[PATCH v1 20/27] x86/fpu: Optimize svml_s_log10f8_core_avx2.S

[Noah Goldstein <goldstein.w.n@gmail.com>]

1. Improve special values case which ends up covering ~half of all
   float bit patterns.
2. Cleanup some missed optimizations in instruction selection /
   unnecissary repeated rodata references.
3. Remove unused rodata.
4. Use common data definitions where possible.

Code Size Change: -43 Bytes (268 - 311)

Input                                 New Time / Old Time
0F          (0x00000000)           -> 0.9483
0F          (0x0000ffff, Denorm)   -> 0.9718
.1F         (0x3dcccccd)           -> 0.9253
5F          (0x40a00000)           -> 0.9218
2315255808F (0x4f0a0000)           -> 0.9199
-NaN        (0xffffffff)           -> 0.8258
---
 .../fpu/multiarch/svml_s_log10f8_core_avx2.S  | 316 ++++++++----------
 1 file changed, 132 insertions(+), 184 deletions(-)

diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_log10f8_core_avx2.S b/sysdeps/x86_64/fpu/multiarch/svml_s_log10f8_core_avx2.S
index 4bdc62e90e..2e9db34f23 100644
--- a/sysdeps/x86_64/fpu/multiarch/svml_s_log10f8_core_avx2.S
+++ b/sysdeps/x86_64/fpu/multiarch/svml_s_log10f8_core_avx2.S
@@ -27,216 +27,164 @@
  *
  */
 
-/* Offsets for data table __svml_slog10_data_internal
- */
-#define MinNorm				0
-#define MaxNorm				32
-#define L2H				64
-#define L2L				96
-#define iBrkValue			128
-#define iOffExpoMask			160
-#define One				192
-#define sPoly				224
-#define L2				512
+#define LOCAL_DATA_NAME	__svml_slog10_data_internal
+#include "svml_s_common_avx2_rodata_offsets.h"
+
+/* Offsets for data table __svml_slog10_data_internal.  */
+#define _Coeff_9	0
+#define _Coeff_8	32
+#define _Coeff_7	64
+#define _Coeff_6	96
+#define _Coeff_5	128
+#define _Coeff_4	160
+#define _Coeff_3	192
+#define _Coeff_2	224
+#define _Coeff_1	256
+#define _L2L	288
+#define _L2H	320
 
 #include <sysdep.h>
 
 	.section .text.avx2, "ax", @progbits
 ENTRY(_ZGVdN8v_log10f_avx2)
-	pushq	%rbp
-	cfi_def_cfa_offset(16)
-	movq	%rsp, %rbp
-	cfi_def_cfa(6, 16)
-	cfi_offset(6, -16)
-	andq	$-32, %rsp
-	subq	$96, %rsp
-
-	/* reduction: compute r, n */
-	vmovups	iBrkValue+__svml_slog10_data_internal(%rip), %ymm4
-	vmovups	sPoly+__svml_slog10_data_internal(%rip), %ymm15
-	vmovups	sPoly+64+__svml_slog10_data_internal(%rip), %ymm9
-	vmovups	sPoly+128+__svml_slog10_data_internal(%rip), %ymm10
-	vmovups	sPoly+192+__svml_slog10_data_internal(%rip), %ymm12
+	/* reduction: compute r, n.  */
+	vmovups	COMMON_DATA(_IBrkValue)(%rip), %ymm4
 	vpsubd	%ymm4, %ymm0, %ymm1
-	vcmplt_oqps MinNorm+__svml_slog10_data_internal(%rip), %ymm0, %ymm5
-	vcmpnle_uqps MaxNorm+__svml_slog10_data_internal(%rip), %ymm0, %ymm6
-	vpand	iOffExpoMask+__svml_slog10_data_internal(%rip), %ymm1, %ymm3
+	vmovups	COMMON_DATA(_NotiOffExpoMask)(%rip), %ymm7
+	vpandn	%ymm1, %ymm7, %ymm3
 	vpsrad	$23, %ymm1, %ymm2
-	vpaddd	%ymm4, %ymm3, %ymm8
+
+	vpsubd	%ymm7, %ymm0, %ymm5
+	vmovups	COMMON_DATA(_ILoRange)(%rip), %ymm7
+	vpcmpgtd %ymm5, %ymm7, %ymm7
+
+
+	vpaddd	%ymm4, %ymm3, %ymm5
 	vcvtdq2ps %ymm2, %ymm1
-	vsubps	One+__svml_slog10_data_internal(%rip), %ymm8, %ymm13
-	vmulps	L2L+__svml_slog10_data_internal(%rip), %ymm1, %ymm14
-	vfmadd213ps sPoly+32+__svml_slog10_data_internal(%rip), %ymm13, %ymm15
-	vfmadd213ps sPoly+96+__svml_slog10_data_internal(%rip), %ymm13, %ymm9
-	vmulps	%ymm13, %ymm13, %ymm11
-	vfmadd213ps sPoly+160+__svml_slog10_data_internal(%rip), %ymm13, %ymm10
-	vfmadd213ps sPoly+224+__svml_slog10_data_internal(%rip), %ymm13, %ymm12
-	vfmadd213ps %ymm9, %ymm11, %ymm15
-	vfmadd213ps %ymm10, %ymm11, %ymm15
-	vfmadd213ps %ymm12, %ymm11, %ymm15
-	vfmadd213ps sPoly+256+__svml_slog10_data_internal(%rip), %ymm13, %ymm15
-	vfmadd213ps %ymm14, %ymm13, %ymm15
-	vorps	%ymm6, %ymm5, %ymm7
-
-	/* combine and get argument value range mask */
+	vsubps	COMMON_DATA(_OneF)(%rip), %ymm5, %ymm5
+
+	vmovups	LOCAL_DATA(_Coeff_9)(%rip), %ymm2
+	vfmadd213ps LOCAL_DATA(_Coeff_8)(%rip), %ymm5, %ymm2
+	vmovups	LOCAL_DATA(_Coeff_7)(%rip), %ymm3
+	vfmadd213ps LOCAL_DATA(_Coeff_6)(%rip), %ymm5, %ymm3
+	vmulps	%ymm5, %ymm5, %ymm4
 	vmovmskps %ymm7, %edx
-	vfmadd132ps L2H+__svml_slog10_data_internal(%rip), %ymm15, %ymm1
-	testl	%edx, %edx
+	vmovups	LOCAL_DATA(_Coeff_5)(%rip), %ymm7
+	vfmadd213ps LOCAL_DATA(_Coeff_4)(%rip), %ymm5, %ymm7
+	vmovups	LOCAL_DATA(_Coeff_3)(%rip), %ymm6
+	vfmadd213ps LOCAL_DATA(_Coeff_2)(%rip), %ymm5, %ymm6
+	vfmadd213ps %ymm3, %ymm4, %ymm2
+	vfmadd213ps %ymm7, %ymm4, %ymm2
+	vfmadd213ps %ymm6, %ymm4, %ymm2
+	vfmadd213ps LOCAL_DATA(_Coeff_1)(%rip), %ymm5, %ymm2
+	vmulps	LOCAL_DATA(_L2L)(%rip), %ymm1, %ymm7
+	vfmadd213ps %ymm7, %ymm5, %ymm2
+
 
-	/* Go to special inputs processing branch */
+
+	vfmadd132ps LOCAL_DATA(_L2H)(%rip), %ymm2, %ymm1
+	testl	%edx, %edx
+	/* Go to special inputs processing branch.  */
 	jne	L(SPECIAL_VALUES_BRANCH)
-	# LOE rbx r12 r13 r14 r15 edx ymm0 ymm1
 
-	/* Restore registers
-	 * and exit the function
-	 */
 
-L(EXIT):
+	/* Restore registers * and exit the function.  */
 	vmovaps	%ymm1, %ymm0
-	movq	%rbp, %rsp
-	popq	%rbp
-	cfi_def_cfa(7, 8)
-	cfi_restore(6)
 	ret
-	cfi_def_cfa(6, 16)
-	cfi_offset(6, -16)
-
-	/* Branch to process
-	 * special inputs
-	 */
 
+	/* Cold case. edx has 1s where there was a special value that
+	   needs to be handled by a atanhf call. Optimize for code size
+	   more so than speed here.  */
 L(SPECIAL_VALUES_BRANCH):
-	vmovups	%ymm0, 32(%rsp)
-	vmovups	%ymm1, 64(%rsp)
-	# LOE rbx r12 r13 r14 r15 edx ymm1
+	/* Use r13 to save/restore the stack. This allows us to use rbp
+	   as callee save register saving code size.  */
+	pushq	%r13
+	cfi_adjust_cfa_offset (8)
+	cfi_offset (r13, -16)
+	/* Need to callee save registers to preserve state across tanhf
+	   calls.  */
+	pushq	%rbx
+	cfi_adjust_cfa_offset (8)
+	cfi_offset (rbx, -24)
+	pushq	%rbp
+	cfi_adjust_cfa_offset (8)
+	cfi_offset (rbp, -32)
+	movq	%rsp, %r13
+	cfi_def_cfa_register (r13)
 
-	xorl	%eax, %eax
-	# LOE rbx r12 r13 r14 r15 eax edx
+	/* Align stack and make room for 2x ymm vectors.  */
+	andq	$-32, %rsp
+	addq	$-64, %rsp
+
+	/* Save all already computed inputs.  */
+	vmovups	%ymm1, (%rsp)
+	/* Save original input (ymm0 unchanged up to this point).  */
+	vmovups	%ymm0, 32(%rsp)
 
 	vzeroupper
-	movq	%r12, 16(%rsp)
-	/*  DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -80; DW_OP_plus)  */
-	.cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xb0, 0xff, 0xff, 0xff, 0x22
-	movl	%eax, %r12d
-	movq	%r13, 8(%rsp)
-	/*  DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -88; DW_OP_plus)  */
-	.cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xa8, 0xff, 0xff, 0xff, 0x22
-	movl	%edx, %r13d
-	movq	%r14, (%rsp)
-	/*  DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -96; DW_OP_plus)  */
-	.cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xa0, 0xff, 0xff, 0xff, 0x22
-	# LOE rbx r15 r12d r13d
-
-	/* Range mask
-	 * bits check
-	 */
-
-L(RANGEMASK_CHECK):
-	btl	%r12d, %r13d
-
-	/* Call scalar math function */
-	jc	L(SCALAR_MATH_CALL)
-	# LOE rbx r15 r12d r13d
-
-	/* Special inputs
-	 * processing loop
-	 */
 
+	/* edx has 1s where there was a special value that needs to be
+	   handled by a atanhf call.  */
+	movl	%edx, %ebx
 L(SPECIAL_VALUES_LOOP):
-	incl	%r12d
-	cmpl	$8, %r12d
-
-	/* Check bits in range mask */
-	jl	L(RANGEMASK_CHECK)
-	# LOE rbx r15 r12d r13d
-
-	movq	16(%rsp), %r12
-	cfi_restore(12)
-	movq	8(%rsp), %r13
-	cfi_restore(13)
-	movq	(%rsp), %r14
-	cfi_restore(14)
-	vmovups	64(%rsp), %ymm1
-
-	/* Go to exit */
-	jmp	L(EXIT)
-	/*  DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -80; DW_OP_plus)  */
-	.cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xb0, 0xff, 0xff, 0xff, 0x22
-	/*  DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -88; DW_OP_plus)  */
-	.cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xa8, 0xff, 0xff, 0xff, 0x22
-	/*  DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -96; DW_OP_plus)  */
-	.cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xa0, 0xff, 0xff, 0xff, 0x22
-	# LOE rbx r12 r13 r14 r15 ymm1
-
-	/* Scalar math fucntion call
-	 * to process special input
-	 */
-
-L(SCALAR_MATH_CALL):
-	movl	%r12d, %r14d
-	vmovss	32(%rsp, %r14, 4), %xmm0
+
+	/* use rbp as index for special value that is saved across calls
+	   to atanhf. We technically don't need a callee save register
+	   here as offset to rsp is always [0, 28] so we can restore
+	   rsp by realigning to 64. Essentially the tradeoff is 1 extra
+	   save/restore vs 2 extra instructions in the loop. Realigning
+	   also costs more code size.  */
+	xorl	%ebp, %ebp
+	tzcntl	%ebx, %ebp
+
+	/* Scalar math fucntion call to process special input.  */
+	vmovss	32(%rsp, %rbp, 4), %xmm0
 	call	log10f@PLT
-	# LOE rbx r14 r15 r12d r13d xmm0
 
-	vmovss	%xmm0, 64(%rsp, %r14, 4)
+	/* No good way to avoid the store-forwarding fault this will
+	   cause on return. `lfence` avoids the SF fault but at greater
+	   cost as it serialized stack/callee save restoration.  */
+	vmovss	%xmm0, (%rsp, %rbp, 4)
+
+	blsrl	%ebx, %ebx
+	jnz	L(SPECIAL_VALUES_LOOP)
+
 
-	/* Process special inputs in loop */
-	jmp	L(SPECIAL_VALUES_LOOP)
-	# LOE rbx r15 r12d r13d
+
+	/* All results have been written to (%rsp).  */
+	vmovups	(%rsp), %ymm0
+	/* Restore rsp.  */
+	movq	%r13, %rsp
+	cfi_def_cfa_register (rsp)
+	/* Restore callee save registers.  */
+	popq	%rbp
+	cfi_adjust_cfa_offset (-8)
+	cfi_restore (rbp)
+	popq	%rbx
+	cfi_adjust_cfa_offset (-8)
+	cfi_restore (rbp)
+	popq	%r13
+	cfi_adjust_cfa_offset (-8)
+	cfi_restore (r13)
+	ret
 END(_ZGVdN8v_log10f_avx2)
 
-	.section .rodata, "a"
+	.section .rodata.avx2, "a"
 	.align	32
 
-#ifdef __svml_slog10_data_internal_typedef
-typedef unsigned int VUINT32;
-typedef struct {
-	__declspec(align(32)) VUINT32 MinNorm[8][1];
-	__declspec(align(32)) VUINT32 MaxNorm[8][1];
-	__declspec(align(32)) VUINT32 L2H[8][1];
-	__declspec(align(32)) VUINT32 L2L[8][1];
-	__declspec(align(32)) VUINT32 iBrkValue[8][1];
-	__declspec(align(32)) VUINT32 iOffExpoMask[8][1];
-	__declspec(align(32)) VUINT32 One[8][1];
-	__declspec(align(32)) VUINT32 sPoly[9][8][1];
-	__declspec(align(32)) VUINT32 L2[8][1];
-} __svml_slog10_data_internal;
-#endif
-__svml_slog10_data_internal:
-	/* MinNorm */
-	.long	0x00800000, 0x00800000, 0x00800000, 0x00800000, 0x00800000, 0x00800000, 0x00800000, 0x00800000
-	/* MaxNorm */
-	.align	32
-	.long	0x7f7fffff, 0x7f7fffff, 0x7f7fffff, 0x7f7fffff, 0x7f7fffff, 0x7f7fffff, 0x7f7fffff, 0x7f7fffff
-	/* L2H */
-	.align	32
-	.long	0x3e9a2100, 0x3e9a2100, 0x3e9a2100, 0x3e9a2100, 0x3e9a2100, 0x3e9a2100, 0x3e9a2100, 0x3e9a2100
-	/* L2L */
-	.align	32
-	.long	0xb64AF600, 0xb64AF600, 0xb64AF600, 0xb64AF600, 0xb64AF600, 0xb64AF600, 0xb64AF600, 0xb64AF600
-	/* iBrkValue = SP 2/3 */
-	.align	32
-	.long	0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab, 0x3f2aaaab
-	/* iOffExpoMask = SP significand mask */
-	.align	32
-	.long	0x007fffff, 0x007fffff, 0x007fffff, 0x007fffff, 0x007fffff, 0x007fffff, 0x007fffff, 0x007fffff
-	/* sOne = SP 1.0 */
-	.align	32
-	.long	0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000
-	/* spoly[9] */
-	.align	32
-	.long	0x3d8063B4, 0x3d8063B4, 0x3d8063B4, 0x3d8063B4, 0x3d8063B4, 0x3d8063B4, 0x3d8063B4, 0x3d8063B4 /* coeff9 */
-	.long	0xbd890073, 0xbd890073, 0xbd890073, 0xbd890073, 0xbd890073, 0xbd890073, 0xbd890073, 0xbd890073 /* coeff8 */
-	.long	0x3d775317, 0x3d775317, 0x3d775317, 0x3d775317, 0x3d775317, 0x3d775317, 0x3d775317, 0x3d775317 /* coeff7 */
-	.long	0xbd91FB27, 0xbd91FB27, 0xbd91FB27, 0xbd91FB27, 0xbd91FB27, 0xbd91FB27, 0xbd91FB27, 0xbd91FB27 /* coeff6 */
-	.long	0x3dB20B96, 0x3dB20B96, 0x3dB20B96, 0x3dB20B96, 0x3dB20B96, 0x3dB20B96, 0x3dB20B96, 0x3dB20B96 /* coeff5 */
-	.long	0xbdDE6E20, 0xbdDE6E20, 0xbdDE6E20, 0xbdDE6E20, 0xbdDE6E20, 0xbdDE6E20, 0xbdDE6E20, 0xbdDE6E20 /* coeff4 */
-	.long	0x3e143CE5, 0x3e143CE5, 0x3e143CE5, 0x3e143CE5, 0x3e143CE5, 0x3e143CE5, 0x3e143CE5, 0x3e143CE5 /* coeff3 */
-	.long	0xbe5E5BC5, 0xbe5E5BC5, 0xbe5E5BC5, 0xbe5E5BC5, 0xbe5E5BC5, 0xbe5E5BC5, 0xbe5E5BC5, 0xbe5E5BC5 /* coeff2 */
-	.long	0x3eDE5BD9, 0x3eDE5BD9, 0x3eDE5BD9, 0x3eDE5BD9, 0x3eDE5BD9, 0x3eDE5BD9, 0x3eDE5BD9, 0x3eDE5BD9 /* coeff1 */
-	/* L2 */
-	.align	32
-	.long	0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b, 0x3e9a209b
-	.align	32
-	.type	__svml_slog10_data_internal, @object
-	.size	__svml_slog10_data_internal, .-__svml_slog10_data_internal
+LOCAL_DATA_NAME:
+
+	DATA_VEC (LOCAL_DATA_NAME, _Coeff_9, 0x3d8063b4)
+	DATA_VEC (LOCAL_DATA_NAME, _Coeff_8, 0xbd890073)
+	DATA_VEC (LOCAL_DATA_NAME, _Coeff_7, 0x3d775317)
+	DATA_VEC (LOCAL_DATA_NAME, _Coeff_6, 0xbd91fb27)
+	DATA_VEC (LOCAL_DATA_NAME, _Coeff_5, 0x3db20b96)
+	DATA_VEC (LOCAL_DATA_NAME, _Coeff_4, 0xbdde6e20)
+	DATA_VEC (LOCAL_DATA_NAME, _Coeff_3, 0x3e143ce5)
+	DATA_VEC (LOCAL_DATA_NAME, _Coeff_2, 0xbe5e5bc5)
+	DATA_VEC (LOCAL_DATA_NAME, _Coeff_1, 0x3ede5bd9)
+	DATA_VEC (LOCAL_DATA_NAME, _L2L, 0xb64af600)
+	DATA_VEC (LOCAL_DATA_NAME, _L2H, 0x3e9a2100)
+
+	.type	LOCAL_DATA_NAME, @object
+	.size	LOCAL_DATA_NAME, .-LOCAL_DATA_NAME
-- 
2.34.1