a105877646
The idea behind this canonicalization is that it allows us to handle less patterns, because we know that some will be canonicalized away. This is indeed very useful to e.g. know that constants are always on the right. However, this is only useful if the canonicalization is actually reliable. This is the case for constants, but not for arguments: Moving these to the right makes it look like the "more complex" expression is guaranteed to be on the left, but this is not actually the case in practice. It fails as soon as you replace the argument with another instruction. The end result is that it looks like things correctly work in tests, while they actually don't. We use the "thwart complexity-based canonicalization" trick to handle this in tests, but it's often a challenge for new contributors to get this right, and based on the regressions this PR originally exposed, we clearly don't get this right in many cases. For this reason, I think that it's better to remove this complexity canonicalization. It will make it much easier to write tests for commuted cases and make sure that they are handled.
131 lines
4.1 KiB
C++
131 lines
4.1 KiB
C++
// RUN: %clang_cc1 -O3 -triple %itanium_abi_triple -emit-llvm -o - %s | FileCheck -check-prefixes=CHECK,DEFAULT %s
|
|
// RUN: %clang_cc1 -O3 -triple %itanium_abi_triple -freciprocal-math -emit-llvm -o - %s | FileCheck -check-prefixes=CHECK,FLAG %s
|
|
|
|
float base(float a, float b, float c) {
|
|
// CHECK-LABEL: _Z4basefff
|
|
// FLAG: %[[A:.+]] = fdiv arcp float %b, %c
|
|
// FLAG: %[[M:.+]] = fdiv arcp float %[[A]], %b
|
|
// FLAG-NEXT: fadd arcp float %c, %[[M]]
|
|
|
|
// DEFAULT: %[[A:.+]] = fdiv float %b, %c
|
|
// DEFAULT: %[[M:.+]] = fdiv float %[[A]], %b
|
|
// DEFAULT-NEXT: fadd float %c, %[[M]]
|
|
a = b / c;
|
|
return a / b + c;
|
|
}
|
|
|
|
// Simple case
|
|
float fp_recip_simple(float a, float b, float c) {
|
|
// CHECK-LABEL: _Z15fp_recip_simplefff
|
|
// CHECK: %[[A:.+]] = fdiv arcp float %b, %c
|
|
// CHECK: %[[M:.+]] = fdiv arcp float %[[A]], %b
|
|
// CHECK-NEXT: fadd arcp float %c, %[[M]]
|
|
#pragma clang fp reciprocal(on)
|
|
a = b / c;
|
|
return a / b + c;
|
|
}
|
|
|
|
// Test interaction with -freciprocal-math
|
|
float fp_recip_disable(float a, float b, float c) {
|
|
// CHECK-LABEL: _Z16fp_recip_disablefff
|
|
// CHECK: %[[A:.+]] = fdiv float %b, %c
|
|
// CHECK: %[[M:.+]] = fdiv float %[[A]], %b
|
|
// CHECK-NEXT: fadd float %c, %[[M]]
|
|
#pragma clang fp reciprocal(off)
|
|
a = b / c;
|
|
return a / b + c;
|
|
}
|
|
|
|
float fp_recip_with_reassoc_simple(float a, float b, float c) {
|
|
// CHECK-LABEL: _Z28fp_recip_with_reassoc_simplefff
|
|
// CHECK: %[[A:.+]] = fmul reassoc arcp float %b, %c
|
|
// CHECK: %[[M:.+]] = fdiv reassoc arcp float %b, %[[A]]
|
|
// CHECK-NEXT: fadd reassoc arcp float %c, %[[M]]
|
|
#pragma clang fp reciprocal(on) reassociate(on)
|
|
a = b / c;
|
|
return a / b + c;
|
|
}
|
|
|
|
// arcp pragma should only apply to its scope
|
|
float fp_recip_scoped(float a, float b, float c) {
|
|
// CHECK-LABEL: _Z15fp_recip_scopedfff
|
|
// DEFAULT: %[[M:.+]] = fdiv float %a, %b
|
|
// DEFAULT-NEXT: fadd float %[[M]], %c
|
|
// FLAG: %[[M:.+]] = fdiv arcp float %a, %b
|
|
// FLAG-NEXT: fadd arcp float %[[M]], %c
|
|
{
|
|
#pragma clang fp reciprocal(on)
|
|
}
|
|
return a / b + c;
|
|
}
|
|
|
|
// arcp pragma should apply to templates as well
|
|
class Foo {};
|
|
Foo operator+(Foo, Foo);
|
|
template <typename T>
|
|
T template_recip(T a, T b, T c) {
|
|
#pragma clang fp reciprocal(on)
|
|
return ((a / b) - c) + c;
|
|
}
|
|
|
|
float fp_recip_template(float a, float b, float c) {
|
|
// CHECK-LABEL: _Z17fp_recip_templatefff
|
|
// CHECK: %[[A1:.+]] = fdiv arcp float %a, %b
|
|
// CHECK-NEXT: %[[A2:.+]] = fsub arcp float %[[A1]], %c
|
|
// CHECK-NEXT: fadd arcp float %c, %[[A2]]
|
|
return template_recip<float>(a, b, c);
|
|
}
|
|
|
|
// File Scoping should work across functions
|
|
#pragma clang fp reciprocal(on)
|
|
float fp_file_scope_on(float a, float b, float c) {
|
|
// CHECK-LABEL: _Z16fp_file_scope_onfff
|
|
// CHECK: %[[M1:.+]] = fdiv arcp float %a, %c
|
|
// CHECK-NEXT: %[[M2:.+]] = fdiv arcp float %b, %c
|
|
// CHECK-NEXT: fadd arcp float %[[M1]], %[[M2]]
|
|
return (a / c) + (b / c);
|
|
}
|
|
|
|
// Inner pragma has precedence
|
|
float fp_file_scope_stop(float a, float b, float c) {
|
|
// CHECK-LABEL: _Z18fp_file_scope_stopfff
|
|
// CHECK: %[[A:.+]] = fdiv arcp float %a, %a
|
|
// CHECK: %[[M1:.+]] = fdiv float %[[A]], %c
|
|
// CHECK-NEXT: %[[M2:.+]] = fdiv float %b, %c
|
|
// CHECK-NEXT: fsub float %[[M1]], %[[M2]]
|
|
a = a / a;
|
|
{
|
|
#pragma clang fp reciprocal(off)
|
|
return (a / c) - (b / c);
|
|
}
|
|
}
|
|
|
|
#pragma clang fp reciprocal(off)
|
|
float fp_recip_off(float a, float b, float c) {
|
|
// CHECK-LABEL: _Z12fp_recip_offfff
|
|
// CHECK: %[[D1:.+]] = fdiv float %a, %c
|
|
// CHECK-NEXT: %[[D2:.+]] = fdiv float %b, %c
|
|
// CHECK-NEXT: fadd float %[[D1]], %[[D2]]
|
|
return (a / c) + (b / c);
|
|
}
|
|
|
|
// Takes latest flag
|
|
float fp_recip_many(float a, float b, float c) {
|
|
// CHECK-LABEL: _Z13fp_recip_manyfff
|
|
// CHECK: %[[D1:.+]] = fdiv arcp float %a, %c
|
|
// CHECK-NEXT: %[[D2:.+]] = fdiv arcp float %b, %c
|
|
// CHECK-NEXT: fadd arcp float %[[D1]], %[[D2]]
|
|
#pragma clang fp reciprocal(off) reciprocal(on)
|
|
return (a / c) + (b / c);
|
|
}
|
|
|
|
// Pragma does not propagate through called functions
|
|
float helper_func(float a, float b, float c) { return a + b + c; }
|
|
float fp_recip_call_helper(float a, float b, float c) {
|
|
// CHECK-LABEL: _Z20fp_recip_call_helperfff
|
|
// CHECK: %[[S1:.+]] = fadd float %a, %b
|
|
// CHECK-NEXT: fadd float %[[S1]], %c
|
|
#pragma clang fp reciprocal(on)
|
|
return helper_func(a, b, c);
|
|
}
|