| Message ID | 1389445524-30623-3-git-send-email-leif.lindholm@linaro.org |
|---|---|
| State | New |
| Headers | show |
On Wed, Jan 22, 2014 at 11:20:55AM +0000, Will Deacon wrote: > > +#ifdef CONFIG_CPU_CP15 > > +/* Macro for setting/clearing bits in sctlr */ > > + .macro update_sctlr, set:req, clear:req, tmp:req, tmp2:req > > + mrc p15, 0, \tmp, c1, c0, 0 > > + ldr \tmp2, =\set > > + orr \tmp, \tmp, \tmp2 > > + ldr \tmp2, =\clear > > + mvn \tmp2, \tmp2 > > + and \tmp, \tmp, \tmp2 > > + mcr p15, 0, \tmp, c1, c0, 0 > > I think this would be cleaner if you force the caller to put set and clear > into registers beforehand, rather than have to do the literal load every > time. Also, I don't think set and clear should be required (and then you can > lose tmp2 as well). I can't figure out how to make register-parameters non-required (i.e. conditionalise on whether an optional parameter was provided), so my attempt of refactoring actually ends up using an additional register: #ifdef CONFIG_CPU_CP15 /* Macro for setting/clearing bits in sctlr */ .macro update_sctlr, set:req, clear:req, tmp:req mrc p15, 0, \tmp, c1, c0, 0 orr \tmp, \set mvn \clear, \clear and \tmp, \tmp, \clear mcr p15, 0, \tmp, c1, c0, 0 .endm #endif If you think that's an improvement I can do that, and I have (just) enough registers to spare. If I'm being daft with my macro issues, do point it out. / Leif
On Wed, 2014-01-29 at 18:28 +0000, Leif Lindholm wrote: > On Wed, Jan 22, 2014 at 11:20:55AM +0000, Will Deacon wrote: > > > +#ifdef CONFIG_CPU_CP15 > > > +/* Macro for setting/clearing bits in sctlr */ > > > + .macro update_sctlr, set:req, clear:req, tmp:req, tmp2:req > > > + mrc p15, 0, \tmp, c1, c0, 0 > > > + ldr \tmp2, =\set > > > + orr \tmp, \tmp, \tmp2 > > > + ldr \tmp2, =\clear > > > + mvn \tmp2, \tmp2 > > > + and \tmp, \tmp, \tmp2 > > > + mcr p15, 0, \tmp, c1, c0, 0 > > > > I think this would be cleaner if you force the caller to put set and clear > > into registers beforehand, rather than have to do the literal load every > > time. Also, I don't think set and clear should be required (and then you can > > lose tmp2 as well). > > I can't figure out how to make register-parameters non-required > (i.e. conditionalise on whether an optional parameter was provided), > so my attempt of refactoring actually ends up using an additional > register: > Register parameters are just strings, so how about this: .macro foo bar=, baz= .ifnc \bar, mov \bar,#0 .endif .ifnc \baz, mov \baz,#1 .endif .endm foo x0 foo foo x1, x2 foo ,x3 Results in: 0000000000000000 <.text>: 0: d2800000 mov x0, #0x0 // #0 4: d2800001 mov x1, #0x0 // #0 8: d2800022 mov x2, #0x1 // #1 c: d2800023 mov x3, #0x1 // #1
On Wed, Jan 29, 2014 at 03:58:44PM -0500, Mark Salter wrote: > > (i.e. conditionalise on whether an optional parameter was provided), > > so my attempt of refactoring actually ends up using an additional > > register: > > > > Register parameters are just strings, so how about this: > > .macro foo bar=, baz= > .ifnc \bar, > mov \bar,#0 > .endif > .ifnc \baz, > mov \baz,#1 > .endif > .endm > > foo x0 > foo > foo x1, x2 > foo ,x3 > > Results in: > > 0000000000000000 <.text>: > 0: d2800000 mov x0, #0x0 // #0 > 4: d2800001 mov x1, #0x0 // #0 > 8: d2800022 mov x2, #0x1 // #1 > c: d2800023 mov x3, #0x1 // #1 Oh, that's neat - thanks! Well, given that, I can think of two less horrible options: 1) .macro update_sctlr, tmp:req, set=, clear= mrc p15, 0, \tmp, c1, c0, 0 .ifnc \set, orr \tmp, \set .endif .ifnc \clear, mvn \clear, \clear and \tmp, \tmp, \clear .endif mcr p15, 0, \tmp, c1, c0, 0 .endm With the two call sites in uefi_phys.S as: ldr r5, =(CR_M) update_sctlr r12, , r5 and ldr r4, =(CR_I | CR_C | CR_M) update_sctlr r12, r4 Which disassembles as: 2c: e3a05001 mov r5, #1 30: ee11cf10 mrc 15, 0, ip, cr1, cr0, {0} 34: e1e05005 mvn r5, r5 38: e00cc005 and ip, ip, r5 3c: ee01cf10 mcr 15, 0, ip, cr1, cr0, {0} and 48: e59f4034 ldr r4, [pc, #52] ; 84 <tmpstack+0x4> 4c: ee11cf10 mrc 15, 0, ip, cr1, cr0, {0} 50: e18cc004 orr ip, ip, r4 54: ee01cf10 mcr 15, 0, ip, cr1, cr0, {0} 2) .macro update_sctlr, tmp:req, tmp2:req, set=, clear= mrc p15, 0, \tmp, c1, c0, 0 .ifnc \set, ldr \tmp2, =\set orr \tmp, \tmp, \tmp2 .endif .ifnc \clear, ldr \tmp2, =\clear mvn \tmp2, \tmp2 and \tmp, \tmp, \tmp2 .endif mcr p15, 0, \tmp, c1, c0, 0 .endm With the two call sites in uefi_phys.S as: update_sctlr r4, r5, , (CR_M) and update_sctlr r4, r5, (CR_I | CR_C | CR_M) Which disassembles as: 2c: ee114f10 mrc 15, 0, r4, cr1, cr0, {0} 30: e3a05001 mov r5, #1 34: e1e05005 mvn r5, r5 38: e0044005 and r4, r4, r5 3c: ee014f10 mcr 15, 0, r4, cr1, cr0, {0} and 48: ee114f10 mrc 15, 0, r4, cr1, cr0, {0} 4c: e59f5030 ldr r5, [pc, #48] ; 84 <tmpstack+0x4> 50: e1844005 orr r4, r4, r5 54: ee014f10 mcr 15, 0, r4, cr1, cr0, {0} The benefit of 2) is a cleaner call site, and one fewer register used if setting and clearing simultaneously. The benefit of 1) is that the macro could then easily be used with the crval mask in mm/proc*.S So, Will, which one do you want? / Leif
On Mon, Feb 03, 2014 at 10:34:15AM +0000, Will Deacon wrote: > On Thu, Jan 30, 2014 at 01:12:47PM +0000, Leif Lindholm wrote: > > Oh, that's neat - thanks! > > > > Well, given that, I can think of two less horrible options: > > 1) > > .macro update_sctlr, tmp:req, set=, clear= > > mrc p15, 0, \tmp, c1, c0, 0 > > .ifnc \set, > > orr \tmp, \set > > .endif > > .ifnc \clear, > > mvn \clear, \clear > > and \tmp, \tmp, \clear > > Can't you use bic here? Yeah. > > .endif > > mcr p15, 0, \tmp, c1, c0, 0 > > .endm > > > > With the two call sites in uefi_phys.S as: > > > > ldr r5, =(CR_M) > > update_sctlr r12, , r5 > > and > > ldr r4, =(CR_I | CR_C | CR_M) > > update_sctlr r12, r4 > > These ldr= could be movs, right? The first one could. The second one could be movw on armv7+. > If so, I definitely prefer this to putting an ldr = into the macro itself > (option 2). And your preference between 1) and 2) is? / Leif
On Mon, Feb 03, 2014 at 04:00:51PM +0000, Will Deacon wrote: > > > > With the two call sites in uefi_phys.S as: > > > > > > > > ldr r5, =(CR_M) > > > > update_sctlr r12, , r5 > > > > and > > > > ldr r4, =(CR_I | CR_C | CR_M) > > > > update_sctlr r12, r4 > > > > > > These ldr= could be movs, right? > > > > The first one could. > > The second one could be movw on armv7+. > > > > > If so, I definitely prefer this to putting an ldr = into the macro itself > > > (option 2). > > > > And your preference between 1) and 2) is? > > (1), using bic and mov[tw] where possible. (1): ok, thanks. bic: sure, that was an oversight. mov[tw]: why? Then we end up battling different available immediate fields in A32/T32 instruction sets and v5/v6/v7 architecture versions. / Leif
On Mon, Feb 03, 2014 at 04:57:18PM +0000, Will Deacon wrote: > > mov[tw]: why? > > Then we end up battling different available immediate fields in A32/T32 > > instruction sets and v5/v6/v7 architecture versions. > > I was making the assumption that UEFI was going to be v7 only... is this not > true? There is no such requirement in the specification. It even mentions requirements for ARMv4 in one place :) But I also don't understand why ldr= should be avoided. This is not performance sensitive (called once on system boot), and it's already executing with the caches off, so even if it ends up being a literal load it does not pollute. / Leif
diff --git a/arch/arm/include/asm/assembler.h b/arch/arm/include/asm/assembler.h index 5c22851..aba6458 100644 --- a/arch/arm/include/asm/assembler.h +++ b/arch/arm/include/asm/assembler.h @@ -383,4 +383,17 @@ THUMB( orr \reg , \reg , #PSR_T_BIT ) #endif .endm +#ifdef CONFIG_CPU_CP15 +/* Macro for setting/clearing bits in sctlr */ + .macro update_sctlr, set:req, clear:req, tmp:req, tmp2:req + mrc p15, 0, \tmp, c1, c0, 0 + ldr \tmp2, =\set + orr \tmp, \tmp, \tmp2 + ldr \tmp2, =\clear + mvn \tmp2, \tmp2 + and \tmp, \tmp, \tmp2 + mcr p15, 0, \tmp, c1, c0, 0 + .endm +#endif + #endif /* __ASM_ASSEMBLER_H__ */
A new macro for setting/clearing bits in the SCTLR. Signed-off-by: Leif Lindholm <leif.lindholm@linaro.org> Suggested-by: Will Deacon <will.deacon@arm.com> --- arch/arm/include/asm/assembler.h | 13 +++++++++++++ 1 file changed, 13 insertions(+)