Enable machine sinking critical edge splitting. e.g.

define double @foo(double %x, double %y, i1 %c) nounwind {
  %a = fdiv double %x, 3.2
  %z = select i1 %c, double %a, double %y
  ret double %z
}

Was:
_foo:
        divsd   LCPI0_0(%rip), %xmm0
        testb   $1, %dil
        jne     LBB0_2
        movaps  %xmm1, %xmm0
LBB0_2:
        ret

Now:
_foo:
        testb   $1, %dil
        je      LBB0_2
        divsd   LCPI0_0(%rip), %xmm0
        ret
LBB0_2:
        movaps  %xmm1, %xmm0
        ret

This avoids the divsd when early exit is taken.
rdar://8454886


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@114372 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Evan Cheng 2010-09-20 22:52:00 +00:00
parent e3c65d1ede
commit 44be1a8d66
5 changed files with 41 additions and 27 deletions

View File

@ -35,10 +35,7 @@ using namespace llvm;
static cl::opt<bool>
SplitEdges("machine-sink-split",
cl::desc("Split critical edges during machine sinking"),
cl::init(false), cl::Hidden);
static cl::opt<unsigned>
SplitLimit("split-limit",
cl::init(~0u), cl::Hidden);
cl::init(true), cl::Hidden);
STATISTIC(NumSunk, "Number of machine instructions sunk");
STATISTIC(NumSplit, "Number of critical edges split");
@ -311,7 +308,7 @@ MachineBasicBlock *MachineSinking::SplitCriticalEdge(MachineInstr *MI,
return 0;
// Avoid breaking back edge. From == To means backedge for single BB loop.
if (!SplitEdges || NumSplit == SplitLimit || FromBB == ToBB)
if (!SplitEdges || FromBB == ToBB)
return 0;
// Check for backedges of more "complex" loops.
@ -561,8 +558,6 @@ bool MachineSinking::SinkInstruction(MachineInstr *MI, bool &SawStore) {
// BreakPHIEdge is true if all the uses are in the successor MBB being
// sunken into and they are all PHI nodes. In this case, machine-sink must
// break the critical edge first.
if (NumSplit == SplitLimit)
return false;
MachineBasicBlock *NewSucc = SplitCriticalEdge(MI, ParentBlock,
SuccToSinkTo, BreakPHIEdge);
if (!NewSucc) {

View File

@ -9,12 +9,12 @@ entry:
volatile store i32 0, i32* %c, align 4
%0 = volatile load i32* %a, align 4 ; <i32> [#uses=1]
%1 = icmp eq i32 %0, 0 ; <i1> [#uses=1]
; CHECK: addiu $4, $zero, 3
; CHECK: addiu $3, $zero, 0
%iftmp.0.0 = select i1 %1, i32 3, i32 0 ; <i32> [#uses=1]
%2 = volatile load i32* %c, align 4 ; <i32> [#uses=1]
%3 = icmp eq i32 %2, 0 ; <i1> [#uses=1]
; CHECK: addu $4, $zero, $3
; CHECK: addu $2, $5, $4
; CHECK: addiu $3, $zero, 3
; CHECK: addu $2, $5, $3
%iftmp.2.0 = select i1 %3, i32 0, i32 5 ; <i32> [#uses=1]
%4 = add nsw i32 %iftmp.2.0, %iftmp.0.0 ; <i32> [#uses=1]
ret i32 %4

View File

@ -5,7 +5,7 @@
; CHECK: oeq_inff:
; CHECK: ucomiss
; CHECK: jae
; CHECK: jb
define float @oeq_inff(float %x, float %y) nounwind readonly {
%t0 = fcmp oeq float %x, 0x7FF0000000000000
%t1 = select i1 %t0, float 1.0, float %y
@ -14,7 +14,7 @@ define float @oeq_inff(float %x, float %y) nounwind readonly {
; CHECK: oeq_inf:
; CHECK: ucomisd
; CHECK: jae
; CHECK: jb
define double @oeq_inf(double %x, double %y) nounwind readonly {
%t0 = fcmp oeq double %x, 0x7FF0000000000000
%t1 = select i1 %t0, double 1.0, double %y
@ -23,7 +23,7 @@ define double @oeq_inf(double %x, double %y) nounwind readonly {
; CHECK: une_inff:
; CHECK: ucomiss
; CHECK: jb
; CHECK: jae
define float @une_inff(float %x, float %y) nounwind readonly {
%t0 = fcmp une float %x, 0x7FF0000000000000
%t1 = select i1 %t0, float 1.0, float %y
@ -32,7 +32,7 @@ define float @une_inff(float %x, float %y) nounwind readonly {
; CHECK: une_inf:
; CHECK: ucomisd
; CHECK: jb
; CHECK: jae
define double @une_inf(double %x, double %y) nounwind readonly {
%t0 = fcmp une double %x, 0x7FF0000000000000
%t1 = select i1 %t0, double 1.0, double %y
@ -41,7 +41,7 @@ define double @une_inf(double %x, double %y) nounwind readonly {
; CHECK: oeq_neg_inff:
; CHECK: ucomiss
; CHECK: jae
; CHECK: jb
define float @oeq_neg_inff(float %x, float %y) nounwind readonly {
%t0 = fcmp oeq float %x, 0xFFF0000000000000
%t1 = select i1 %t0, float 1.0, float %y
@ -50,7 +50,7 @@ define float @oeq_neg_inff(float %x, float %y) nounwind readonly {
; CHECK: oeq_neg_inf:
; CHECK: ucomisd
; CHECK: jae
; CHECK: jb
define double @oeq_neg_inf(double %x, double %y) nounwind readonly {
%t0 = fcmp oeq double %x, 0xFFF0000000000000
%t1 = select i1 %t0, double 1.0, double %y
@ -59,7 +59,7 @@ define double @oeq_neg_inf(double %x, double %y) nounwind readonly {
; CHECK: une_neg_inff:
; CHECK: ucomiss
; CHECK: jb
; CHECK: jae
define float @une_neg_inff(float %x, float %y) nounwind readonly {
%t0 = fcmp une float %x, 0xFFF0000000000000
%t1 = select i1 %t0, float 1.0, float %y
@ -68,7 +68,7 @@ define float @une_neg_inff(float %x, float %y) nounwind readonly {
; CHECK: une_neg_inf:
; CHECK: ucomisd
; CHECK: jb
; CHECK: jae
define double @une_neg_inf(double %x, double %y) nounwind readonly {
%t0 = fcmp une double %x, 0xFFF0000000000000
%t1 = select i1 %t0, double 1.0, double %y

View File

@ -6,10 +6,11 @@
; that it's conditionally evaluated.
; CHECK: foo:
; CHECK: divsd
; CHECK-NEXT: testb $1, %dil
; CHECK-NEXT: jne
; CHECK-NEXT: je
; CHECK-NEXT: divsd
; CHECK-NEXT: ret
; CHECK: divsd
define double @foo(double %x, double %y, i1 %c) nounwind {
%a = fdiv double %x, 3.2
@ -18,6 +19,24 @@ define double @foo(double %x, double %y, i1 %c) nounwind {
ret double %z
}
; Make sure the critical edge is broken so the divsd is sunken below
; the conditional branch.
; rdar://8454886
; CHECK: split:
; CHECK-NEXT: testb $1, %dil
; CHECK-NEXT: je
; CHECK-NEXT: divsd
; CHECK-NEXT: ret
; CHECK: movaps
; CHECK-NEXT: ret
define double @split(double %x, double %y, i1 %c) nounwind {
%a = fdiv double %x, 3.2
%z = select i1 %c, double %a, double %y
ret double %z
}
; Hoist floating-point constant-pool loads out of loops.
; CHECK: bar:
@ -68,9 +87,9 @@ return:
; Codegen should hoist and CSE these constants.
; CHECK: vv:
; CHECK: LCPI2_0(%rip), %xmm0
; CHECK: LCPI2_1(%rip), %xmm1
; CHECK: LCPI2_2(%rip), %xmm2
; CHECK: LCPI3_0(%rip), %xmm0
; CHECK: LCPI3_1(%rip), %xmm1
; CHECK: LCPI3_2(%rip), %xmm2
; CHECK: align
; CHECK-NOT: LCPI
; CHECK: ret

View File

@ -50,9 +50,9 @@ define i32 @f3(i32 %a) {
ret i32 %2
}
; CHECK: f3:
; CHECK-NEXT: ashr r1, r0, 32
; CHECK-NEXT: ashr r0, r0, 32
; CHECK-NEXT: bf r0
; CHECK-NEXT: ldc r0, 10
; CHECK-NEXT: bt r1
; CHECK: ldc r0, 17
define i32 @f4(i32 %a) {
@ -61,9 +61,9 @@ define i32 @f4(i32 %a) {
ret i32 %2
}
; CHECK: f4:
; CHECK-NEXT: ashr r1, r0, 32
; CHECK-NEXT: ashr r0, r0, 32
; CHECK-NEXT: bf r0
; CHECK-NEXT: ldc r0, 17
; CHECK-NEXT: bt r1
; CHECK: ldc r0, 10
define i32 @f5(i32 %a) {