https://github.com/hvdijk created
https://github.com/llvm/llvm-project/pull/145698
For a kernel such as
kernel void foo(__global double3 *z) {
double3 x = {0.6631661088,0.6612268107,0.1513627528};
int3 y = {-1980459213,-660855407,615708204};
*z = pown(x, y);
}
we were not storing anything to z, because the implementation of pown relied on
an floating-point-to-integer conversion where the floating-point value was
outside of the integer's range. Although in LLVM IR we permit that operation so
long as we end up ignoring its result -- that is the general rule for poison --
one thing we are not permitted to do is have conditional branches that depend
on it, and through the call to __clc_ldexp, we did have that.
To fix this, rather than changing expv at the end to INFINITY/0, we can change
v at the start to values that we know will produce INFINITY/0 without
performing such out-of-range conversions.
Tested with
clang --target=nvptx64 -S -O3 -o - test.cl \
-Xclang -mlink-builtin-bitcode \
-Xclang runtimes/runtimes-bins/libclc/nvptx64--.bc
A grep showed that this exact same code existed in three more places, so I
changed it there too, though I did not do a broader search for other similar
code that potentially has the same problem.
>From 6e37101f93b285d6f0c5313d4367089daeb7097f Mon Sep 17 00:00:00 2001
From: Harald van Dijk <harald.vand...@codeplay.com>
Date: Wed, 25 Jun 2025 14:22:15 +0100
Subject: [PATCH] [libclc] Avoid out-of-range float-to-int.
For a kernel such as
kernel void foo(__global double3 *z) {
double3 x = {0.6631661088,0.6612268107,0.1513627528};
int3 y = {-1980459213,-660855407,615708204};
*z = pown(x, y);
}
we were not storing anything to z, because the implementation of pown
relied on an floating-point-to-integer conversion where the
floating-point value was outside of the integer's range. Although in
LLVM IR we permit that operation so long as we end up ignoring its
result -- that is the general rule for poison -- one thing we are not
permitted to do is have conditional branches that depend on it, and
through the call to __clc_ldexp, we did have that.
To fix this, rather than changing expv at the end to INFINITY/0, we can
change v at the start to values that we know will produce INFINITY/0
without performing such out-of-range conversions.
Tested with
clang --target=nvptx64 -S -O3 -o - test.cl \
-Xclang -mlink-builtin-bitcode \
-Xclang runtimes/runtimes-bins/libclc/nvptx64--.bc
A grep showed that this exact same code existed in three more places, so
I changed it there too, though I did not do a broader search for other
similar code that potentially has the same problem.
---
libclc/clc/lib/generic/math/clc_pow.inc | 13 +++++++++----
libclc/clc/lib/generic/math/clc_pown.inc | 13 +++++++++----
libclc/clc/lib/generic/math/clc_powr.inc | 13 +++++++++----
libclc/clc/lib/generic/math/clc_rootn.inc | 13 +++++++++----
4 files changed, 36 insertions(+), 16 deletions(-)
diff --git a/libclc/clc/lib/generic/math/clc_pow.inc
b/libclc/clc/lib/generic/math/clc_pow.inc
index 98e154984aaa3..8b1f820268ba0 100644
--- a/libclc/clc/lib/generic/math/clc_pow.inc
+++ b/libclc/clc/lib/generic/math/clc_pow.inc
@@ -330,6 +330,15 @@ _CLC_DEF _CLC_OVERLOAD __CLC_GENTYPE
__clc_pow(__CLC_GENTYPE x,
const __CLC_GENTYPE lnof2_by_64_head = 0.010830424260348081;
const __CLC_GENTYPE lnof2_by_64_tail = -4.359010638708991e-10;
+ // If v is so large that we need to return INFINITY, or so small that we
+ // need to return 0, set v to known values that will produce that result.
Do
+ // not try to continue the computation with the original v and patch it up
+ // afterwards because v may be so large that temp is out of range of int,
in
+ // which case that conversion, and a value based on that conversion being
+ // passed to __clc_ldexp, results in undefined behavior.
+ v = v > max_exp_arg ? 1000.0 : v;
+ v = v < min_exp_arg ? -1000.0 : v;
+
__CLC_GENTYPE temp = v * sixtyfour_by_lnof2;
__CLC_INTN n = __CLC_CONVERT_INTN(temp);
__CLC_GENTYPE dn = __CLC_CONVERT_GENTYPE(n);
@@ -357,10 +366,6 @@ _CLC_DEF _CLC_OVERLOAD __CLC_GENTYPE
__clc_pow(__CLC_GENTYPE x,
expv = __clc_fma(f, q, f2) + f1;
expv = __clc_ldexp(expv, m);
-
- expv = v > max_exp_arg ?
__CLC_AS_GENTYPE((__CLC_ULONGN)0x7FF0000000000000L)
- : expv;
- expv = v < min_exp_arg ? 0.0 : expv;
}
// See whether y is an integer.
diff --git a/libclc/clc/lib/generic/math/clc_pown.inc
b/libclc/clc/lib/generic/math/clc_pown.inc
index 8bdc407e9ac82..483fd2faf2717 100644
--- a/libclc/clc/lib/generic/math/clc_pown.inc
+++ b/libclc/clc/lib/generic/math/clc_pown.inc
@@ -317,6 +317,15 @@ _CLC_DEF _CLC_OVERLOAD __CLC_GENTYPE
__clc_pown(__CLC_GENTYPE x,
const __CLC_GENTYPE lnof2_by_64_head = 0.010830424260348081;
const __CLC_GENTYPE lnof2_by_64_tail = -4.359010638708991e-10;
+ // If v is so large that we need to return INFINITY, or so small that we
+ // need to return 0, set v to known values that will produce that result.
Do
+ // not try to continue the computation with the original v and patch it up
+ // afterwards because v may be so large that temp is out of range of int,
in
+ // which case that conversion, and a value based on that conversion being
+ // passed to __clc_ldexp, results in undefined behavior.
+ v = v > max_exp_arg ? 1000.0 : v;
+ v = v < min_exp_arg ? -1000.0 : v;
+
__CLC_GENTYPE temp = v * sixtyfour_by_lnof2;
__CLC_INTN n = __CLC_CONVERT_INTN(temp);
__CLC_GENTYPE dn = __CLC_CONVERT_GENTYPE(n);
@@ -344,10 +353,6 @@ _CLC_DEF _CLC_OVERLOAD __CLC_GENTYPE
__clc_pown(__CLC_GENTYPE x,
expv = __clc_fma(f, q, f2) + f1;
expv = __clc_ldexp(expv, m);
-
- expv = v > max_exp_arg ?
__CLC_AS_GENTYPE((__CLC_ULONGN)0x7FF0000000000000L)
- : expv;
- expv = v < min_exp_arg ? 0.0 : expv;
}
// See whether y is an integer.
diff --git a/libclc/clc/lib/generic/math/clc_powr.inc
b/libclc/clc/lib/generic/math/clc_powr.inc
index fbdf3d85de2b7..1244f7f6ac5d6 100644
--- a/libclc/clc/lib/generic/math/clc_powr.inc
+++ b/libclc/clc/lib/generic/math/clc_powr.inc
@@ -316,6 +316,15 @@ _CLC_DEF _CLC_OVERLOAD __CLC_GENTYPE
__clc_powr(__CLC_GENTYPE x,
const __CLC_GENTYPE lnof2_by_64_head = 0.010830424260348081;
const __CLC_GENTYPE lnof2_by_64_tail = -4.359010638708991e-10;
+ // If v is so large that we need to return INFINITY, or so small that we
+ // need to return 0, set v to known values that will produce that result.
Do
+ // not try to continue the computation with the original v and patch it up
+ // afterwards because v may be so large that temp is out of range of int,
in
+ // which case that conversion, and a value based on that conversion being
+ // passed to __clc_ldexp, results in undefined behavior.
+ v = v > max_exp_arg ? 1000.0 : v;
+ v = v < min_exp_arg ? -1000.0 : v;
+
__CLC_GENTYPE temp = v * sixtyfour_by_lnof2;
__CLC_INTN n = __CLC_CONVERT_INTN(temp);
__CLC_GENTYPE dn = __CLC_CONVERT_GENTYPE(n);
@@ -343,10 +352,6 @@ _CLC_DEF _CLC_OVERLOAD __CLC_GENTYPE
__clc_powr(__CLC_GENTYPE x,
expv = __clc_fma(f, q, f2) + f1;
expv = __clc_ldexp(expv, m);
-
- expv = v > max_exp_arg ?
__CLC_AS_GENTYPE((__CLC_ULONGN)0x7FF0000000000000L)
- : expv;
- expv = v < min_exp_arg ? 0.0 : expv;
}
// See whether y is an integer.
diff --git a/libclc/clc/lib/generic/math/clc_rootn.inc
b/libclc/clc/lib/generic/math/clc_rootn.inc
index 0c459ae5c3cbb..996f88f145357 100644
--- a/libclc/clc/lib/generic/math/clc_rootn.inc
+++ b/libclc/clc/lib/generic/math/clc_rootn.inc
@@ -323,6 +323,15 @@ _CLC_DEF _CLC_OVERLOAD __CLC_GENTYPE
__clc_rootn(__CLC_GENTYPE x,
const __CLC_GENTYPE lnof2_by_64_head = 0.010830424260348081;
const __CLC_GENTYPE lnof2_by_64_tail = -4.359010638708991e-10;
+ // If v is so large that we need to return INFINITY, or so small that we
+ // need to return 0, set v to known values that will produce that result.
Do
+ // not try to continue the computation with the original v and patch it up
+ // afterwards because v may be so large that temp is out of range of int,
in
+ // which case that conversion, and a value based on that conversion being
+ // passed to __clc_ldexp, results in undefined behavior.
+ v = v > max_exp_arg ? 1000.0 : v;
+ v = v < min_exp_arg ? -1000.0 : v;
+
__CLC_GENTYPE temp = v * sixtyfour_by_lnof2;
__CLC_INTN n = __CLC_CONVERT_INTN(temp);
__CLC_GENTYPE dn = __CLC_CONVERT_GENTYPE(n);
@@ -350,10 +359,6 @@ _CLC_DEF _CLC_OVERLOAD __CLC_GENTYPE
__clc_rootn(__CLC_GENTYPE x,
expv = __clc_fma(f, q, f2) + f1;
expv = __clc_ldexp(expv, m);
-
- expv = v > max_exp_arg ?
__CLC_AS_GENTYPE((__CLC_ULONGN)0x7FF0000000000000L)
- : expv;
- expv = v < min_exp_arg ? 0.0 : expv;
}
// See whether y is an integer.
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
