUILD_VECTOR
|
|
; CHECK: Legally typed node: [[VTWOB]]: v2f64 = BUILD_VECTOR
|
|
; CHECK: Legalizing node: t31: v2f64 = fmaxnum nnan reassoc [[VTWOB]], [[VTWOA]]
|
|
|
|
target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
|
|
target triple = "aarch64--linux-gnu"
|
|
|
|
|
|
; Function Attrs: norecurse nounwind
|
|
define fastcc double @test(double %a0, double %a1, double %a2, double %a3) unnamed_addr #1 {
|
|
entry:
|
|
%0 = insertelement <4 x double> undef, double %a0, i32 0
|
|
%1 = insertelement <4 x double> %0, double %a1, i32 1
|
|
%2 = insertelement <4 x double> %1, double %a2, i32 2
|
|
%3 = insertelement <4 x double> %2, double %a3, i32 3
|
|
%4 = call nnan reassoc double @llvm.vector.reduce.fmax.v4f64(<4 x double> %3)
|
|
ret double %4
|
|
}
|
|
|
|
declare double @llvm.vector.reduce.fmax.v4f64(<4 x double>)
|