mirror of
https://github.com/RPCS3/llvm.git
synced 2024-12-28 23:43:50 +00:00
6a7770b7ae
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. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@192750 91177308-0d34-0410-b5e6-96231b3b80d8
47 lines
2.6 KiB
LLVM
47 lines
2.6 KiB
LLVM
; RUN: llc < %s -march=x86-64 | grep lea | count 13
|
|
|
|
; This testcase was written to demonstrate an instruction-selection problem,
|
|
; however it also happens to expose a limitation in the DAGCombiner's
|
|
; expression reassociation which causes it to miss opportunities for
|
|
; constant folding due to the intermediate adds having multiple uses.
|
|
; The Reassociate pass has similar limitations. If these limitations are
|
|
; fixed, the test commands above will need to be updated to expect fewer
|
|
; lea instructions.
|
|
|
|
@g0 = weak global [1000 x i32] zeroinitializer, align 32 ; <[1000 x i32]*> [#uses=8]
|
|
@g1 = weak global [1000 x i32] zeroinitializer, align 32 ; <[1000 x i32]*> [#uses=7]
|
|
|
|
define void @foo() {
|
|
entry:
|
|
%tmp4 = load i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 0) ; <i32> [#uses=1]
|
|
%tmp8 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 0) ; <i32> [#uses=1]
|
|
%tmp9 = add i32 %tmp4, 1 ; <i32> [#uses=1]
|
|
%tmp10 = add i32 %tmp9, %tmp8 ; <i32> [#uses=2]
|
|
store i32 %tmp10, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 1)
|
|
%tmp8.1 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 1) ; <i32> [#uses=1]
|
|
%tmp9.1 = add i32 %tmp10, 1 ; <i32> [#uses=1]
|
|
%tmp10.1 = add i32 %tmp9.1, %tmp8.1 ; <i32> [#uses=2]
|
|
store i32 %tmp10.1, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 2)
|
|
%tmp8.2 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 2) ; <i32> [#uses=1]
|
|
%tmp9.2 = add i32 %tmp10.1, 1 ; <i32> [#uses=1]
|
|
%tmp10.2 = add i32 %tmp9.2, %tmp8.2 ; <i32> [#uses=2]
|
|
store i32 %tmp10.2, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 3)
|
|
%tmp8.3 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 3) ; <i32> [#uses=1]
|
|
%tmp9.3 = add i32 %tmp10.2, 1 ; <i32> [#uses=1]
|
|
%tmp10.3 = add i32 %tmp9.3, %tmp8.3 ; <i32> [#uses=2]
|
|
store i32 %tmp10.3, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 4)
|
|
%tmp8.4 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 4) ; <i32> [#uses=1]
|
|
%tmp9.4 = add i32 %tmp10.3, 1 ; <i32> [#uses=1]
|
|
%tmp10.4 = add i32 %tmp9.4, %tmp8.4 ; <i32> [#uses=2]
|
|
store i32 %tmp10.4, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 5)
|
|
%tmp8.5 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 5) ; <i32> [#uses=1]
|
|
%tmp9.5 = add i32 %tmp10.4, 1 ; <i32> [#uses=1]
|
|
%tmp10.5 = add i32 %tmp9.5, %tmp8.5 ; <i32> [#uses=2]
|
|
store i32 %tmp10.5, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 6)
|
|
%tmp8.6 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 6) ; <i32> [#uses=1]
|
|
%tmp9.6 = add i32 %tmp10.5, 1 ; <i32> [#uses=1]
|
|
%tmp10.6 = add i32 %tmp9.6, %tmp8.6 ; <i32> [#uses=1]
|
|
store i32 %tmp10.6, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 7)
|
|
ret void
|
|
}
|