LLVM immediate offset missed optimization for ARM thumb register access

output

Ubuntu clang version 3.5.0-4ubuntu2 (tags/RELEASE_350/final) (based on LLVM 3.5.0)
00008000 <main>:
    8000:	4802      	ldr	r0, [pc, #8]	; (800c <main+0xc>)
    8002:	2101      	movs	r1, #1
    8004:	6001      	str	r1, [r0, #0]
    8006:	4802      	ldr	r0, [pc, #8]	; (8010 <main+0x10>)
    8008:	6001      	str	r1, [r0, #0]
    800a:	4770      	bx	lr
    800c:	10001004 	.word	0x10001004
    8010:	10001008 	.word	0x10001008


gcc version 4.8.3 20140228 (release) [ARM/embedded-4_8-branch revision 208322] (GNU Tools for ARM Embedded Processors)
00008000 <main>:
    8000:	4b02      	ldr	r3, [pc, #8]	; (800c <main+0xc>)
    8002:	2201      	movs	r2, #1
    8004:	605a      	str	r2, [r3, #4]
    8006:	609a      	str	r2, [r3, #8]
    8008:	4770      	bx	lr
    800a:	46c0      	nop			; (mov r8, r8)
    800c:	10001000 	.word	0x10001000

run.sh

clang-3.5 -v 2>&1 | head -n1
clang-3.5 --target=arm-none-eabi -O3 test.c -ffreestanding -nostdlib -mthumb -o test-clang
arm-none-eabi-objdump -S test-clang

arm-none-eabi-gcc -v 2>&1 | tail -n 1
arm-none-eabi-gcc -O3 -ffreestanding -nostdlib test.c -mthumb -o test-gcc
arm-none-eabi-objdump -S test-gcc

test.c

typedef struct foo {
	volatile unsigned a;
	volatile unsigned b;
	volatile unsigned c;
} foo;

void main() {
	foo* f = (foo*) 0x10001000;
	f->b = 1;
	f->c = 1;
}

Seems to happen when LLVM constant-folds the GEP instruction. This patch for x86 seems relevant:
http://llvm.org/viewvc/llvm-project?view=revision&revision=204739

I'm using the workaround of putting the register base addresses in the linker script, which hides them from LLVM.