[X86] Custom lower UINT_TO_FP from v4f32 to v4i32, and for v8f32 to v8i32 if
AVX2 is available.
According to IACA, the new lowering has a throughput of 8 cycles instead of 13
with the previous one.
Althought this lowering kicks in some SPECs benchmarks, the performance
improvement was within the noise.
Correctness testing has been done for the whole range of uint32_t with the
following program:
uint4 v = (uint4) {0,1,2,3};
uint32_t i;
//Check correctness over entire range for uint4 -> float4 conversion
for( i = 0; i < 1U << (32-2); i++ )
{
float4 t = test(v);
float4 c = correct(v);
if( 0xf != _mm_movemask_ps( t == c ))
{
printf( "Error @ %vx: %vf vs. %vf\n", v, c, t);
return -1;
}
v += 4;
}
Where "correct" is the old lowering and "test" the new one.
The patch adds a test case for the two custom lowering instruction.
It also modifies the vector cost model, which is why cast.ll and uitofp.ll are
modified.
2009-02-26-MachineLICMBug.ll is also modified because we now hoist 7
instructions instead of 4 (3 more constant loads).
rdar://problem/18153096>
llvm-svn: 221657
2014-11-11 10:23:47 +08:00
|
|
|
; RUN: llc < %s -mtriple=x86_64-apple-macosx | FileCheck --check-prefix=CHECK --check-prefix=SSE --check-prefix=CST %s
|
|
|
|
|
; RUN: llc < %s -mtriple=x86_64-apple-macosx -mattr=+sse4.1 | FileCheck --check-prefix=CHECK --check-prefix=SSE41 --check-prefix=CST %s
|
|
|
|
|
; RUN: llc < %s -mtriple=x86_64-apple-macosx -mattr=+avx | FileCheck --check-prefix=CHECK --check-prefix=AVX --check-prefix=CST %s
|
|
|
|
|
; RUN: llc < %s -mtriple=x86_64-apple-macosx -mattr=+avx2 | FileCheck --check-prefix=CHECK --check-prefix=AVX2 %s
|
|
|
|
|
|
|
|
|
|
; Check that the constant used in the vectors are the right ones.
|
|
|
|
|
; SSE: [[MASKCSTADDR:LCPI0_[0-9]+]]:
|
|
|
|
|
; SSE-NEXT: .long 65535 ## 0xffff
|
|
|
|
|
; SSE-NEXT: .long 65535 ## 0xffff
|
|
|
|
|
; SSE-NEXT: .long 65535 ## 0xffff
|
|
|
|
|
; SSE-NEXT: .long 65535 ## 0xffff
|
|
|
|
|
|
|
|
|
|
; CST: [[LOWCSTADDR:LCPI0_[0-9]+]]:
|
|
|
|
|
; CST-NEXT: .long 1258291200 ## 0x4b000000
|
|
|
|
|
; CST-NEXT: .long 1258291200 ## 0x4b000000
|
|
|
|
|
; CST-NEXT: .long 1258291200 ## 0x4b000000
|
|
|
|
|
; CST-NEXT: .long 1258291200 ## 0x4b000000
|
|
|
|
|
|
|
|
|
|
; CST: [[HIGHCSTADDR:LCPI0_[0-9]+]]:
|
|
|
|
|
; CST-NEXT: .long 1392508928 ## 0x53000000
|
|
|
|
|
; CST-NEXT: .long 1392508928 ## 0x53000000
|
|
|
|
|
; CST-NEXT: .long 1392508928 ## 0x53000000
|
|
|
|
|
; CST-NEXT: .long 1392508928 ## 0x53000000
|
|
|
|
|
|
|
|
|
|
; CST: [[MAGICCSTADDR:LCPI0_[0-9]+]]:
|
|
|
|
|
; CST-NEXT: .long 3539992704 ## float -5.497642e+11
|
|
|
|
|
; CST-NEXT: .long 3539992704 ## float -5.497642e+11
|
|
|
|
|
; CST-NEXT: .long 3539992704 ## float -5.497642e+11
|
|
|
|
|
; CST-NEXT: .long 3539992704 ## float -5.497642e+11
|
|
|
|
|
|
|
|
|
|
; AVX2: [[LOWCSTADDR:LCPI0_[0-9]+]]:
|
|
|
|
|
; AVX2-NEXT: .long 1258291200 ## 0x4b000000
|
|
|
|
|
|
|
|
|
|
; AVX2: [[HIGHCSTADDR:LCPI0_[0-9]+]]:
|
|
|
|
|
; AVX2-NEXT: .long 1392508928 ## 0x53000000
|
|
|
|
|
|
|
|
|
|
; AVX2: [[MAGICCSTADDR:LCPI0_[0-9]+]]:
|
|
|
|
|
; AVX2-NEXT: .long 3539992704 ## float -5.49764202E+11
|
2011-03-19 21:09:10 +08:00
|
|
|
|
|
|
|
|
define <4 x float> @test1(<4 x i32> %A) nounwind {
|
2013-07-14 04:38:47 +08:00
|
|
|
; CHECK-LABEL: test1:
|
[X86] Custom lower UINT_TO_FP from v4f32 to v4i32, and for v8f32 to v8i32 if
AVX2 is available.
According to IACA, the new lowering has a throughput of 8 cycles instead of 13
with the previous one.
Althought this lowering kicks in some SPECs benchmarks, the performance
improvement was within the noise.
Correctness testing has been done for the whole range of uint32_t with the
following program:
uint4 v = (uint4) {0,1,2,3};
uint32_t i;
//Check correctness over entire range for uint4 -> float4 conversion
for( i = 0; i < 1U << (32-2); i++ )
{
float4 t = test(v);
float4 c = correct(v);
if( 0xf != _mm_movemask_ps( t == c ))
{
printf( "Error @ %vx: %vf vs. %vf\n", v, c, t);
return -1;
}
v += 4;
}
Where "correct" is the old lowering and "test" the new one.
The patch adds a test case for the two custom lowering instruction.
It also modifies the vector cost model, which is why cast.ll and uitofp.ll are
modified.
2009-02-26-MachineLICMBug.ll is also modified because we now hoist 7
instructions instead of 4 (3 more constant loads).
rdar://problem/18153096>
llvm-svn: 221657
2014-11-11 10:23:47 +08:00
|
|
|
;
|
|
|
|
|
; SSE: movdqa [[MASKCSTADDR]](%rip), [[MASK:%xmm[0-9]+]]
|
|
|
|
|
; SSE-NEXT: pand %xmm0, [[MASK]]
|
|
|
|
|
; After this instruction, MASK will have the value of the low parts
|
|
|
|
|
; of the vector.
|
|
|
|
|
; SSE-NEXT: por [[LOWCSTADDR]](%rip), [[MASK]]
|
|
|
|
|
; SSE-NEXT: psrld $16, %xmm0
|
|
|
|
|
; SSE-NEXT: por [[HIGHCSTADDR]](%rip), %xmm0
|
|
|
|
|
; SSE-NEXT: addps [[MAGICCSTADDR]](%rip), %xmm0
|
|
|
|
|
; SSE-NEXT: addps [[MASK]], %xmm0
|
|
|
|
|
; SSE-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; Currently we commute the arguments of the first blend, but this could be
|
|
|
|
|
; improved to match the lowering of the second blend.
|
|
|
|
|
; SSE41: movdqa [[LOWCSTADDR]](%rip), [[LOWVEC:%xmm[0-9]+]]
|
|
|
|
|
; SSE41-NEXT: pblendw $85, %xmm0, [[LOWVEC]]
|
|
|
|
|
; SSE41-NEXT: psrld $16, %xmm0
|
|
|
|
|
; SSE41-NEXT: pblendw $170, [[HIGHCSTADDR]](%rip), %xmm0
|
|
|
|
|
; SSE41-NEXT: addps [[MAGICCSTADDR]](%rip), %xmm0
|
|
|
|
|
; SSE41-NEXT: addps [[LOWVEC]], %xmm0
|
|
|
|
|
; SSE41-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX: vpblendw $170, [[LOWCSTADDR]](%rip), %xmm0, [[LOWVEC:%xmm[0-9]+]]
|
|
|
|
|
; AVX-NEXT: vpsrld $16, %xmm0, [[SHIFTVEC:%xmm[0-9]+]]
|
|
|
|
|
; AVX-NEXT: vpblendw $170, [[HIGHCSTADDR]](%rip), [[SHIFTVEC]], [[HIGHVEC:%xmm[0-9]+]]
|
|
|
|
|
; AVX-NEXT: vaddps [[MAGICCSTADDR]](%rip), [[HIGHVEC]], [[TMP:%xmm[0-9]+]]
|
|
|
|
|
; AVX-NEXT: vaddps [[TMP]], [[LOWVEC]], %xmm0
|
|
|
|
|
; AVX-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; The lowering for AVX2 is a bit messy, because we select broadcast
|
|
|
|
|
; instructions, instead of folding the constant loads.
|
|
|
|
|
; AVX2: vpbroadcastd [[LOWCSTADDR]](%rip), [[LOWCST:%xmm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vpblendw $170, [[LOWCST]], %xmm0, [[LOWVEC:%xmm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vpsrld $16, %xmm0, [[SHIFTVEC:%xmm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vpbroadcastd [[HIGHCSTADDR]](%rip), [[HIGHCST:%xmm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vpblendw $170, [[HIGHCST]], [[SHIFTVEC]], [[HIGHVEC:%xmm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vbroadcastss [[MAGICCSTADDR]](%rip), [[MAGICCST:%xmm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vaddps [[MAGICCST]], [[HIGHVEC]], [[TMP:%xmm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vaddps [[TMP]], [[LOWVEC]], %xmm0
|
|
|
|
|
; AVX2-NEXT: retq
|
2011-03-19 21:09:10 +08:00
|
|
|
%C = uitofp <4 x i32> %A to <4 x float>
|
|
|
|
|
ret <4 x float> %C
|
|
|
|
|
}
|
|
|
|
|
|
[X86] Custom lower UINT_TO_FP from v4f32 to v4i32, and for v8f32 to v8i32 if
AVX2 is available.
According to IACA, the new lowering has a throughput of 8 cycles instead of 13
with the previous one.
Althought this lowering kicks in some SPECs benchmarks, the performance
improvement was within the noise.
Correctness testing has been done for the whole range of uint32_t with the
following program:
uint4 v = (uint4) {0,1,2,3};
uint32_t i;
//Check correctness over entire range for uint4 -> float4 conversion
for( i = 0; i < 1U << (32-2); i++ )
{
float4 t = test(v);
float4 c = correct(v);
if( 0xf != _mm_movemask_ps( t == c ))
{
printf( "Error @ %vx: %vf vs. %vf\n", v, c, t);
return -1;
}
v += 4;
}
Where "correct" is the old lowering and "test" the new one.
The patch adds a test case for the two custom lowering instruction.
It also modifies the vector cost model, which is why cast.ll and uitofp.ll are
modified.
2009-02-26-MachineLICMBug.ll is also modified because we now hoist 7
instructions instead of 4 (3 more constant loads).
rdar://problem/18153096>
llvm-svn: 221657
2014-11-11 10:23:47 +08:00
|
|
|
; Match the AVX2 constants used in the next function
|
|
|
|
|
; AVX2: [[LOWCSTADDR:LCPI1_[0-9]+]]:
|
|
|
|
|
; AVX2-NEXT: .long 1258291200 ## 0x4b000000
|
|
|
|
|
|
|
|
|
|
; AVX2: [[HIGHCSTADDR:LCPI1_[0-9]+]]:
|
|
|
|
|
; AVX2-NEXT: .long 1392508928 ## 0x53000000
|
|
|
|
|
|
|
|
|
|
; AVX2: [[MAGICCSTADDR:LCPI1_[0-9]+]]:
|
|
|
|
|
; AVX2-NEXT: .long 3539992704 ## float -5.49764202E+11
|
|
|
|
|
|
|
|
|
|
define <8 x float> @test2(<8 x i32> %A) nounwind {
|
|
|
|
|
; CHECK-LABEL: test2:
|
|
|
|
|
; Legalization will break the thing is 2 x <4 x i32> on anthing prior AVX.
|
|
|
|
|
; The constant used for in the vector instruction are shared between the
|
|
|
|
|
; two sequences of instructions.
|
|
|
|
|
;
|
|
|
|
|
; SSE: movdqa {{.*#+}} [[MASK:xmm[0-9]+]] = [65535,65535,65535,65535]
|
|
|
|
|
; SSE-NEXT: movdqa %xmm0, [[VECLOW:%xmm[0-9]+]]
|
|
|
|
|
; SSE-NEXT: pand %[[MASK]], [[VECLOW]]
|
|
|
|
|
; SSE-NEXT: movdqa {{.*#+}} [[LOWCST:xmm[0-9]+]] = [1258291200,1258291200,1258291200,1258291200]
|
|
|
|
|
; SSE-NEXT: por %[[LOWCST]], [[VECLOW]]
|
|
|
|
|
; SSE-NEXT: psrld $16, %xmm0
|
|
|
|
|
; SSE-NEXT: movdqa {{.*#+}} [[HIGHCST:xmm[0-9]+]] = [1392508928,1392508928,1392508928,1392508928]
|
|
|
|
|
; SSE-NEXT: por %[[HIGHCST]], %xmm0
|
|
|
|
|
; SSE-NEXT: movaps {{.*#+}} [[MAGICCST:xmm[0-9]+]] = [-5.497642e+11,-5.497642e+11,-5.497642e+11,-5.497642e+11]
|
|
|
|
|
; SSE-NEXT: addps %[[MAGICCST]], %xmm0
|
|
|
|
|
; SSE-NEXT: addps [[VECLOW]], %xmm0
|
|
|
|
|
; MASK is the low vector of the second part after this point.
|
|
|
|
|
; SSE-NEXT: pand %xmm1, %[[MASK]]
|
|
|
|
|
; SSE-NEXT: por %[[LOWCST]], %[[MASK]]
|
|
|
|
|
; SSE-NEXT: psrld $16, %xmm1
|
|
|
|
|
; SSE-NEXT: por %[[HIGHCST]], %xmm1
|
|
|
|
|
; SSE-NEXT: addps %[[MAGICCST]], %xmm1
|
|
|
|
|
; SSE-NEXT: addps %[[MASK]], %xmm1
|
|
|
|
|
; SSE-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; SSE41: movdqa {{.*#+}} [[LOWCST:xmm[0-9]+]] = [1258291200,1258291200,1258291200,1258291200]
|
|
|
|
|
; SSE41-NEXT: movdqa %xmm0, [[VECLOW:%xmm[0-9]+]]
|
|
|
|
|
; SSE41-NEXT: pblendw $170, %[[LOWCST]], [[VECLOW]]
|
|
|
|
|
; SSE41-NEXT: psrld $16, %xmm0
|
|
|
|
|
; SSE41-NEXT: movdqa {{.*#+}} [[HIGHCST:xmm[0-9]+]] = [1392508928,1392508928,1392508928,1392508928]
|
|
|
|
|
; SSE41-NEXT: pblendw $170, %[[HIGHCST]], %xmm0
|
|
|
|
|
; SSE41-NEXT: movaps {{.*#+}} [[MAGICCST:xmm[0-9]+]] = [-5.497642e+11,-5.497642e+11,-5.497642e+11,-5.497642e+11]
|
|
|
|
|
; SSE41-NEXT: addps %[[MAGICCST]], %xmm0
|
|
|
|
|
; SSE41-NEXT: addps [[VECLOW]], %xmm0
|
|
|
|
|
; LOWCST is the low vector of the second part after this point.
|
|
|
|
|
; The operands of the blend are inverted because we reuse xmm1
|
|
|
|
|
; in the next shift.
|
|
|
|
|
; SSE41-NEXT: pblendw $85, %xmm1, %[[LOWCST]]
|
|
|
|
|
; SSE41-NEXT: psrld $16, %xmm1
|
|
|
|
|
; SSE41-NEXT: pblendw $170, %[[HIGHCST]], %xmm1
|
|
|
|
|
; SSE41-NEXT: addps %[[MAGICCST]], %xmm1
|
|
|
|
|
; SSE41-NEXT: addps %[[LOWCST]], %xmm1
|
|
|
|
|
; SSE41-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; Test that we are not lowering uinttofp to scalars
|
|
|
|
|
; AVX-NOT: cvtsd2ss
|
|
|
|
|
; AVX: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX2: vpbroadcastd [[LOWCSTADDR]](%rip), [[LOWCST:%ymm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vpblendw $170, [[LOWCST]], %ymm0, [[LOWVEC:%ymm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vpsrld $16, %ymm0, [[SHIFTVEC:%ymm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vpbroadcastd [[HIGHCSTADDR]](%rip), [[HIGHCST:%ymm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vpblendw $170, [[HIGHCST]], [[SHIFTVEC]], [[HIGHVEC:%ymm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vbroadcastss [[MAGICCSTADDR]](%rip), [[MAGICCST:%ymm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vaddps [[MAGICCST]], [[HIGHVEC]], [[TMP:%ymm[0-9]+]]
|
|
|
|
|
; AVX2-NEXT: vaddps [[TMP]], [[LOWVEC]], %ymm0
|
|
|
|
|
; AVX2-NEXT: retq
|
|
|
|
|
%C = uitofp <8 x i32> %A to <8 x float>
|
|
|
|
|
ret <8 x float> %C
|
|
|
|
|
}
|