Skip to content

Commit 27e5f50

Browse files
author
Andy Kaylor
authored
[Driver] Make ffp-model=fast honor non-finite-values, introduce ffp-model=aggressive (#100453)
This change modifies -ffp-model=fast to select options that more closely match -funsafe-math-optimizations, and introduces a new model, -ffp-model=aggressive which matches the existing behavior (except for a minor change in the fp-contract behavior). The primary motivation for this change is to make -ffp-model=fast more user friendly, particularly in light of LLVM's aggressive optimizations when -fno-honor-nans and -fno-honor-infinites are used. This was previously proposed here: https://discourse.llvm.org/t/making-ffp-model-fast-more-user-friendly/78402
1 parent fd83b86 commit 27e5f50

File tree

6 files changed

+185
-97
lines changed

6 files changed

+185
-97
lines changed

‎clang/docs/ReleaseNotes.rst

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -176,6 +176,14 @@ Modified Compiler Flags
176176

177177
- The compiler flag `-fbracket-depth` default value is increased from 256 to 2048.
178178

179+
- The ``-ffp-model`` option has been updated to enable a more limited set of
180+
optimizations when the ``fast`` argument is used and to accept a new argument,
181+
``aggressive``. The behavior of ``-ffp-model=aggressive`` is equivalent
182+
to the previous behavior of ``-ffp-model=fast``. The updated
183+
``-ffp-model=fast`` behavior no longer assumes finite math only and uses
184+
the ``promoted`` algorithm for complex division when possible rather than the
185+
less basic (limited range) algorithm.
186+
179187
Removed Compiler Flags
180188
-------------------------
181189

‎clang/docs/UsersManual.rst

Lines changed: 25 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -1452,28 +1452,30 @@ describes the various floating point semantic modes and the corresponding option
14521452
"fhonor-infinities", "{on, off}"
14531453
"fsigned-zeros", "{on, off}"
14541454
"freciprocal-math", "{on, off}"
1455-
"allow_approximate_fns", "{on, off}"
1455+
"fallow-approximate-fns", "{on, off}"
14561456
"fassociative-math", "{on, off}"
1457+
"fcomplex-arithmetic", "{basic, improved, full, promoted}"
14571458

14581459
This table describes the option settings that correspond to the three
14591460
floating point semantic models: precise (the default), strict, and fast.
14601461

14611462

14621463
.. csv-table:: Floating Point Models
1463-
:header: "Mode", "Precise", "Strict", "Fast"
1464-
:widths: 25, 15, 15, 15
1465-
1466-
"except_behavior", "ignore", "strict", "ignore"
1467-
"fenv_access", "off", "on", "off"
1468-
"rounding_mode", "tonearest", "dynamic", "tonearest"
1469-
"contract", "on", "off", "fast"
1470-
"support_math_errno", "on", "on", "off"
1471-
"no_honor_nans", "off", "off", "on"
1472-
"no_honor_infinities", "off", "off", "on"
1473-
"no_signed_zeros", "off", "off", "on"
1474-
"allow_reciprocal", "off", "off", "on"
1475-
"allow_approximate_fns", "off", "off", "on"
1476-
"allow_reassociation", "off", "off", "on"
1464+
:header: "Mode", "Precise", "Strict", "Fast", "Aggressive"
1465+
:widths: 25, 25, 25, 25, 25
1466+
1467+
"except_behavior", "ignore", "strict", "ignore", "ignore"
1468+
"fenv_access", "off", "on", "off", "off"
1469+
"rounding_mode", "tonearest", "dynamic", "tonearest", "tonearest"
1470+
"contract", "on", "off", "fast", "fast"
1471+
"support_math_errno", "on", "on", "off", "off"
1472+
"no_honor_nans", "off", "off", "off", "on"
1473+
"no_honor_infinities", "off", "off", "off", "on"
1474+
"no_signed_zeros", "off", "off", "on", "on"
1475+
"allow_reciprocal", "off", "off", "on", "on"
1476+
"allow_approximate_fns", "off", "off", "on", "on"
1477+
"allow_reassociation", "off", "off", "on", "on"
1478+
"complex_arithmetic", "full", "full", "promoted", "basic"
14771479

14781480
The ``-ffp-model`` option does not modify the ``fdenormal-fp-math``
14791481
setting, but it does have an impact on whether ``crtfastmath.o`` is
@@ -1492,9 +1494,9 @@ for more details.
14921494
* Floating-point math obeys regular algebraic rules for real numbers (e.g.
14931495
``+`` and ``*`` are associative, ``x/y == x * (1/y)``, and
14941496
``(a + b) * c == a * c + b * c``),
1495-
* Operands to floating-point operations are not equal to ``NaN`` and
1496-
``Inf``, and
1497-
* ``+0`` and ``-0`` are interchangeable.
1497+
* No ``NaN`` or infinite values will be operands or results of
1498+
floating-point operations,
1499+
* ``+0`` and ``-0`` may be treated as interchangeable.
14981500

14991501
``-ffast-math`` also defines the ``__FAST_MATH__`` preprocessor
15001502
macro. Some math libraries recognize this macro and change their behavior.
@@ -1753,7 +1755,7 @@ for more details.
17531755
Specify floating point behavior. ``-ffp-model`` is an umbrella
17541756
option that encompasses functionality provided by other, single
17551757
purpose, floating point options. Valid values are: ``precise``, ``strict``,
1756-
and ``fast``.
1758+
``fast``, and ``aggressive``.
17571759
Details:
17581760

17591761
* ``precise`` Disables optimizations that are not value-safe on
@@ -1766,7 +1768,10 @@ for more details.
17661768
``STDC FENV_ACCESS``: by default ``FENV_ACCESS`` is disabled. This option
17671769
setting behaves as though ``#pragma STDC FENV_ACCESS ON`` appeared at the
17681770
top of the source file.
1769-
* ``fast`` Behaves identically to specifying both ``-ffast-math`` and
1771+
* ``fast`` Behaves identically to specifying ``-funsafe-math-optimizations``,
1772+
``-fno-math-errno`` and ``-fcomplex-arithmetic=promoted``
1773+
``ffp-contract=fast``
1774+
* ``aggressive`` Behaves identically to specifying both ``-ffast-math`` and
17701775
``ffp-contract=fast``
17711776

17721777
Note: If your command line specifies multiple instances

‎clang/lib/Driver/ToolChain.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1366,7 +1366,7 @@ bool ToolChain::isFastMathRuntimeAvailable(const ArgList &Args,
13661366
Default = false;
13671367
if (A && A->getOption().getID() == options::OPT_ffp_model_EQ) {
13681368
StringRef Model = A->getValue();
1369-
if (Model != "fast")
1369+
if (Model != "fast" && Model != "aggressive")
13701370
Default = false;
13711371
}
13721372
}

‎clang/lib/Driver/ToolChains/Clang.cpp

Lines changed: 33 additions & 21 deletions
Original file line numberDiff line numberDiff line change
@@ -2885,10 +2885,29 @@ static void RenderFloatingPointOptions(const ToolChain &TC, const Driver &D,
28852885
std::string ComplexRangeStr = "";
28862886
std::string GccRangeComplexOption = "";
28872887

2888+
auto setComplexRange = [&](LangOptions::ComplexRangeKind NewRange) {
2889+
// Warn if user expects to perform full implementation of complex
2890+
// multiplication or division in the presence of nnan or ninf flags.
2891+
if (Range != NewRange)
2892+
EmitComplexRangeDiag(D,
2893+
!GccRangeComplexOption.empty()
2894+
? GccRangeComplexOption
2895+
: ComplexArithmeticStr(Range),
2896+
ComplexArithmeticStr(NewRange));
2897+
Range = NewRange;
2898+
};
2899+
28882900
// Lambda to set fast-math options. This is also used by -ffp-model=fast
2889-
auto applyFastMath = [&]() {
2890-
HonorINFs = false;
2891-
HonorNaNs = false;
2901+
auto applyFastMath = [&](bool Aggressive) {
2902+
if (Aggressive) {
2903+
HonorINFs = false;
2904+
HonorNaNs = false;
2905+
setComplexRange(LangOptions::ComplexRangeKind::CX_Basic);
2906+
} else {
2907+
HonorINFs = true;
2908+
HonorNaNs = true;
2909+
setComplexRange(LangOptions::ComplexRangeKind::CX_Promoted);
2910+
}
28922911
MathErrno = false;
28932912
AssociativeMath = true;
28942913
ReciprocalMath = true;
@@ -2897,21 +2916,7 @@ static void RenderFloatingPointOptions(const ToolChain &TC, const Driver &D,
28972916
TrappingMath = false;
28982917
RoundingFPMath = false;
28992918
FPExceptionBehavior = "";
2900-
// If fast-math is set then set the fp-contract mode to fast.
29012919
FPContract = "fast";
2902-
// ffast-math enables basic range rules for complex multiplication and
2903-
// division.
2904-
// Warn if user expects to perform full implementation of complex
2905-
// multiplication or division in the presence of nan or ninf flags.
2906-
if (Range == LangOptions::ComplexRangeKind::CX_Full ||
2907-
Range == LangOptions::ComplexRangeKind::CX_Improved ||
2908-
Range == LangOptions::ComplexRangeKind::CX_Promoted)
2909-
EmitComplexRangeDiag(
2910-
D, ComplexArithmeticStr(Range),
2911-
!GccRangeComplexOption.empty()
2912-
? GccRangeComplexOption
2913-
: ComplexArithmeticStr(LangOptions::ComplexRangeKind::CX_Basic));
2914-
Range = LangOptions::ComplexRangeKind::CX_Basic;
29152920
SeenUnsafeMathModeOption = true;
29162921
};
29172922

@@ -3039,8 +3044,8 @@ static void RenderFloatingPointOptions(const ToolChain &TC, const Driver &D,
30393044
SignedZeros = true;
30403045

30413046
StringRef Val = A->getValue();
3042-
if (OFastEnabled && Val != "fast") {
3043-
// Only -ffp-model=fast is compatible with OFast, ignore.
3047+
if (OFastEnabled && Val != "aggressive") {
3048+
// Only -ffp-model=aggressive is compatible with -OFast, ignore.
30443049
D.Diag(clang::diag::warn_drv_overriding_option)
30453050
<< Args.MakeArgString("-ffp-model=" + Val) << "-Ofast";
30463051
break;
@@ -3052,13 +3057,19 @@ static void RenderFloatingPointOptions(const ToolChain &TC, const Driver &D,
30523057
<< Args.MakeArgString("-ffp-model=" + Val);
30533058
if (Val == "fast") {
30543059
FPModel = Val;
3055-
applyFastMath();
3060+
applyFastMath(false);
30563061
// applyFastMath sets fp-contract="fast"
30573062
LastFpContractOverrideOption = "-ffp-model=fast";
3063+
} else if (Val == "aggressive") {
3064+
FPModel = Val;
3065+
applyFastMath(true);
3066+
// applyFastMath sets fp-contract="fast"
3067+
LastFpContractOverrideOption = "-ffp-model=aggressive";
30583068
} else if (Val == "precise") {
30593069
FPModel = Val;
30603070
FPContract = "on";
30613071
LastFpContractOverrideOption = "-ffp-model=precise";
3072+
setComplexRange(LangOptions::ComplexRangeKind::CX_Full);
30623073
} else if (Val == "strict") {
30633074
StrictFPModel = true;
30643075
FPExceptionBehavior = "strict";
@@ -3067,6 +3078,7 @@ static void RenderFloatingPointOptions(const ToolChain &TC, const Driver &D,
30673078
LastFpContractOverrideOption = "-ffp-model=strict";
30683079
TrappingMath = true;
30693080
RoundingFPMath = true;
3081+
setComplexRange(LangOptions::ComplexRangeKind::CX_Full);
30703082
} else
30713083
D.Diag(diag::err_drv_unsupported_option_argument)
30723084
<< A->getSpelling() << Val;
@@ -3247,7 +3259,7 @@ static void RenderFloatingPointOptions(const ToolChain &TC, const Driver &D,
32473259
continue;
32483260
[[fallthrough]];
32493261
case options::OPT_ffast_math:
3250-
applyFastMath();
3262+
applyFastMath(true);
32513263
if (A->getOption().getID() == options::OPT_Ofast)
32523264
LastFpContractOverrideOption = "-Ofast";
32533265
else

‎clang/test/CodeGen/ffp-model.c

Lines changed: 27 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -3,6 +3,9 @@
33
// RUN: %clang -S -emit-llvm -fenable-matrix -ffp-model=fast %s -o - \
44
// RUN: | FileCheck %s --check-prefixes=CHECK,CHECK-FAST
55

6+
// RUN: %clang -S -emit-llvm -fenable-matrix -ffp-model=aggressive %s -o - \
7+
// RUN: | FileCheck %s --check-prefixes=CHECK,CHECK-AGGRESSIVE
8+
69
// RUN: %clang -S -emit-llvm -fenable-matrix -ffp-model=precise %s -o - \
710
// RUN: | FileCheck %s --check-prefixes=CHECK,CHECK-PRECISE
811

@@ -20,9 +23,13 @@ float mymuladd(float x, float y, float z) {
2023
// CHECK: define{{.*}} float @mymuladd
2124
return x * y + z;
2225

23-
// CHECK-FAST: fmul fast float
26+
// CHECK-AGGRESSIVE: fmul fast float
27+
// CHECK-AGGRESSIVE: load float, ptr
28+
// CHECK-AGGRESSIVE: fadd fast float
29+
30+
// CHECK-FAST: fmul reassoc nsz arcp contract afn float
2431
// CHECK-FAST: load float, ptr
25-
// CHECK-FAST: fadd fast float
32+
// CHECK-FAST: fadd reassoc nsz arcp contract afn float
2633

2734
// CHECK-PRECISE: load float, ptr
2835
// CHECK-PRECISE: load float, ptr
@@ -54,9 +61,13 @@ void my_vec_muladd(v2f x, float y, v2f z, v2f *res) {
5461
// CHECK: define{{.*}}@my_vec_muladd
5562
*res = x * y + z;
5663

57-
// CHECK-FAST: fmul fast <2 x float>
64+
// CHECK-AGGRESSIVE: fmul fast <2 x float>
65+
// CHECK-AGGRESSIVE: load <2 x float>, ptr
66+
// CHECK-AGGRESSIVE: fadd fast <2 x float>
67+
68+
// CHECK-FAST: fmul reassoc nsz arcp contract afn <2 x float>
5869
// CHECK-FAST: load <2 x float>, ptr
59-
// CHECK-FAST: fadd fast <2 x float>
70+
// CHECK-FAST: fadd reassoc nsz arcp contract afn <2 x float>
6071

6172
// CHECK-PRECISE: load <2 x float>, ptr
6273
// CHECK-PRECISE: load float, ptr
@@ -88,9 +99,13 @@ void my_m21_muladd(m21f x, float y, m21f z, m21f *res) {
8899
// CHECK: define{{.*}}@my_m21_muladd
89100
*res = x * y + z;
90101

91-
// CHECK-FAST: fmul fast <2 x float>
102+
// CHECK-AGGRESSIVE: fmul fast <2 x float>
103+
// CHECK-AGGRESSIVE: load <2 x float>, ptr
104+
// CHECK-AGGRESSIVE: fadd fast <2 x float>
105+
106+
// CHECK-FAST: fmul reassoc nsz arcp contract afn <2 x float>
92107
// CHECK-FAST: load <2 x float>, ptr
93-
// CHECK-FAST: fadd fast <2 x float>
108+
// CHECK-FAST: fadd reassoc nsz arcp contract afn <2 x float>
94109

95110
// CHECK-PRECISE: load <2 x float>, ptr
96111
// CHECK-PRECISE: load float, ptr
@@ -122,9 +137,13 @@ void my_m22_muladd(m22f x, float y, m22f z, m22f *res) {
122137
// CHECK: define{{.*}}@my_m22_muladd
123138
*res = x * y + z;
124139

125-
// CHECK-FAST: fmul fast <4 x float>
140+
// CHECK-AGGRESSIVE: fmul fast <4 x float>
141+
// CHECK-AGGRESSIVE: load <4 x float>, ptr
142+
// CHECK-AGGRESSIVE: fadd fast <4 x float>
143+
144+
// CHECK-FAST: fmul reassoc nsz arcp contract afn <4 x float>
126145
// CHECK-FAST: load <4 x float>, ptr
127-
// CHECK-FAST: fadd fast <4 x float>
146+
// CHECK-FAST: fadd reassoc nsz arcp contract afn <4 x float>
128147

129148
// CHECK-PRECISE: load <4 x float>, ptr
130149
// CHECK-PRECISE: load float, ptr

0 commit comments

Comments
 (0)