[SelectionDAG] Lower llvm.ldexp.f32 to ldexp() on Windows. #95301
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This reduces codesize. As discussed in llvm#92707.
|
@llvm/pr-subscribers-backend-aarch64 @llvm/pr-subscribers-llvm-selectiondag Author: Eli Friedman (efriedma-quic) ChangesThis reduces codesize. As discussed in #92707. Patch is 34.12 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/95301.diff 2 Files Affected:
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index 8cd2bb60d81f2..1d9f2fe65e6fb 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -3650,6 +3650,27 @@ bool SelectionDAGLegalize::ExpandNode(SDNode *Node) {
// FIXME: Use separate LibCall action.
if (TLI.getLibcallName(LC))
break;
+ if (Node->getOpcode() == ISD::FLDEXP && VT == MVT::f32 &&
+ TLI.isTypeLegal(MVT::f64) &&
+ TLI.getLibcallName(RTLIB::getLDEXP(MVT::f64))) {
+ // On Windows, it's common to be missing the 32-bit libcall, but have
+ // the 64-bit libcall. Expand to the 64-bit libcall. (Note that ldexp
+ // involves a rounding step if the result is subnormal, but that isn't
+ // relevant here because any subnormal result will round to zero when
+ // it's truncated.)
+ //
+ // FIXME: Consider doing something similar for f16/bf16. But be very
+ // careful handling bf16: expanding bf16->f64 is fine, but expanding
+ // bf16->f32 would produce incorrect subnormal results.
+ SDValue Extended =
+ DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Node->getOperand(0));
+ SDValue LdExp =
+ DAG.getNode(ISD::FLDEXP, dl, MVT::f64, Extended, Node->getOperand(1));
+ Results.push_back(
+ DAG.getNode(ISD::FP_ROUND, dl, VT, LdExp,
+ DAG.getIntPtrConstant(0, dl, /*isTarget=*/true)));
+ break;
+ }
if (SDValue Expanded = expandLdexp(Node)) {
Results.push_back(Expanded);
diff --git a/llvm/test/CodeGen/X86/ldexp.ll b/llvm/test/CodeGen/X86/ldexp.ll
index 2be5dec156690..d3b02dc8f9b7c 100644
--- a/llvm/test/CodeGen/X86/ldexp.ll
+++ b/llvm/test/CodeGen/X86/ldexp.ll
@@ -1,5 +1,6 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
; RUN: llc -mtriple=x86_64-unknown-unknown -verify-machineinstrs < %s | FileCheck -check-prefixes=X64 %s
+; RUN: llc -mtriple=x86_64-pc-win32 -verify-machineinstrs < %s | FileCheck -check-prefixes=WIN64 %s
; RUN: llc -mtriple=i386-pc-win32 -verify-machineinstrs < %s | FileCheck -check-prefix=WIN32 %s
define float @ldexp_f32(i8 zeroext %x) {
@@ -8,75 +9,30 @@ define float @ldexp_f32(i8 zeroext %x) {
; X64-NEXT: movss {{.*#+}} xmm0 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
; X64-NEXT: jmp ldexpf@PLT # TAILCALL
;
+; WIN64-LABEL: ldexp_f32:
+; WIN64: # %bb.0:
+; WIN64-NEXT: subq $40, %rsp
+; WIN64-NEXT: .seh_stackalloc 40
+; WIN64-NEXT: .seh_endprologue
+; WIN64-NEXT: movzbl %cl, %edx
+; WIN64-NEXT: movsd {{.*#+}} xmm0 = [1.0E+0,0.0E+0]
+; WIN64-NEXT: callq ldexp
+; WIN64-NEXT: cvtsd2ss %xmm0, %xmm0
+; WIN64-NEXT: addq $40, %rsp
+; WIN64-NEXT: retq
+; WIN64-NEXT: .seh_endproc
+;
; WIN32-LABEL: ldexp_f32:
; WIN32: # %bb.0:
-; WIN32-NEXT: pushl %eax
-; WIN32-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
-; WIN32-NEXT: cmpl $381, %ecx # imm = 0x17D
-; WIN32-NEXT: movl %ecx, %eax
-; WIN32-NEXT: jl LBB0_2
-; WIN32-NEXT: # %bb.1:
-; WIN32-NEXT: movl $381, %eax # imm = 0x17D
-; WIN32-NEXT: LBB0_2:
-; WIN32-NEXT: addl $-254, %eax
-; WIN32-NEXT: leal -127(%ecx), %edx
-; WIN32-NEXT: cmpl $255, %ecx
-; WIN32-NEXT: jae LBB0_4
-; WIN32-NEXT: # %bb.3:
-; WIN32-NEXT: movl %edx, %eax
-; WIN32-NEXT: LBB0_4:
-; WIN32-NEXT: flds __real@7f800000
-; WIN32-NEXT: flds __real@7f000000
-; WIN32-NEXT: jae LBB0_6
-; WIN32-NEXT: # %bb.5:
-; WIN32-NEXT: fstp %st(1)
-; WIN32-NEXT: fldz
-; WIN32-NEXT: LBB0_6:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: cmpl $-329, %ecx # imm = 0xFEB7
-; WIN32-NEXT: movl %ecx, %edx
-; WIN32-NEXT: jge LBB0_8
-; WIN32-NEXT: # %bb.7:
-; WIN32-NEXT: movl $-330, %edx # imm = 0xFEB6
-; WIN32-NEXT: LBB0_8:
-; WIN32-NEXT: cmpl $-228, %ecx
-; WIN32-NEXT: fldz
-; WIN32-NEXT: flds __real@0c800000
-; WIN32-NEXT: jb LBB0_9
-; WIN32-NEXT: # %bb.10:
-; WIN32-NEXT: fstp %st(1)
-; WIN32-NEXT: leal 102(%ecx), %edx
-; WIN32-NEXT: cmpl $-126, %ecx
-; WIN32-NEXT: jge LBB0_12
-; WIN32-NEXT: jmp LBB0_13
-; WIN32-NEXT: LBB0_9:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: addl $204, %edx
-; WIN32-NEXT: cmpl $-126, %ecx
-; WIN32-NEXT: jl LBB0_13
-; WIN32-NEXT: LBB0_12:
-; WIN32-NEXT: movl %ecx, %edx
-; WIN32-NEXT: LBB0_13:
+; WIN32-NEXT: subl $16, %esp
+; WIN32-NEXT: movzbl {{[0-9]+}}(%esp), %eax
+; WIN32-NEXT: movl %eax, {{[0-9]+}}(%esp)
; WIN32-NEXT: fld1
-; WIN32-NEXT: jl LBB0_15
-; WIN32-NEXT: # %bb.14:
-; WIN32-NEXT: fstp %st(1)
-; WIN32-NEXT: fldz
-; WIN32-NEXT: LBB0_15:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: cmpl $127, %ecx
-; WIN32-NEXT: jg LBB0_17
-; WIN32-NEXT: # %bb.16:
-; WIN32-NEXT: fstp %st(1)
-; WIN32-NEXT: movl %edx, %eax
-; WIN32-NEXT: fldz
-; WIN32-NEXT: LBB0_17:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: shll $23, %eax
-; WIN32-NEXT: addl $1065353216, %eax # imm = 0x3F800000
-; WIN32-NEXT: movl %eax, (%esp)
-; WIN32-NEXT: fmuls (%esp)
-; WIN32-NEXT: popl %eax
+; WIN32-NEXT: fstpl (%esp)
+; WIN32-NEXT: calll _ldexp
+; WIN32-NEXT: fstps {{[0-9]+}}(%esp)
+; WIN32-NEXT: flds {{[0-9]+}}(%esp)
+; WIN32-NEXT: addl $16, %esp
; WIN32-NEXT: retl
%zext = zext i8 %x to i32
%ldexp = call float @llvm.ldexp.f32.i32(float 1.000000e+00, i32 %zext)
@@ -89,6 +45,12 @@ define double @ldexp_f64(i8 zeroext %x) {
; X64-NEXT: movsd {{.*#+}} xmm0 = [1.0E+0,0.0E+0]
; X64-NEXT: jmp ldexp@PLT # TAILCALL
;
+; WIN64-LABEL: ldexp_f64:
+; WIN64: # %bb.0:
+; WIN64-NEXT: movzbl %cl, %edx
+; WIN64-NEXT: movsd {{.*#+}} xmm0 = [1.0E+0,0.0E+0]
+; WIN64-NEXT: jmp ldexp # TAILCALL
+;
; WIN32-LABEL: ldexp_f64:
; WIN32: # %bb.0:
; WIN32-NEXT: subl $12, %esp
@@ -127,152 +89,60 @@ define <2 x float> @ldexp_v2f32(<2 x float> %val, <2 x i32> %exp) {
; X64-NEXT: .cfi_def_cfa_offset 8
; X64-NEXT: retq
;
+; WIN64-LABEL: ldexp_v2f32:
+; WIN64: # %bb.0:
+; WIN64-NEXT: pushq %rsi
+; WIN64-NEXT: .seh_pushreg %rsi
+; WIN64-NEXT: subq $64, %rsp
+; WIN64-NEXT: .seh_stackalloc 64
+; WIN64-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; WIN64-NEXT: .seh_savexmm %xmm7, 48
+; WIN64-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; WIN64-NEXT: .seh_savexmm %xmm6, 32
+; WIN64-NEXT: .seh_endprologue
+; WIN64-NEXT: movaps (%rcx), %xmm7
+; WIN64-NEXT: movl (%rdx), %eax
+; WIN64-NEXT: movl 4(%rdx), %esi
+; WIN64-NEXT: cvtss2sd %xmm7, %xmm0
+; WIN64-NEXT: movl %eax, %edx
+; WIN64-NEXT: callq ldexp
+; WIN64-NEXT: xorps %xmm6, %xmm6
+; WIN64-NEXT: cvtsd2ss %xmm0, %xmm6
+; WIN64-NEXT: shufps {{.*#+}} xmm7 = xmm7[1,1,1,1]
+; WIN64-NEXT: xorps %xmm0, %xmm0
+; WIN64-NEXT: cvtss2sd %xmm7, %xmm0
+; WIN64-NEXT: movl %esi, %edx
+; WIN64-NEXT: callq ldexp
+; WIN64-NEXT: cvtsd2ss %xmm0, %xmm0
+; WIN64-NEXT: unpcklps {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1]
+; WIN64-NEXT: movaps %xmm6, %xmm0
+; WIN64-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; WIN64-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; WIN64-NEXT: addq $64, %rsp
+; WIN64-NEXT: popq %rsi
+; WIN64-NEXT: retq
+; WIN64-NEXT: .seh_endproc
+;
; WIN32-LABEL: ldexp_v2f32:
; WIN32: # %bb.0:
-; WIN32-NEXT: pushl %edi
; WIN32-NEXT: pushl %esi
-; WIN32-NEXT: subl $8, %esp
-; WIN32-NEXT: flds {{[0-9]+}}(%esp)
+; WIN32-NEXT: subl $20, %esp
+; WIN32-NEXT: movl {{[0-9]+}}(%esp), %esi
; WIN32-NEXT: movl {{[0-9]+}}(%esp), %eax
-; WIN32-NEXT: cmpl $-329, %eax # imm = 0xFEB7
-; WIN32-NEXT: movl %eax, %edx
-; WIN32-NEXT: jge LBB2_2
-; WIN32-NEXT: # %bb.1:
-; WIN32-NEXT: movl $-330, %edx # imm = 0xFEB6
-; WIN32-NEXT: LBB2_2:
-; WIN32-NEXT: addl $204, %edx
-; WIN32-NEXT: leal 102(%eax), %ecx
-; WIN32-NEXT: cmpl $-228, %eax
-; WIN32-NEXT: jb LBB2_4
-; WIN32-NEXT: # %bb.3:
-; WIN32-NEXT: movl %ecx, %edx
-; WIN32-NEXT: LBB2_4:
-; WIN32-NEXT: flds __real@0c800000
-; WIN32-NEXT: fld %st(1)
-; WIN32-NEXT: fmul %st(1), %st
-; WIN32-NEXT: fld %st(0)
-; WIN32-NEXT: fmul %st(2), %st
-; WIN32-NEXT: jb LBB2_6
-; WIN32-NEXT: # %bb.5:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: fldz
-; WIN32-NEXT: fxch %st(1)
-; WIN32-NEXT: LBB2_6:
-; WIN32-NEXT: fstp %st(1)
-; WIN32-NEXT: cmpl $-126, %eax
-; WIN32-NEXT: jl LBB2_8
-; WIN32-NEXT: # %bb.7:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: fld %st(1)
-; WIN32-NEXT: movl %eax, %edx
-; WIN32-NEXT: LBB2_8:
-; WIN32-NEXT: cmpl $381, %eax # imm = 0x17D
-; WIN32-NEXT: movl %eax, %esi
-; WIN32-NEXT: jl LBB2_10
-; WIN32-NEXT: # %bb.9:
-; WIN32-NEXT: movl $381, %esi # imm = 0x17D
-; WIN32-NEXT: LBB2_10:
-; WIN32-NEXT: flds __real@7f000000
-; WIN32-NEXT: fmul %st, %st(3)
-; WIN32-NEXT: fld %st(3)
-; WIN32-NEXT: fmul %st(1), %st
-; WIN32-NEXT: leal -127(%eax), %ecx
-; WIN32-NEXT: cmpl $255, %eax
-; WIN32-NEXT: jae LBB2_11
-; WIN32-NEXT: # %bb.12:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: jmp LBB2_13
-; WIN32-NEXT: LBB2_11:
-; WIN32-NEXT: fstp %st(4)
-; WIN32-NEXT: addl $-254, %esi
-; WIN32-NEXT: movl %esi, %ecx
-; WIN32-NEXT: LBB2_13:
-; WIN32-NEXT: cmpl $127, %eax
+; WIN32-NEXT: movl %eax, {{[0-9]+}}(%esp)
; WIN32-NEXT: flds {{[0-9]+}}(%esp)
-; WIN32-NEXT: movl {{[0-9]+}}(%esp), %esi
-; WIN32-NEXT: jg LBB2_15
-; WIN32-NEXT: # %bb.14:
-; WIN32-NEXT: movl %edx, %ecx
-; WIN32-NEXT: LBB2_15:
-; WIN32-NEXT: cmpl $381, %esi # imm = 0x17D
-; WIN32-NEXT: movl %esi, %edx
-; WIN32-NEXT: jl LBB2_17
-; WIN32-NEXT: # %bb.16:
-; WIN32-NEXT: movl $381, %edx # imm = 0x17D
-; WIN32-NEXT: LBB2_17:
-; WIN32-NEXT: addl $-254, %edx
-; WIN32-NEXT: leal -127(%esi), %edi
-; WIN32-NEXT: cmpl $255, %esi
-; WIN32-NEXT: jae LBB2_19
-; WIN32-NEXT: # %bb.18:
-; WIN32-NEXT: movl %edi, %edx
-; WIN32-NEXT: LBB2_19:
-; WIN32-NEXT: fld %st(0)
-; WIN32-NEXT: fmul %st(2), %st
-; WIN32-NEXT: fmul %st, %st(2)
-; WIN32-NEXT: jae LBB2_21
-; WIN32-NEXT: # %bb.20:
-; WIN32-NEXT: fstp %st(2)
-; WIN32-NEXT: fldz
-; WIN32-NEXT: LBB2_21:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: cmpl $-329, %esi # imm = 0xFEB7
-; WIN32-NEXT: movl %esi, %edi
-; WIN32-NEXT: jge LBB2_23
-; WIN32-NEXT: # %bb.22:
-; WIN32-NEXT: movl $-330, %edi # imm = 0xFEB6
-; WIN32-NEXT: LBB2_23:
-; WIN32-NEXT: fld %st(0)
-; WIN32-NEXT: fmul %st(4), %st
-; WIN32-NEXT: fmul %st, %st(4)
-; WIN32-NEXT: cmpl $-228, %esi
-; WIN32-NEXT: jb LBB2_24
-; WIN32-NEXT: # %bb.25:
-; WIN32-NEXT: fstp %st(4)
-; WIN32-NEXT: leal 102(%esi), %edi
-; WIN32-NEXT: cmpl $-126, %esi
-; WIN32-NEXT: jge LBB2_27
-; WIN32-NEXT: jmp LBB2_28
-; WIN32-NEXT: LBB2_24:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: addl $204, %edi
-; WIN32-NEXT: cmpl $-126, %esi
-; WIN32-NEXT: jl LBB2_28
-; WIN32-NEXT: LBB2_27:
-; WIN32-NEXT: fstp %st(3)
-; WIN32-NEXT: movl %esi, %edi
-; WIN32-NEXT: fldz
-; WIN32-NEXT: LBB2_28:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: cmpl $127, %esi
-; WIN32-NEXT: jg LBB2_30
-; WIN32-NEXT: # %bb.29:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: movl %edi, %edx
-; WIN32-NEXT: fldz
-; WIN32-NEXT: fxch %st(2)
-; WIN32-NEXT: LBB2_30:
-; WIN32-NEXT: fstp %st(2)
-; WIN32-NEXT: cmpl $127, %eax
-; WIN32-NEXT: jg LBB2_32
-; WIN32-NEXT: # %bb.31:
-; WIN32-NEXT: fstp %st(2)
-; WIN32-NEXT: fldz
-; WIN32-NEXT: LBB2_32:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: shll $23, %ecx
-; WIN32-NEXT: addl $1065353216, %ecx # imm = 0x3F800000
-; WIN32-NEXT: movl %ecx, (%esp)
-; WIN32-NEXT: shll $23, %edx
-; WIN32-NEXT: addl $1065353216, %edx # imm = 0x3F800000
-; WIN32-NEXT: movl %edx, {{[0-9]+}}(%esp)
-; WIN32-NEXT: fxch %st(1)
-; WIN32-NEXT: fmuls (%esp)
-; WIN32-NEXT: fxch %st(1)
-; WIN32-NEXT: fmuls {{[0-9]+}}(%esp)
-; WIN32-NEXT: addl $8, %esp
+; WIN32-NEXT: fstpl (%esp)
+; WIN32-NEXT: calll _ldexp
+; WIN32-NEXT: movl %esi, {{[0-9]+}}(%esp)
+; WIN32-NEXT: flds {{[0-9]+}}(%esp)
+; WIN32-NEXT: fstpl (%esp)
+; WIN32-NEXT: fstps {{[0-9]+}}(%esp)
+; WIN32-NEXT: calll _ldexp
+; WIN32-NEXT: fstps {{[0-9]+}}(%esp)
+; WIN32-NEXT: flds {{[0-9]+}}(%esp)
+; WIN32-NEXT: flds {{[0-9]+}}(%esp)
+; WIN32-NEXT: addl $20, %esp
; WIN32-NEXT: popl %esi
-; WIN32-NEXT: popl %edi
; WIN32-NEXT: retl
%1 = call <2 x float> @llvm.ldexp.v2f32.v2i32(<2 x float> %val, <2 x i32> %exp)
ret <2 x float> %1
@@ -319,335 +189,106 @@ define <4 x float> @ldexp_v4f32(<4 x float> %val, <4 x i32> %exp) {
; X64-NEXT: .cfi_def_cfa_offset 8
; X64-NEXT: retq
;
+; WIN64-LABEL: ldexp_v4f32:
+; WIN64: # %bb.0:
+; WIN64-NEXT: pushq %rsi
+; WIN64-NEXT: .seh_pushreg %rsi
+; WIN64-NEXT: subq $80, %rsp
+; WIN64-NEXT: .seh_stackalloc 80
+; WIN64-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; WIN64-NEXT: .seh_savexmm %xmm8, 64
+; WIN64-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; WIN64-NEXT: .seh_savexmm %xmm7, 48
+; WIN64-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; WIN64-NEXT: .seh_savexmm %xmm6, 32
+; WIN64-NEXT: .seh_endprologue
+; WIN64-NEXT: movq %rdx, %rsi
+; WIN64-NEXT: movaps (%rcx), %xmm7
+; WIN64-NEXT: movl 12(%rdx), %edx
+; WIN64-NEXT: movaps %xmm7, %xmm0
+; WIN64-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm7[3,3]
+; WIN64-NEXT: cvtss2sd %xmm0, %xmm0
+; WIN64-NEXT: callq ldexp
+; WIN64-NEXT: xorps %xmm6, %xmm6
+; WIN64-NEXT: cvtsd2ss %xmm0, %xmm6
+; WIN64-NEXT: movl 8(%rsi), %edx
+; WIN64-NEXT: movaps %xmm7, %xmm0
+; WIN64-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm7[1]
+; WIN64-NEXT: cvtss2sd %xmm0, %xmm0
+; WIN64-NEXT: callq ldexp
+; WIN64-NEXT: xorps %xmm8, %xmm8
+; WIN64-NEXT: cvtsd2ss %xmm0, %xmm8
+; WIN64-NEXT: unpcklps {{.*#+}} xmm8 = xmm8[0],xmm6[0],xmm8[1],xmm6[1]
+; WIN64-NEXT: movl (%rsi), %edx
+; WIN64-NEXT: movl 4(%rsi), %esi
+; WIN64-NEXT: xorps %xmm0, %xmm0
+; WIN64-NEXT: cvtss2sd %xmm7, %xmm0
+; WIN64-NEXT: callq ldexp
+; WIN64-NEXT: xorps %xmm6, %xmm6
+; WIN64-NEXT: cvtsd2ss %xmm0, %xmm6
+; WIN64-NEXT: shufps {{.*#+}} xmm7 = xmm7[1,1,1,1]
+; WIN64-NEXT: xorps %xmm0, %xmm0
+; WIN64-NEXT: cvtss2sd %xmm7, %xmm0
+; WIN64-NEXT: movl %esi, %edx
+; WIN64-NEXT: callq ldexp
+; WIN64-NEXT: cvtsd2ss %xmm0, %xmm0
+; WIN64-NEXT: unpcklps {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1]
+; WIN64-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm8[0]
+; WIN64-NEXT: movaps %xmm6, %xmm0
+; WIN64-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; WIN64-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; WIN64-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
+; WIN64-NEXT: addq $80, %rsp
+; WIN64-NEXT: popq %rsi
+; WIN64-NEXT: retq
+; WIN64-NEXT: .seh_endproc
+;
; WIN32-LABEL: ldexp_v4f32:
; WIN32: # %bb.0:
; WIN32-NEXT: pushl %ebp
; WIN32-NEXT: pushl %ebx
; WIN32-NEXT: pushl %edi
; WIN32-NEXT: pushl %esi
-; WIN32-NEXT: subl $32, %esp
+; WIN32-NEXT: subl $44, %esp
+; WIN32-NEXT: movl {{[0-9]+}}(%esp), %esi
+; WIN32-NEXT: movl {{[0-9]+}}(%esp), %edi
+; WIN32-NEXT: movl {{[0-9]+}}(%esp), %ebx
+; WIN32-NEXT: movl {{[0-9]+}}(%esp), %ebp
+; WIN32-NEXT: movl {{[0-9]+}}(%esp), %eax
+; WIN32-NEXT: movl %eax, {{[0-9]+}}(%esp)
; WIN32-NEXT: flds {{[0-9]+}}(%esp)
-; WIN32-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; WIN32-NEXT: fstpl (%esp)
+; WIN32-NEXT: calll _ldexp
+; WIN32-NEXT: fstpl {{[-0-9]+}}(%e{{[sb]}}p) # 8-byte Folded Spill
+; WIN32-NEXT: movl %ebp, {{[0-9]+}}(%esp)
; WIN32-NEXT: flds {{[0-9]+}}(%esp)
-; WIN32-NEXT: flds __real@7f000000
-; WIN32-NEXT: fld %st(1)
-; WIN32-NEXT: fmul %st(1), %st
-; WIN32-NEXT: fld %st(0)
-; WIN32-NEXT: fmul %st(2), %st
-; WIN32-NEXT: cmpl $255, %ecx
-; WIN32-NEXT: jae LBB3_2
-; WIN32-NEXT: # %bb.1:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: fldz
-; WIN32-NEXT: fxch %st(1)
-; WIN32-NEXT: LBB3_2:
-; WIN32-NEXT: fstp %st(1)
-; WIN32-NEXT: cmpl $-329, %ecx # imm = 0xFEB7
-; WIN32-NEXT: movl %ecx, %esi
-; WIN32-NEXT: jge LBB3_4
-; WIN32-NEXT: # %bb.3:
-; WIN32-NEXT: movl $-330, %esi # imm = 0xFEB6
-; WIN32-NEXT: LBB3_4:
-; WIN32-NEXT: addl $204, %esi
-; WIN32-NEXT: leal 102(%ecx), %eax
-; WIN32-NEXT: cmpl $-228, %ecx
-; WIN32-NEXT: jb LBB3_6
-; WIN32-NEXT: # %bb.5:
-; WIN32-NEXT: movl %eax, %esi
-; WIN32-NEXT: LBB3_6:
-; WIN32-NEXT: flds __real@0c800000
-; WIN32-NEXT: fld %st(3)
-; WIN32-NEXT: fmul %st(1), %st
-; WIN32-NEXT: fld %st(0)
-; WIN32-NEXT: fmul %st(2), %st
-; WIN32-NEXT: jb LBB3_8
-; WIN32-NEXT: # %bb.7:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: fldz
-; WIN32-NEXT: fxch %st(1)
-; WIN32-NEXT: LBB3_8:
-; WIN32-NEXT: fstp %st(1)
-; WIN32-NEXT: cmpl $-126, %ecx
-; WIN32-NEXT: jl LBB3_10
-; WIN32-NEXT: # %bb.9:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: fldz
-; WIN32-NEXT: fxch %st(4)
-; WIN32-NEXT: LBB3_10:
-; WIN32-NEXT: fstp %st(4)
-; WIN32-NEXT: movl {{[0-9]+}}(%esp), %edi
-; WIN32-NEXT: movl %ecx, %edx
-; WIN32-NEXT: subl $127, %edx
-; WIN32-NEXT: jg LBB3_12
-; WIN32-NEXT: # %bb.11:
-; WIN32-NEXT: fstp %st(1)
-; WIN32-NEXT: fldz
-; WIN32-NEXT: fxch %st(3)
-; WIN32-NEXT: fxch %st(1)
-; WIN32-NEXT: LBB3_12:
-; WIN32-NEXT: fstp %st(3)
-; WIN32-NEXT: fld %st(3)
-; WIN32-NEXT: fmul %st(2), %st
-; WIN32-NEXT: fld %st(0)
-; WIN32-NEXT: fmul %st(3), %st
-; WIN32-NEXT: cmpl $255, %edi
-; WIN32-NEXT: jae LBB3_14
-; WIN32-NEXT: # %bb.13:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: fldz
-; WIN32-NEXT: fxch %st(1)
-; WIN32-NEXT: LBB3_14:
-; WIN32-NEXT: fstp %st(1)
-; WIN32-NEXT: fxch %st(1)
-; WIN32-NEXT: fstps {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
-; WIN32-NEXT: cmpl $-329, %edi # imm = 0xFEB7
-; WIN32-NEXT: movl %edi, %eax
-; WIN32-NEXT: jge LBB3_16
-; WIN32-NEXT: # %bb.15:
-; WIN32-NEXT: movl $-330, %eax # imm = 0xFEB6
-; WIN32-NEXT: LBB3_16:
-; WIN32-NEXT: fld %st(3)
-; WIN32-NEXT: fmul %st(3), %st
-; WIN32-NEXT: fld %st(0)
-; WIN32-NEXT: fmul %st(4), %st
-; WIN32-NEXT: cmpl $-228, %edi
-; WIN32-NEXT: jb LBB3_17
-; WIN32-NEXT: # %bb.18:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: leal 102(%edi), %eax
-; WIN32-NEXT: cmpl $-126, %edi
-; WIN32-NEXT: jge LBB3_20
-; WIN32-NEXT: jmp LBB3_21
-; WIN32-NEXT: LBB3_17:
-; WIN32-NEXT: fstp %st(1)
-; WIN32-NEXT: addl $204, %eax
-; WIN32-NEXT: cmpl $-126, %edi
-; WIN32-NEXT: jl LBB3_21
-; WIN32-NEXT: LBB3_20:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: movl %edi, %eax
-; WIN32-NEXT: fldz
-; WIN32-NEXT: fxch %st(4)
-; WIN32-NEXT: LBB3_21:
-; WIN32-NEXT: fstp %st(4)
-; WIN32-NEXT: movl %eax, (%esp) # 4-byte Spill
-; WIN32-NEXT: movl %edi, %ebx
-; WIN32-NEXT: subl $127, %ebx
-; WIN32-NEXT: jg LBB3_23
-; WIN32-NEXT: # %bb.22:
-; WIN32-NEXT: fstp %st(0)
-; WIN32-NEXT: fldz
-; WIN32-NEXT: fxch %st(3)
-; WIN32-NEXT: LBB3_23:
-; WIN32-NEXT: fstp %st(3)
-; WIN32-NEXT: cmpl $381, %edi # imm = 0x17D
-; WIN32-NEXT: movl %edi, %eax
-; WIN32-NEXT: jge LBB3_24
-; WIN32-NEXT: # %bb.25:
-; WIN32-NEXT: cmpl $255, %edi
-; WIN32-NEXT: jae LBB3_26
-; WIN32-NEXT: LBB3_27:
-; WIN32-NEXT: cmpl $-126, %ecx
-; WIN32-NEXT: jl LBB3_29
-; WIN32-NEXT: LBB3_28:
-; WIN32-NEXT: movl %ecx, %esi
-; WIN32-NEXT: LBB3_29:
-; WIN32-NEXT: cmpl $381, %ecx # imm = 0x17D
-; WIN32-NEXT: movl %ecx, %eax
-; WIN32-NEXT: jl LBB3_31
-; WIN32-NEXT: # %bb.30:
-; WIN32-NEXT: movl $381, %eax # imm = 0x17D
-; WIN32-NEXT: LBB3_31:
-; WIN32-NEXT: cmpl $255, %ecx
+; WIN32-NEXT: fstpl (%esp)
+; WIN32-NEXT: calll _ldexp
+; WIN32-NEXT: fstpl {{[-0-9]+}}(%e{{[sb]}}p) # 8-byte Folded Spill
+; WIN32-NEXT: movl %ebx, {{[0-9]+}}(%esp)
; WIN32-NEXT: flds {{[0-9]+}}(%esp)
-; WIN32-NEXT: movl {{[0-9]+}}(%esp), %ebp
-; WIN32-NEXT: jb LBB3_33
-; WIN32-NEXT: # %bb.32:
-; WIN32-NEXT: addl $-254, %eax
-; WIN32-NEXT: movl %eax, %edx
-; WIN32-NEXT: LBB3_33:
-; WIN32-NEXT: fxc...
[truncated]
|
arsenm
reviewed
Jun 12, 2024
|
Updated patch to mark "promote". I don't know of any way to do that cross-target, so this is now done per-target; not sure if this is what you expected. |
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
arsenm
approved these changes
Jun 25, 2024
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/30/builds/716 Here is the relevant piece of the build log for the reference: |
AlexisPerry
pushed a commit
to llvm-project-tlp/llvm-project
that referenced
this pull request
Jul 9, 2024
This reduces codesize. As discussed in llvm#92707.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This reduces codesize. As discussed in #92707.