Re: [musl] Q: dealing with missing removal of excess precision

[Rich Felker <dalias@libc.org>]

[-- Attachment #1: Type: text/plain, Size: 2760 bytes --]

On Fri, Mar 20, 2020 at 02:12:50PM -0400, Rich Felker wrote:
> On Sun, Feb 23, 2020 at 07:14:08PM +0300, Alexander Monakov wrote:
> > On Sat, 22 Feb 2020, Rich Felker wrote:
> > 
> > > I'm not well acquainted with SSE, and only so-so with x87, so pretty
> > > much I'm reading them for higher-level issues with tooling
> > > compatibility (like the concerns I already raised and looked up and
> > > seem to have resolved about x87 constraints and non-GCC compilers) and
> > > logic, then planning to apply and test them. I think being aware of
> > > non-obvious mistake modes that have already been found would be a lot
> > > more useful than staring at things, especially if the bugs you've
> > > found are in subtleties of the insn behavior or constraint behavior.
> > 
> > Okay, thanks. I found two issues:
> > 
> > 1. i386 lrint* functions mistakenly used fistpll instead of fistpl (I
> > posted the fix for x32 asm after noticing my own mistake).
> > 
> > 2. Some functions bind a 32-bit lvalue as output for fnstsw %ax, which
> > as the operand says writes only 16 bits. They should be changed to either
> > use a 16-bit lvalue, or a zero-initialized 32-bit lvalue with "+a" constraint.
> > 
> > Plus, not bugs, but still worth mentioning:
> > 
> > 3. The new remquol in C could alternatively be implemented by using fxam
> > to extract sign bits instead of loading them from stack slots. The current
> > approach makes sense given the ABI, but an implementation aiming for better
> > code after inlining could choose to use fxam instead of forcing a spill.
> > 
> > 4. I did not manage to find a copy of Figueroa's "When is double rounding
> > innocuous", but I could cite e.g. "Innocuous Double Rounding of Basic
> > Arithmetic Operations" by Pierre Roux instead (in i386/sqrtf.c).
> > 
> > If you like you can fetch a Git tree with my patches from 
> > 
> >     https://git.sr.ht/~amonakov/musl
> > 
> > (issue 1 is already corrected in that repo)
> 
> Hi. I'm trying to catch up on this and other patches after being sick
> and not getting much done for a while. Is there anything else I should
> be aware of before going forward with these?
> 
> One minor cosmetic change I'd like to make to commit names if you
> don't object is changing "x86-family" to "x87-family" for the commits
> that are "i386 + long double functions on x86_64" since I found it
> confusing when there's one commit for "x86 family" then a separate one
> for x86_64 versions of the same function. Let me know if you think of
> any alternative that might be more clear than "x87"

With the fixups mentioned (included in attached patches), i386 and
x86_64 are passing libc-test and seem fine. OK if I merge them
(rebasing the fixups in as fixups)? With s/x86/x87/ naming?

Rich

[-- Attachment #2: 0001-fabsf-i386-fixup.patch --]
[-- Type: text/plain, Size: 552 bytes --]

From 3f99f4595aa3ce528008ee22d65512768376a9a6 Mon Sep 17 00:00:00 2001
From: Rich Felker <dalias@aerifal.cx>
Date: Sat, 21 Mar 2020 16:23:51 -0400
Subject: [PATCH 1/3] fabsf i386 fixup

---
 src/math/i386/fabsf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/math/i386/fabsf.c b/src/math/i386/fabsf.c
index d07be321..d882eee3 100644
--- a/src/math/i386/fabsf.c
+++ b/src/math/i386/fabsf.c
@@ -1,6 +1,6 @@
 #include <math.h>
 
-float fabs(float x)
+float fabsf(float x)
 {
 	__asm__ ("fabs" : "+t"(x));
 	return x;
-- 
2.21.0


[-- Attachment #3: 0002-sqrt-i386-fixup.patch --]
[-- Type: text/plain, Size: 747 bytes --]

From 7fd1bc3b8ecf61d758498edd7f4a75a079ee8df5 Mon Sep 17 00:00:00 2001
From: Rich Felker <dalias@aerifal.cx>
Date: Sat, 21 Mar 2020 16:24:04 -0400
Subject: [PATCH 2/3] sqrt i386 fixup

---
 src/math/i386/sqrt.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/math/i386/sqrt.c b/src/math/i386/sqrt.c
index 619df056..934fbcca 100644
--- a/src/math/i386/sqrt.c
+++ b/src/math/i386/sqrt.c
@@ -10,6 +10,6 @@ double sqrt(double x)
 	/* Rounding to double would have encountered an exact halfway case.
 	   Adjust mantissa downwards if fsqrt rounded up, else upwards.
 	   (result of fsqrt could not have been exact) */
-	ux.i.m ^= (fpsr & 0x200) + 0x200;
+	ux.i.m ^= (fpsr & 0x200) + 0x300;
 	return (double)ux.f;
 }
-- 
2.21.0


[-- Attachment #4: 0003-sqrt-i386-fixup-2.patch --]
[-- Type: text/plain, Size: 649 bytes --]

From b60ddebd251ff28559b48aa758026da8fbbab3f7 Mon Sep 17 00:00:00 2001
From: Rich Felker <dalias@aerifal.cx>
Date: Sat, 21 Mar 2020 16:57:38 -0400
Subject: [PATCH 3/3] sqrt i386 fixup 2

---
 src/math/i386/sqrt.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/math/i386/sqrt.c b/src/math/i386/sqrt.c
index 934fbcca..a34f4ce8 100644
--- a/src/math/i386/sqrt.c
+++ b/src/math/i386/sqrt.c
@@ -3,7 +3,7 @@
 double sqrt(double x)
 {
 	union ldshape ux;
-	unsigned fpsr;
+	unsigned short fpsr;
 	__asm__ ("fsqrt; fnstsw %%ax": "=t"(ux.f), "=a"(fpsr) : "0"(x));
 	if ((ux.i.m & 0x7ff) != 0x400)
 		return (double)ux.f;
-- 
2.21.0