Emit armv8.4 ldapur* for volatile loads with contained offsets

[EgorBo]

Contributes to #64457

volatile int a;
volatile int b;
volatile int c;
volatile int d;

[Benchmark]
public int Test()
{
    return a + b + c + d;
}

Current codegen for Test:

; Assembly listing for method P:Test():int:this (Tier1)		
            stp     fp, lr, [sp, #-0x10]!		
            mov     fp, sp		
            add     x1, x0, #8		
            ldapr   w1, [x1]		
            add     x2, x0, #12		
            ldapr   w2, [x2]		
            add     w1, w1, w2		
            add     x2, x0, #16		
            ldapr   w2, [x2]		
            add     w1, w1, w2		
            add     x0, x0, #20		
            ldapr   w0, [x0]		
            add     w0, w1, w0		
            ldp     fp, lr, [sp], #0x10		
            ret     lr		
; Total bytes of code 60

New codegen for Test:

; Assembly listing for method P:Test():int:this (Tier1)
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
            ldapur  w1, [x0, #0x08]
            ldapur  w2, [x0, #0x0C]
            add     w1, w1, w2
            ldapur  w2, [x0, #0x10]
            add     w1, w1, w2
            ldapur  w0, [x0, #0x14]
            add     w0, w1, w0
            ldp     fp, lr, [sp], #0x10
            ret     lr
; Total bytes of code 44

Apple M2 Max:

Method	Job	Toolchain	Mean	Error	StdDev	Ratio
Test	Job-LOXHRO	/Core_Root_PR/corerun	0.4724 ns	0.0050 ns	0.0045 ns	0.89
Test	Job-MPJUHL	/Core_Root_main/corerun	0.5327 ns	0.0063 ns	0.0059 ns	1.00

Perf difference is not super exciting but it is what it is 🙂

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

volatile int a;
volatile int b;
volatile int c;
volatile int d;

[Benchmark]
public int Test()
{
    return a + b + c + d;
}

Current codegen for Test:

; Assembly listing for method P:Test():int:this (Tier1)		
            stp     fp, lr, [sp, #-0x10]!		
            mov     fp, sp		
            add     x1, x0, #8		
            ldapr   w1, [x1]		
            add     x2, x0, #12		
            ldapr   w2, [x2]		
            add     w1, w1, w2		
            add     x2, x0, #16		
            ldapr   w2, [x2]		
            add     w1, w1, w2		
            add     x0, x0, #20		
            ldapr   w0, [x0]		
            add     w0, w1, w0		
            ldp     fp, lr, [sp], #0x10		
            ret     lr		
; Total bytes of code 60

New codegen for Test:

; Assembly listing for method P:Test():int:this (Tier1)
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
            ldapur  w1, [x0, #0x08]
            ldapur  w2, [x0, #0x0C]
            add     w1, w1, w2
            ldapur  w2, [x0, #0x10]
            add     w1, w1, w2
            ldapur  w0, [x0, #0x14]
            add     w0, w1, w0
            ldp     fp, lr, [sp], #0x10
            ret     lr
; Total bytes of code 44

Method	Job	Toolchain	Mean	Error	StdDev	Ratio
Test	Job-LOXHRO	/Core_Root_PR/corerun	0.4724 ns	0.0050 ns	0.0045 ns	0.89
Test	Job-MPJUHL	/Core_Root_main/corerun	0.5327 ns	0.0063 ns	0.0059 ns	1.00

Perf difference is not super exciting but it is what it is 🙂

Author:	EgorBo
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[EgorBo]

I added a bit here since we now have more than 512 instructions (and that for sure will be increasing in future) but I also removed an used (on arm) _idCallRegPtr bit so the total size of the struct remains the same - 16b

[EgorBo]

@TIHan @dotnet/jit-contrib PTAL, simple change, emits ldapur where we previously emitted ldapr+add because ldapr doesn't support IMM offsets

[TIHan]

Why was this removed for arm32?

[EgorBo]

I mentioned it in https://github.com/dotnet/runtime/pull/89681/files#r1278410873 comment, basically, it sets a bit that is unused for arm, so I removed that bit because I needed to extend _idIns's width

[TIHan]

LGTM, only one comment on the idSetIsCallingRegPtr.