ARM bad code: constants materialization inside loops

[llvmbot]

Bugzilla Link	1298
Resolution	DUPLICATE
Resolved on	Feb 22, 2010 12:44
Version	unspecified
OS	All
Reporter	LLVM Bugzilla Contributor

Extended Description

code generated for sieve test:

    mov r2, #​0
    ldr r1, .LCPI1_0

.LBB1_3: @​bb7
add r0, r1, r2
mov r12, #​1
strb r12, [r0, #+2]
add r2, r2, #​1
mov r0, #​255
orr r0, r0, #​31, 24 @ 7936
cmp r2, r0
bne .LBB1_3 @​bb7

can be:

    mov r2, #​0
    ldr r1, .LCPI1_0
    mov r12, #​1
    mov r0, #​255
    orr r0, r0, #​31, 24 @ 7936

.LBB1_3: @​bb7
add r3, r1, r2
strb r12, [r3, #+2]
add r2, r2, #​1
cmp r2, r0
bne .LBB1_3 @​bb7

[lattner]

Absolutely. We need to either do whole-function selection dags (with a code layout approach similar to
the one described in cliff click's thesis) or we need to have a machine code LICM pass.

-Chris

*** This bug has been marked as a duplicate of 893 ***

[lattner]

mentioned in issue #1265