llvm-6502/test/CodeGen/NVPTX/machine-sink.ll
Jingyue Wu 75d77cd179 [MachineSink] Use the real post dominator tree
Summary:
Fixes a FIXME in MachineSinking. Instead of using the simple heuristics in
isPostDominatedBy, use the real MachinePostDominatorTree and MachineLoopInfo.
The old heuristics caused instructions to sink unnecessarily, and might create
register pressure.

This is the second try of the fix. The first one (D4814) caused a performance
regression due to failing to sink instructions out of loops (PR21115). This
patch fixes PR21115 by sinking an instruction from a deeper loop to a shallower
one regardless of whether the target block post-dominates the source.

Thanks Alexey Volkov for reporting PR21115! 

Test Plan:
Added a NVPTX codegen test to verify that our change prevents the backend from
over-sinking. It also shows the unnecessary register pressure caused by
over-sinking.

Added an X86 test to verify we can sink instructions out of loops regardless of
the dominance relationship. This test is reduced from Alexey's test in PR21115.

Updated an affected test in X86.

Also ran SPEC CINT2006 and llvm-test-suite for compilation time and runtime
performance. Results are attached separately in the review thread.

Reviewers: Jiangning, resistor, hfinkel

Reviewed By: hfinkel

Subscribers: hfinkel, bruno, volkalexey, llvm-commits, meheff, eliben, jholewinski

Differential Revision: http://reviews.llvm.org/D5633

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219773 91177308-0d34-0410-b5e6-96231b3b80d8
2014-10-15 03:27:43 +00:00

41 lines
1.4 KiB
LLVM

; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
@scalar1 = internal addrspace(3) global float 0.000000e+00, align 4
@scalar2 = internal addrspace(3) global float 0.000000e+00, align 4
; We shouldn't sink mul.rn.f32 to BB %merge because BB %merge post-dominates
; BB %entry. Over-sinking created more register pressure on this example. The
; backend would sink the fmuls to BB %merge, but not the loads for being
; conservative on sinking memory accesses. As a result, the loads and
; the two fmuls would be separated to two basic blocks, causing two
; cross-BB live ranges.
define float @post_dominate(float %x, i1 %cond) {
; CHECK-LABEL: post_dominate(
entry:
%0 = load float* addrspacecast (float addrspace(3)* @scalar1 to float*), align 4
%1 = load float* addrspacecast (float addrspace(3)* @scalar2 to float*), align 4
; CHECK: ld.shared.f32
; CHECK: ld.shared.f32
%2 = fmul float %0, %0
%3 = fmul float %1, %2
; CHECK-NOT: bra
; CHECK: mul.rn.f32
; CHECK: mul.rn.f32
br i1 %cond, label %then, label %merge
then:
%z = fadd float %x, %x
br label %then2
then2:
%z2 = fadd float %z, %z
br label %merge
merge:
%y = phi float [ 0.0, %entry ], [ %z2, %then2 ]
%w = fadd float %y, %3
ret float %w
}