mirror of
https://github.com/RPCS3/llvm-mirror.git
synced 2024-11-24 03:33:20 +01:00
e3e67d4a0a
This changes the SelectionDAG scheduling preference to source order. Soon, the SelectionDAG scheduler can be bypassed saving a nice chunk of compile time. Performance differences that result from this change are often a consequence of register coalescing. The register coalescer is far from perfect. Bugs can be filed for deficiencies. On x86 SandyBridge/Haswell, the source order schedule is often preserved, particularly for small blocks. Register pressure is generally improved over the SD scheduler's ILP mode. However, we are still able to handle large blocks that require latency hiding, unlike the SD scheduler's BURR mode. MI scheduler also attempts to discover the critical path in single-block loops and adjust heuristics accordingly. The MI scheduler relies on the new machine model. This is currently unimplemented for AVX, so we may not be generating the best code yet. Unit tests are updated so they don't depend on SD scheduling heuristics. llvm-svn: 192750
52 lines
1.7 KiB
LLVM
52 lines
1.7 KiB
LLVM
; RUN: llc -march=x86 -mcpu=generic -mattr=+sse4.2 < %s | FileCheck %s
|
|
; RUN: llc -march=x86 -mcpu=atom < %s | FileCheck -check-prefix=ATOM %s
|
|
|
|
; CHECK: movl
|
|
; CHECK: paddd
|
|
; CHECK: movlpd
|
|
|
|
; Scheduler causes produce a different instruction order
|
|
; ATOM: movl
|
|
; ATOM: paddd
|
|
; ATOM: movlpd
|
|
|
|
; bitcast a v4i16 to v2i32
|
|
|
|
define void @convert(<2 x i32>* %dst, <4 x i16>* %src) nounwind {
|
|
entry:
|
|
%dst.addr = alloca <2 x i32>* ; <<2 x i32>**> [#uses=2]
|
|
%src.addr = alloca <4 x i16>* ; <<4 x i16>**> [#uses=2]
|
|
%i = alloca i32, align 4 ; <i32*> [#uses=6]
|
|
store <2 x i32>* %dst, <2 x i32>** %dst.addr
|
|
store <4 x i16>* %src, <4 x i16>** %src.addr
|
|
store i32 0, i32* %i
|
|
br label %forcond
|
|
|
|
forcond: ; preds = %forinc, %entry
|
|
%tmp = load i32* %i ; <i32> [#uses=1]
|
|
%cmp = icmp slt i32 %tmp, 4 ; <i1> [#uses=1]
|
|
br i1 %cmp, label %forbody, label %afterfor
|
|
|
|
forbody: ; preds = %forcond
|
|
%tmp1 = load i32* %i ; <i32> [#uses=1]
|
|
%tmp2 = load <2 x i32>** %dst.addr ; <<2 x i32>*> [#uses=1]
|
|
%arrayidx = getelementptr <2 x i32>* %tmp2, i32 %tmp1 ; <<2 x i32>*> [#uses=1]
|
|
%tmp3 = load i32* %i ; <i32> [#uses=1]
|
|
%tmp4 = load <4 x i16>** %src.addr ; <<4 x i16>*> [#uses=1]
|
|
%arrayidx5 = getelementptr <4 x i16>* %tmp4, i32 %tmp3 ; <<4 x i16>*> [#uses=1]
|
|
%tmp6 = load <4 x i16>* %arrayidx5 ; <<4 x i16>> [#uses=1]
|
|
%add = add <4 x i16> %tmp6, < i16 1, i16 1, i16 1, i16 1 > ; <<4 x i16>> [#uses=1]
|
|
%conv = bitcast <4 x i16> %add to <2 x i32> ; <<2 x i32>> [#uses=1]
|
|
store <2 x i32> %conv, <2 x i32>* %arrayidx
|
|
br label %forinc
|
|
|
|
forinc: ; preds = %forbody
|
|
%tmp7 = load i32* %i ; <i32> [#uses=1]
|
|
%inc = add i32 %tmp7, 1 ; <i32> [#uses=1]
|
|
store i32 %inc, i32* %i
|
|
br label %forcond
|
|
|
|
afterfor: ; preds = %forcond
|
|
ret void
|
|
}
|