Remat VLDRD from constpool. Clean up some instruction property specifications.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@89478 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Evan Cheng 2009-11-20 19:57:15 +00:00
parent 87b75ba75e
commit 4aedb61d03
7 changed files with 77 additions and 13 deletions

View File

@ -581,7 +581,6 @@ def PICADD : AXI1<0b0100, (outs GPR:$dst), (ins GPR:$a, pclabel:$cp, pred:$p),
[(set GPR:$dst, (ARMpic_add GPR:$a, imm:$cp))]>;
let AddedComplexity = 10 in {
let canFoldAsLoad = 1 in
def PICLDR : AXI2ldw<(outs GPR:$dst), (ins addrmodepc:$addr, pred:$p),
Pseudo, IIC_iLoadr, "\n${addr:label}:\n\tldr$p\t$dst, $addr",
[(set GPR:$dst, (load addrmodepc:$addr))]>;
@ -801,13 +800,14 @@ let isBranch = 1, isTerminator = 1 in {
//
// Load
let canFoldAsLoad = 1, isReMaterializable = 1 in
let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
def LDR : AI2ldw<(outs GPR:$dst), (ins addrmode2:$addr), LdFrm, IIC_iLoadr,
"ldr", "\t$dst, $addr",
[(set GPR:$dst, (load addrmode2:$addr))]>;
// Special LDR for loads from non-pc-relative constpools.
let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1 in
let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1,
mayHaveSideEffects = 1 in
def LDRcp : AI2ldw<(outs GPR:$dst), (ins addrmode2:$addr), LdFrm, IIC_iLoadr,
"ldr", "\t$dst, $addr", []>;

View File

@ -296,7 +296,7 @@ let isBranch = 1, isTerminator = 1 in {
// Load Store Instructions.
//
let canFoldAsLoad = 1 in
let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
def tLDR : T1pI4<(outs tGPR:$dst), (ins t_addrmode_s4:$addr), IIC_iLoadr,
"ldr", "\t$dst, $addr",
[(set tGPR:$dst, (load t_addrmode_s4:$addr))]>;
@ -332,13 +332,14 @@ def tRestore : T1pIs<(outs tGPR:$dst), (ins t_addrmode_sp:$addr), IIC_iLoadi,
// Load tconstpool
// FIXME: Use ldr.n to work around a Darwin assembler bug.
let canFoldAsLoad = 1 in
let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
def tLDRpci : T1pIs<(outs tGPR:$dst), (ins i32imm:$addr), IIC_iLoadi,
"ldr", ".n\t$dst, $addr",
[(set tGPR:$dst, (load (ARMWrapper tconstpool:$addr)))]>;
// Special LDR for loads from non-pc-relative constpools.
let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1 in
let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1,
mayHaveSideEffects = 1 in
def tLDRcp : T1pIs<(outs tGPR:$dst), (ins i32imm:$addr), IIC_iLoadi,
"ldr", "\t$dst, $addr", []>;

View File

@ -471,7 +471,7 @@ def t2SUBrSPs_ : PseudoInst<(outs GPR:$dst), (ins GPR:$sp, t2_so_reg:$rhs),
//
// Load
let canFoldAsLoad = 1 in
let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
defm t2LDR : T2I_ld<"ldr", UnOpFrag<(load node:$Src)>>;
// Loads with zero extension
@ -1183,7 +1183,7 @@ def t2MOVi32imm : T2Ix2<(outs GPR:$dst), (ins i32imm:$src), IIC_iMOVi,
// Pseudo instruction that combines ldr from constpool and add pc. This should
// be expanded into two instructions late to allow if-conversion and
// scheduling.
let isReMaterializable = 1 in
let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
def t2LDRpci_pic : PseudoInst<(outs GPR:$dst), (ins i32imm:$addr, pclabel:$cp),
NoItinerary, "@ ldr.w\t$dst, $addr\n$cp:\n\tadd\t$dst, pc",
[(set GPR:$dst, (ARMpic_add (load (ARMWrapper tconstpool:$addr)),

View File

@ -54,7 +54,7 @@ def vfp_f64imm : Operand<f64>,
// Load / store Instructions.
//
let canFoldAsLoad = 1 in {
let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in {
def VLDRD : ADI5<0b1101, 0b01, (outs DPR:$dst), (ins addrmode5:$addr),
IIC_fpLoad64, "vldr", ".64\t$dst, $addr",
[(set DPR:$dst, (load addrmode5:$addr))]>;

View File

@ -0,0 +1,65 @@
; RUN: llc < %s -march=arm -mattr=+v6,+vfp2 -stats -info-output-file - | grep "Number of re-materialization"
define arm_apcscc i32 @main(i32 %argc, i8** nocapture %argv) nounwind {
entry:
br i1 undef, label %smvp.exit, label %bb.i3
bb.i3: ; preds = %bb.i3, %bb134
br i1 undef, label %smvp.exit, label %bb.i3
smvp.exit: ; preds = %bb.i3
%0 = fmul double undef, 2.400000e-03 ; <double> [#uses=2]
br i1 undef, label %bb138.preheader, label %bb159
bb138.preheader: ; preds = %smvp.exit
br label %bb138
bb138: ; preds = %bb138, %bb138.preheader
br i1 undef, label %bb138, label %bb145.loopexit
bb142: ; preds = %bb.nph218.bb.nph218.split_crit_edge, %phi0.exit
%1 = fmul double undef, -1.200000e-03 ; <double> [#uses=1]
%2 = fadd double undef, %1 ; <double> [#uses=1]
%3 = fmul double %2, undef ; <double> [#uses=1]
%4 = fsub double 0.000000e+00, %3 ; <double> [#uses=1]
br i1 %14, label %phi1.exit, label %bb.i35
bb.i35: ; preds = %bb142
%5 = call arm_apcscc double @sin(double %15) nounwind readonly ; <double> [#uses=1]
%6 = fmul double %5, 0x4031740AFA84AD8A ; <double> [#uses=1]
%7 = fsub double 1.000000e+00, undef ; <double> [#uses=1]
%8 = fdiv double %7, 6.000000e-01 ; <double> [#uses=1]
br label %phi1.exit
phi1.exit: ; preds = %bb.i35, %bb142
%.pn = phi double [ %6, %bb.i35 ], [ 0.000000e+00, %bb142 ] ; <double> [#uses=0]
%9 = phi double [ %8, %bb.i35 ], [ 0.000000e+00, %bb142 ] ; <double> [#uses=1]
%10 = fmul double undef, %9 ; <double> [#uses=0]
br i1 %14, label %phi0.exit, label %bb.i
bb.i: ; preds = %phi1.exit
unreachable
phi0.exit: ; preds = %phi1.exit
%11 = fsub double %4, undef ; <double> [#uses=1]
%12 = fadd double 0.000000e+00, %11 ; <double> [#uses=1]
store double %12, double* undef, align 4
br label %bb142
bb145.loopexit: ; preds = %bb138
br i1 undef, label %bb.nph218.bb.nph218.split_crit_edge, label %bb159
bb.nph218.bb.nph218.split_crit_edge: ; preds = %bb145.loopexit
%13 = fmul double %0, 0x401921FB54442D18 ; <double> [#uses=1]
%14 = fcmp ugt double %0, 6.000000e-01 ; <i1> [#uses=2]
%15 = fdiv double %13, 6.000000e-01 ; <double> [#uses=1]
br label %bb142
bb159: ; preds = %bb145.loopexit, %smvp.exit, %bb134
unreachable
bb166: ; preds = %bb127
unreachable
}
declare arm_apcscc double @sin(double) nounwind readonly

View File

@ -1,4 +1,4 @@
; RUN: llc < %s -mtriple=thumbv7-apple-darwin9 -mcpu=cortex-a8 | grep vmov.f32 | count 4
; RUN: llc < %s -mtriple=thumbv7-apple-darwin9 -mcpu=cortex-a8 | grep vmov.f32 | count 6
define arm_apcscc void @fht(float* nocapture %fz, i16 signext %n) nounwind {
entry:

View File

@ -22,8 +22,7 @@
define arm_apcscc %union.rec* @Manifest(%union.rec* %x, %union.rec* %env, %struct.STYLE* %style, %union.rec** %bthr, %union.rec** %fthr, %union.rec** %target, %union.rec** %crs, i32 %ok, i32 %need_expand, %union.rec** %enclose, i32 %fcr) nounwind {
entry:
; CHECK: ldr.w r9, [r7, #+32]
; CHECK-NEXT : str.w r9, [sp, #+28]
; CHECK: ldr.w r9, [r7, #+28]
%xgaps.i = alloca [32 x %union.rec*], align 4 ; <[32 x %union.rec*]*> [#uses=0]
%ycomp.i = alloca [32 x %union.rec*], align 4 ; <[32 x %union.rec*]*> [#uses=0]
br i1 false, label %bb, label %bb20
@ -53,7 +52,6 @@ bb420: ; preds = %bb20, %bb20
; CHECK: str r{{[0-7]}}, [sp]
; CHECK: str r{{[0-7]}}, [sp, #+4]
; CHECK: str r{{[0-7]}}, [sp, #+8]
; CHECK: ldr r{{[0-7]}}, [sp, #+28]
; CHECK: str r{{[0-7]}}, [sp, #+24]
store %union.rec* null, %union.rec** @zz_hold, align 4
store %union.rec* null, %union.rec** @zz_res, align 4