189900eb14
This stops reporting CostPerUse 1 for `R8`-`R15` and `XMM8`-`XMM31`.
This was previously done because instruction encoding require a REX
prefix when using them resulting in longer instruction encodings. I
found that this regresses the quality of the register allocation as the
costs impose an ordering on eviction candidates. I also feel that there
is a bit of an impedance mismatch as the actual costs occure when
encoding instructions using those registers, but the order of VReg
assignments is not primarily ordered by number of Defs+Uses.
I did extensive measurements with the llvm-test-suite wiht SPEC2006 +
SPEC2017 included, internal services showed similar patterns. Generally
there are a log of improvements but also a lot of regression. But on
average the allocation quality seems to improve at a small code size
regression.
Results for measuring static and dynamic instruction counts:
Dynamic Counts (scaled by execution frequency) / Optimization Remarks:
Spills+FoldedSpills -5.6%
Reloads+FoldedReloads -4.2%
Copies -0.1%
Static / LLVM Statistics:
regalloc.NumSpills mean -1.6%, geomean -2.8%
regalloc.NumReloads mean -1.7%, geomean -3.1%
size..text mean +0.4%, geomean +0.4%
Static / LLVM Statistics:
mean -2.2%, geomean -3.1%) regalloc.NumSpills
mean -2.6%, geomean -3.9%) regalloc.NumReloads
mean +0.6%, geomean +0.6%) size..text
Static / LLVM Statistics:
regalloc.NumSpills mean -3.0%
regalloc.NumReloads mean -3.3%
size..text mean +0.3%, geomean +0.3%
Differential Revision: https://reviews.llvm.org/D133902
111 lines
4.5 KiB
LLVM
111 lines
4.5 KiB
LLVM
; RUN: llc < %s -frame-pointer=all -mtriple x86_64-apple-darwin11 -mcpu corei7 | FileCheck -check-prefix=ASM %s
|
|
; RUN: llc < %s -frame-pointer=all -mtriple x86_64-apple-darwin11 -mcpu corei7 -filetype=obj -o - \
|
|
; RUN: | llvm-objdump --triple=x86_64-apple-darwin11 --unwind-info - \
|
|
; RUN: | FileCheck -check-prefix=CU %s
|
|
; RUN: llc < %s -frame-pointer=all -mtriple x86_64-apple-darwin11 -mcpu corei7 \
|
|
; RUN: | llvm-mc -triple x86_64-apple-darwin11 -filetype=obj -o - \
|
|
; RUN: | llvm-objdump --triple=x86_64-apple-darwin11 --unwind-info - \
|
|
; RUN: | FileCheck -check-prefix=FROM-ASM %s
|
|
|
|
; RUN: llc < %s -mtriple x86_64-apple-macosx10.8.0 -mcpu corei7 -filetype=obj -o - \
|
|
; RUN: | llvm-objdump --triple=x86_64-apple-macosx10.8.0 --unwind-info - \
|
|
; RUN: | FileCheck -check-prefix=NOFP-CU %s
|
|
; RUN: llc < %s -mtriple x86_64-apple-darwin11 -mcpu corei7 \
|
|
; RUN: | llvm-mc -triple x86_64-apple-darwin11 -filetype=obj -o - \
|
|
; RUN: | llvm-objdump --triple=x86_64-apple-darwin11 --unwind-info - \
|
|
; RUN: | FileCheck -check-prefix=NOFP-FROM-ASM %s
|
|
|
|
%ty = type { ptr }
|
|
|
|
@gv = external global i32
|
|
|
|
; This is aligning the stack with a push of a random register.
|
|
; ASM: pushq %rax
|
|
|
|
; Even though we can't encode %rax into the compact unwind, We still want to be
|
|
; able to generate a compact unwind encoding in this particular case.
|
|
|
|
; CU: Contents of __compact_unwind section:
|
|
; CU-NEXT: Entry at offset 0x0:
|
|
; CU-NEXT: start: 0x0 _test0
|
|
; CU-NEXT: length: 0x1e
|
|
; CU-NEXT: compact encoding: 0x01010001
|
|
|
|
; FROM-ASM: Contents of __compact_unwind section:
|
|
; FROM-ASM-NEXT: Entry at offset 0x0:
|
|
; FROM-ASM-NEXT: start: 0x0 _test0
|
|
; FROM-ASM-NEXT: length: 0x1e
|
|
; FROM-ASM-NEXT: compact encoding: 0x01010001
|
|
|
|
define ptr @test0(i64 %size) {
|
|
%addr = alloca i64, align 8
|
|
%tmp20 = load i32, ptr @gv, align 4
|
|
%tmp21 = call i32 @bar()
|
|
%tmp25 = load i64, ptr %addr, align 8
|
|
%tmp26 = inttoptr i64 %tmp25 to ptr
|
|
%tmp34 = load ptr, ptr %tmp26, align 8
|
|
%tmp35 = getelementptr inbounds i8, ptr %tmp34, i64 %size
|
|
store ptr %tmp35, ptr %tmp26, align 8
|
|
ret ptr null
|
|
}
|
|
|
|
declare i32 @bar()
|
|
|
|
%"struct.dyld::MappedRanges" = type { [400 x %struct.anon], ptr }
|
|
%struct.anon = type { ptr, i64, i64 }
|
|
%class.ImageLoader = type { ptr, ptr, ptr, i32, i64, i64, i32, i32, ptr, i16, i16, [4 x i8] }
|
|
%"struct.ImageLoader::recursive_lock" = type { i32, i32 }
|
|
|
|
@G1 = external hidden global %"struct.dyld::MappedRanges", align 8
|
|
|
|
declare void @OSMemoryBarrier() optsize
|
|
|
|
; Test the code below uses UNWIND_X86_64_MODE_STACK_IMMD compact unwind
|
|
; encoding.
|
|
|
|
; NOFP-CU: Entry at offset 0x20:
|
|
; NOFP-CU-NEXT: start: 0x1d _test1
|
|
; NOFP-CU-NEXT: length: 0x44
|
|
; NOFP-CU-NEXT: compact encoding: 0x02040c0a
|
|
|
|
; NOFP-FROM-ASM: Entry at offset 0x20:
|
|
; NOFP-FROM-ASM-NEXT: start: 0x1d _test1
|
|
; NOFP-FROM-ASM-NEXT: length: 0x44
|
|
; NOFP-FROM-ASM-NEXT: compact encoding: 0x02040c0a
|
|
|
|
define void @test1(ptr %image) optsize ssp uwtable {
|
|
entry:
|
|
br label %for.cond1.preheader
|
|
|
|
for.cond1.preheader: ; preds = %for.inc10, %entry
|
|
%p.019 = phi ptr [ @G1, %entry ], [ %1, %for.inc10 ]
|
|
br label %for.body3
|
|
|
|
for.body3: ; preds = %for.inc, %for.cond1.preheader
|
|
%indvars.iv = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next, %for.inc ]
|
|
%image4 = getelementptr inbounds %"struct.dyld::MappedRanges", ptr %p.019, i64 0, i32 0, i64 %indvars.iv, i32 0
|
|
%0 = load ptr, ptr %image4, align 8
|
|
%cmp5 = icmp eq ptr %0, %image
|
|
br i1 %cmp5, label %if.then, label %for.inc
|
|
|
|
if.then: ; preds = %for.body3
|
|
tail call void @OSMemoryBarrier() optsize
|
|
store ptr null, ptr %image4, align 8
|
|
br label %for.inc
|
|
|
|
for.inc: ; preds = %if.then, %for.body3
|
|
%indvars.iv.next = add i64 %indvars.iv, 1
|
|
%lftr.wideiv = trunc i64 %indvars.iv.next to i32
|
|
%exitcond = icmp eq i32 %lftr.wideiv, 400
|
|
br i1 %exitcond, label %for.inc10, label %for.body3
|
|
|
|
for.inc10: ; preds = %for.inc
|
|
%next = getelementptr inbounds %"struct.dyld::MappedRanges", ptr %p.019, i64 0, i32 1
|
|
%1 = load ptr, ptr %next, align 8
|
|
%cmp = icmp eq ptr %1, null
|
|
br i1 %cmp, label %for.end11, label %for.cond1.preheader
|
|
|
|
for.end11: ; preds = %for.inc10
|
|
ret void
|
|
}
|