Hi Bohan, I observed ~5% regression on several libm functions on SpacemiT K3. And rebuilding with `-falign-functions=16`, the regression drops to ~1% (the affected functions are not exactly the same).
-2.40% bench-modff modff / workload-integral / latency -1.50% bench-sinhf sinhf / workload-random / reciprocal-throughput -1.32% bench-modf modf / workload-0_1 / reciprocal-throughput -1.31% bench-modf modf / workload-0_1 / latency -1.26% bench-sinhf sinhf / workload-random / latency -1.22% bench-log2f log2f / workload-spec2017.wrf / reciprocal-throughput -1.20% bench-fmodf fmodf / normal / mean -1.19% bench-coshf coshf / workload-random / latency -1.14% bench-ilogbf ilogbf / normal / mean -1.08% bench-log2f log2f / workload-spec2017.wrf / latency I think this is performance jitter caused by address layout changes, and suggest ignoring this K3-specific regression for this patch and moving forward with the merge. Thanks. > From: "Bohan Lei"<[email protected]> > Date: Wed, Nov 11, 2026, 10:16 > Subject: [PATCH v2] RISC-V: Allow IRA to allocate FPR to integer modes > To: <[email protected]> > Cc: "Jeff Law"<[email protected]>, "Bohan > Lei"<[email protected]>, "Zhijin Zeng"<[email protected]> > This is based on Zhijin Zeng's v3 patch > (https://gcc.gnu.org/pipermail/gcc-patches/2025-January/673406.html). > This commit modifies riscv_can_change_mode_class and > riscv_hard_regno_mode_ok to allow IRA to allocate an FPR to a pseudo > that is used in a interger mode in some places. > > A typical scenario would be cases with bit-level uses of floating-point > variables. The current IRA can only allocate them in GPRs, and any FP > use of it would create a new FPR reload, which can lead to lots of > redundant fmv.d.x instructions in the function. The attached case > (LLM-generated) is a reduced one that can still show the underlying > issue. > > I am not using Zhijin's original testcase in this version because > register renaming has accidentally helped eliminate the redundant fmv > instructions in that case. > > Also note that the tiebility change in that patch is not included in > this version. We found that change can lead to more regressions and > would require special handling. Nonetheless, we found one redundant > fmv.x.d instruction in our complete lgamma case downstream, not verified > with the trunk though, which can be eliminated with the tieability > change. (After all, FP and integer modes being tieable makes it hard > track separate live ranges of integer uses and FP uses of the same > variable.) We may submit a separate tieability patch after gathering > enough evidence that this could lead to real performance issues. > > The earlier version failed to handle xtheadfmemidx correctly, and this > version contains a fix for that route. A local Sashiko review found a > preexisting issue in th_memidx_classify_address_index, for which after > confirming I will send a separate patch as it is not directly related to > this. > > Regtested on rv64gcv and rv32gcv, no regression. > > Assisted-by: Codex > Co-authored-by: Zhijin Zeng <[email protected]> > > gcc/ChangeLog: > > * config/riscv/constraints.md (th_m_fpr): New constraint for FPR > load/store. > * config/riscv/riscv.cc (riscv_hard_regno_mode_ok): Allow SImode > and DImode in GPRs. > (riscv_can_change_mode_class): Allow mode changing between > integer and FP modes. > * config/riscv/riscv.md: Add f,f constraints to SI and HI moves. > * config/riscv/thead.cc (th_fmemidx_output_index): Take into > account the scenario where MODE is integral. > (th_output_move): Remove instruction class conditions. > * config/riscv/vector.md: Restore X constraint for vr,vr LRA > move patterns to prevent ICE caused by FPR-allocated scratch. > > gcc/testsuite/ChangeLog: > > * gcc.target/riscv/fmv-lgamma.c: New test. > * gcc.target/riscv/xtheadmemidx-fpr-int-1.c: New test. > * gcc.target/riscv/xtheadmemidx-fpr-int-2.c: New test. > * gcc.target/riscv/xtheadmemidx-fpr-int.h: New test. > --- > gcc/config/riscv/constraints.md | 9 ++++++ > gcc/config/riscv/riscv.cc | 29 ++++++++++++++++++- > gcc/config/riscv/riscv.md | 28 +++++++++--------- > gcc/config/riscv/thead.cc | 21 +++++++++----- > gcc/config/riscv/vector.md | 6 ++-- > gcc/testsuite/gcc.target/riscv/fmv-lgamma.c | 20 +++++++++++++ > .../gcc.target/riscv/xtheadmemidx-fpr-int-1.c | 8 +++++ > .../gcc.target/riscv/xtheadmemidx-fpr-int-2.c | 7 +++++ > .../gcc.target/riscv/xtheadmemidx-fpr-int.h | 25 ++++++++++++++++ > 9 files changed, 128 insertions(+), 25 deletions(-) > create mode 100644 gcc/testsuite/gcc.target/riscv/fmv-lgamma.c > create mode 100644 gcc/testsuite/gcc.target/riscv/xtheadmemidx-fpr-int-1.c > create mode 100644 gcc/testsuite/gcc.target/riscv/xtheadmemidx-fpr-int-2.c > create mode 100644 gcc/testsuite/gcc.target/riscv/xtheadmemidx-fpr-int.h > > diff --git a/gcc/config/riscv/constraints.md b/gcc/config/riscv/constraints.md > index 666d5a246ce..73534cf7710 100644 > --- a/gcc/config/riscv/constraints.md > +++ b/gcc/config/riscv/constraints.md > @@ -354,6 +354,15 @@ (define_memory_constraint "th_m_noi" > (and (match_test "!th_memidx_legitimate_index_p (op, > false)") > (match_test "!th_memidx_legitimate_index_p (op, > true)")))))) > > +(define_memory_constraint "th_m_fpr" > + "@internal > + A memory operand usable by an FPR. XTheadMemIdx writeback addresses > + require a GPR, and indexed FPR accesses require XTheadFMemIdx." > + (and (match_operand 0 "memory_operand") > + (ior (match_test "satisfies_constraint_th__m__noi (op)") > + (and (match_test "TARGET_XTHEADFMEMIDX") > + (match_test "th_memidx_legitimate_index_p (op)"))))) > + > ;; CORE-V Constraints > (define_constraint "CV_alu_pow2" > "@internal > diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc > index 9f1f4df3ea6..ada8c706e1f 100644 > --- a/gcc/config/riscv/riscv.cc > +++ b/gcc/config/riscv/riscv.cc > @@ -11740,7 +11740,10 @@ riscv_hard_regno_mode_ok (unsigned int regno, > machine_mode mode) > return false; > > if (GET_MODE_CLASS (mode) != MODE_FLOAT > - && GET_MODE_CLASS (mode) != MODE_COMPLEX_FLOAT) > + && GET_MODE_CLASS (mode) != MODE_COMPLEX_FLOAT > + && !((mode == DImode && TARGET_DOUBLE_FLOAT && TARGET_64BIT) > + || (mode == SImode && TARGET_HARD_FLOAT) > + || (mode == HImode && TARGET_ZFHMIN))) > return false; > > /* Only use callee-saved registers if a potential callee is guaranteed > @@ -13378,6 +13381,30 @@ riscv_can_change_mode_class (machine_mode from, > machine_mode to, > && !ordered_p (GET_MODE_PRECISION (from), GET_MODE_PRECISION (to))) > return false; > > + if ((SCALAR_INT_MODE_P (to) || SCALAR_FLOAT_MODE_P (to)) > + && (rclass == FP_REGS || rclass == RVC_FP_REGS > + || rclass == GR_REGS || rclass == RVC_GR_REGS > + || rclass == JALR_REGS)) > + { > + /* fmv.x.w or fmv.w.x. */ > + if (TARGET_HARD_FLOAT) > + if ((from == SFmode && to == SImode) > + || (from == SImode && to == SFmode)) > + return true; > + > + /* fmv.x.d or fmv.d.x. */ > + if (TARGET_DOUBLE_FLOAT && TARGET_64BIT) > + if ((from == DFmode && to == DImode) > + || (from == DImode && to == DFmode)) > + return true; > + > + /* fmv.x.h or fmv.h.x. */ > + if (TARGET_ZFHMIN) > + if ((from == HFmode && to == HImode) > + || (from == HImode && to == HFmode)) > + return true; > + } > + > return !reg_classes_intersect_p (FP_REGS, rclass); > } > > diff --git a/gcc/config/riscv/riscv.md b/gcc/config/riscv/riscv.md > index 1fabe9260e4..2866e26f2c3 100644 > --- a/gcc/config/riscv/riscv.md > +++ b/gcc/config/riscv/riscv.md > @@ -2633,8 +2633,8 @@ (define_expand "movdi" > }) > > (define_insn "*movdi_32bit" > - [(set (match_operand:DI 0 "nonimmediate_operand" "=r,r,r, m, > *f,*f,*r,*f,*m,r") > - (match_operand:DI 1 "move_operand" " > r,i,m,rJ,*J*r,*m,*f,*f,*f,vp"))] > + [(set (match_operand:DI 0 "nonimmediate_operand" "=r,r,r, m, > *f,*f,*r,*f,*th_m_fpr,r") > + (match_operand:DI 1 "move_operand" " > r,i,m,rJ,*J*r,*th_m_fpr,*f,*f,*f,vp"))] > "!TARGET_64BIT > && (register_operand (operands[0], DImode) > || reg_or_0_operand (operands[1], DImode))" > @@ -2645,8 +2645,8 @@ (define_insn "*movdi_32bit" > (set_attr "ext" "base,base,base,base,d,d,d,d,d,vector")]) > > (define_insn "*movdi_64bit" > - [(set (match_operand:DI 0 "nonimmediate_operand" "=r,r,r, m, > *f,*f,*r,*f,*m,r") > - (match_operand:DI 1 "move_operand" " > r,T,m,rJ,*r*J,*m,*f,*f,*f,vp"))] > + [(set (match_operand:DI 0 "nonimmediate_operand" "=r,r,r, m, > *f,*f,*r,*f,*th_m_fpr,r") > + (match_operand:DI 1 "move_operand" " > r,T,m,rJ,*r*J,*th_m_fpr,*f,*f,*f,vp"))] > "TARGET_64BIT > && (register_operand (operands[0], DImode) > || reg_or_0_operand (operands[1], DImode))" > @@ -2679,16 +2679,16 @@ (define_expand "mov<mode>" > }) > > (define_insn "*movsi_internal" > - [(set (match_operand:SI 0 "nonimmediate_operand" "=r,r,r, m, > *f,*f,*r,*m,r") > - (match_operand:SI 1 "move_operand" " > r,T,m,rJ,*r*J,*m,*f,*f,vp"))] > + [(set (match_operand:SI 0 "nonimmediate_operand" "=r,r,r, m, > *f,*f,*r,*f,*th_m_fpr,r") > + (match_operand:SI 1 "move_operand" " > r,T,m,rJ,*r*J,*th_m_fpr,*f,*f,*f,vp"))] > "(register_operand (operands[0], SImode) > || reg_or_0_operand (operands[1], SImode)) > && !(REG_P (operands[1]) && VL_REG_P (REGNO (operands[1])))" > { return riscv_output_move (operands[0], operands[1]); } > - [(set_attr "move_type" > "move,const,load,store,mtc,fpload,mfc,fpstore,rdvlenb") > + [(set_attr "move_type" > "move,const,load,store,mtc,fpload,mfc,fmove,fpstore,rdvlenb") > (set_attr "mode" "SI") > - (set_attr "type" "move,move,load,store,mtc,fpload,mfc,fpstore,move") > - (set_attr "ext" "base,base,base,base,f,f,f,f,vector")]) > + (set_attr "type" "move,move,load,store,mtc,fpload,mfc,fmove,fpstore,move") > + (set_attr "ext" "base,base,base,base,f,f,f,f,f,vector")]) > > ;; 16-bit Integer moves > > @@ -2707,15 +2707,15 @@ (define_expand "movhi" > }) > > (define_insn "*movhi_internal" > - [(set (match_operand:HI 0 "nonimmediate_operand" "=r,r,r, m, *f,*r,r") > - (match_operand:HI 1 "move_operand" " > r,T,m,rJ,*r*J,*f,vp"))] > + [(set (match_operand:HI 0 "nonimmediate_operand" "=r,r,r, m, *f,*r,*f,r") > + (match_operand:HI 1 "move_operand" " > r,T,m,rJ,*r*J,*f,*f,vp"))] > "(register_operand (operands[0], HImode) > || reg_or_0_operand (operands[1], HImode))" > { return riscv_output_move (operands[0], operands[1]); } > - [(set_attr "move_type" "move,const,load,store,mtc,mfc,rdvlenb") > + [(set_attr "move_type" "move,const,load,store,mtc,mfc,fmove,rdvlenb") > (set_attr "mode" "HI") > - (set_attr "type" "move,move,load,store,mtc,mfc,move") > - (set_attr "ext" "base,base,base,base,f,f,vector")]) > + (set_attr "type" "move,move,load,store,mtc,mfc,fmove,move") > + (set_attr "ext" "base,base,base,base,f,f,f,vector")]) > > ;; HImode constant generation; see riscv_move_integer for details. > ;; si+si->hi without truncation is legal because of > diff --git a/gcc/config/riscv/thead.cc b/gcc/config/riscv/thead.cc > index 33d9dc28dd4..59d9f4b6590 100644 > --- a/gcc/config/riscv/thead.cc > +++ b/gcc/config/riscv/thead.cc > @@ -840,6 +840,16 @@ th_memidx_output_index (rtx dest, rtx src, machine_mode > mode, bool load) > static const char * > th_fmemidx_output_index (rtx dest, rtx src, machine_mode mode, bool load) > { > + /* Integer-mode values can also reside in FPRs. */ > + machine_mode fmode = mode; > + if (mode == SImode) > + fmode = SFmode; > + else if (mode == DImode) > + fmode = DFmode; > + > + if (!TARGET_XTHEADFMEMIDX || !is_fmemidx_mode (fmode)) > + return NULL; > + > struct riscv_address_info info; > char format[24]; > rtx output_operands[2]; > @@ -920,15 +930,14 @@ th_output_move (rtx dest, rtx src) > > if (dest_code == REG && src_code == MEM) > { > - if (GET_MODE_CLASS (mode) == MODE_INT > - || (GET_MODE_CLASS (mode) == MODE_FLOAT && GP_REG_P (REGNO > (dest)))) > + if (GP_REG_P (REGNO (dest))) > { > if ((insn = th_memidx_output_index (dest, src, mode, true))) > return insn; > if ((insn = th_memidx_output_modify (dest, src, mode, true))) > return insn; > } > - else if (GET_MODE_CLASS (mode) == MODE_FLOAT && HARDFP_REG_P (REGNO > (dest))) > + else if (HARDFP_REG_P (REGNO (dest))) > { > if ((insn = th_fmemidx_output_index (dest, src, mode, true))) > return insn; > @@ -936,16 +945,14 @@ th_output_move (rtx dest, rtx src) > } > else if (dest_code == MEM && (src_code == REG || src == CONST0_RTX (mode))) > { > - if (GET_MODE_CLASS (mode) == MODE_INT > - || src == CONST0_RTX (mode) > - || (GET_MODE_CLASS (mode) == MODE_FLOAT && GP_REG_P (REGNO (src)))) > + if (src == CONST0_RTX (mode) || GP_REG_P (REGNO (src))) > { > if ((insn = th_memidx_output_index (dest, src, mode, false))) > return insn; > if ((insn = th_memidx_output_modify (dest, src, mode, false))) > return insn; > } > - else if (GET_MODE_CLASS (mode) == MODE_FLOAT && HARDFP_REG_P (REGNO > (src))) > + else if (HARDFP_REG_P (REGNO (src))) > { > if ((insn = th_fmemidx_output_index (dest, src, mode, false))) > return insn; > diff --git a/gcc/config/riscv/vector.md b/gcc/config/riscv/vector.md > index 7fbee404694..cbe39fd3cc9 100644 > --- a/gcc/config/riscv/vector.md > +++ b/gcc/config/riscv/vector.md > @@ -1357,7 +1357,7 @@ (define_expand "@mov<VB:mode><P:mode>_lra" > (define_insn_and_split "*mov<V_FRACT:mode><P:mode>_lra" > [(set (match_operand:V_FRACT 0 "reg_or_mem_operand" "=vr, m,vr") > (match_operand:V_FRACT 1 "reg_or_mem_operand" " m,vr,vr")) > - (clobber (match_scratch:P 2 "=&r,&r,r"))] > + (clobber (match_operand:P 2 "" "=&r,&r,X"))] > "TARGET_VECTOR && (lra_in_progress || reload_completed)" > "#" > "&& reload_completed" > @@ -1379,7 +1379,7 @@ (define_insn_and_split "*mov<V_FRACT:mode><P:mode>_lra" > (define_insn_and_split "*mov<VB:mode><P:mode>_lra" > [(set (match_operand:VB 0 "reg_or_mem_operand" "=vr, m,vr") > (match_operand:VB 1 "reg_or_mem_operand" " m,vr,vr")) > - (clobber (match_scratch:P 2 "=&r,&r,r"))] > + (clobber (match_operand:P 2 "" "=&r,&r,X"))] > "TARGET_VECTOR && (lra_in_progress || reload_completed)" > "#" > "&& reload_completed" > @@ -1507,7 +1507,7 @@ (define_expand "@mov<VLS_AVL_REG:mode><P:mode>_lra" > (define_insn_and_split "*mov<VLS_AVL_REG:mode><P:mode>_lra" > [(set (match_operand:VLS_AVL_REG 0 "reg_or_mem_operand" "=vr, m,vr") > (match_operand:VLS_AVL_REG 1 "reg_or_mem_operand" " m,vr,vr")) > - (clobber (match_scratch:P 2 "=&r,&r,r"))] > + (clobber (match_operand:P 2 "" "=&r,&r,X"))] > "TARGET_VECTOR && (lra_in_progress || reload_completed) > && (register_operand (operands[0], <VLS_AVL_REG:MODE>mode) > || register_operand (operands[1], <VLS_AVL_REG:MODE>mode))" > diff --git a/gcc/testsuite/gcc.target/riscv/fmv-lgamma.c > b/gcc/testsuite/gcc.target/riscv/fmv-lgamma.c > new file mode 100644 > index 00000000000..dabd143d212 > --- /dev/null > +++ b/gcc/testsuite/gcc.target/riscv/fmv-lgamma.c > @@ -0,0 +1,20 @@ > +/* { dg-do compile { target { rv64 } } } */ > +/* { dg-options "-O2 -march=rv64gc -mabi=lp64d -mtune=xt-c9501fdvt" } */ > + > +/* Keep the floating-point value in an FPR when its bits are also used by > + an integer comparison. Reduced from lgamma. */ > + > +double > +f (double x) > +{ > + if (((union { double d; unsigned long u; }) { x }).u) > + return 0; > + > + return x * (x * 3); > +} > + > +/* The floating-point operations must not reload x from a GPR or a spill > + slot after the integer comparison. Allow fmv.d.x from zero, which > + constructs the return value of zero. */ > +/* { dg-final { scan-assembler-not {\tfmv\.d\.x\t[^,\n]+,(?!zero)} } } */ > +/* { dg-final { scan-assembler-not {\tfsd\t} } } */ > diff --git a/gcc/testsuite/gcc.target/riscv/xtheadmemidx-fpr-int-1.c > b/gcc/testsuite/gcc.target/riscv/xtheadmemidx-fpr-int-1.c > new file mode 100644 > index 00000000000..78e98f1af3e > --- /dev/null > +++ b/gcc/testsuite/gcc.target/riscv/xtheadmemidx-fpr-int-1.c > @@ -0,0 +1,8 @@ > +/* { dg-do assemble { target { rv64 } } } */ > +/* { dg-options "-O3 -march=rv64imafdc_xtheadfmemidx_xtheadmemidx > -mabi=lp64d -save-temps" } */ > + > +#include "xtheadmemidx-fpr-int.h" > + > +/* { dg-final { scan-assembler {\mth\.fsrd\s} } } */ > +/* { dg-final { scan-assembler {\mth\.fsrw\s} } } */ > +/* { dg-final { scan-assembler-not {\mth\.[ls][a-z]+\s+f[ast][0-9]+,} } } */ > diff --git a/gcc/testsuite/gcc.target/riscv/xtheadmemidx-fpr-int-2.c > b/gcc/testsuite/gcc.target/riscv/xtheadmemidx-fpr-int-2.c > new file mode 100644 > index 00000000000..8b54158ae4a > --- /dev/null > +++ b/gcc/testsuite/gcc.target/riscv/xtheadmemidx-fpr-int-2.c > @@ -0,0 +1,7 @@ > +/* { dg-do assemble { target { rv64 } } } */ > +/* { dg-options "-O3 -march=rv64imafdc_xtheadmemidx -mabi=lp64d -save-temps" > } */ > + > +#include "xtheadmemidx-fpr-int.h" > + > +/* { dg-final { scan-assembler-not {\mth\.f[ls]} } } */ > +/* { dg-final { scan-assembler-not {\mth\.[ls][a-z]+\s+f[ast][0-9]+,} } } */ > diff --git a/gcc/testsuite/gcc.target/riscv/xtheadmemidx-fpr-int.h > b/gcc/testsuite/gcc.target/riscv/xtheadmemidx-fpr-int.h > new file mode 100644 > index 00000000000..ddbce59bb6d > --- /dev/null > +++ b/gcc/testsuite/gcc.target/riscv/xtheadmemidx-fpr-int.h > @@ -0,0 +1,25 @@ > +/* Keep integer-mode values in FPRs across an empty asm, so the tests do > + not depend on IRA choosing a particular class for an unconstrained copy. > */ > + > +#define STORE(NAME, ITYPE, FTYPE) \ > +void NAME (ITYPE *dst, long index, FTYPE value) \ > +{ \ > + union { FTYPE f; ITYPE i; } bits = { value }; \ > + asm ("" : "+f" (bits.i)); \ > + dst[index] = bits.i; \ > +} > + > +STORE (store_di, unsigned long long, double) > +STORE (store_si, unsigned int, float) > + > +void > +store_post (unsigned long long *dst, double value, int n) > +{ > + union { double f; unsigned long long i; } bits = { value }; > + asm ("" : "+f" (bits.i)); > + for (int i = 0; i < n; ++i) > + { > + *dst++ = bits.i; > + asm volatile ("" ::: "memory"); > + } > +} > -- > 2.47.1 > This message and any attachment are confidential and may be privileged or otherwise protected from disclosure. If you are not an intended recipient of this message, please delete it and any attachment from your system and notify the sender immediately by reply e-mail. Unintended recipients should not use, copy, disclose or take any action based on this message or any information contained in this message. Emails cannot be guaranteed to be secure or error free as they can be intercepted, amended, lost or destroyed, and you should take full responsibility for security checking. 本邮件及其任何附件具有保密性质,并可能受其他保护或不允许被披露给第三方。如阁下误收到本邮件,敬请立即以回复电子邮件的方式通知发件人,并将本邮件及其任何附件从阁下系统中予以删除。如阁下并非本邮件写明之收件人,敬请切勿使用、复制、披露本邮件或其任何内容,亦请切勿依本邮件或其任何内容而采取任何行动。电子邮件无法保证是一种安全和不会出现任何差错的通信方式,可能会被拦截、修改、丢失或损坏,收件人需自行负责做好安全检查。
